Advances in Intelligent Systems and Computing 847

Lakhmi C. Jain · Valentina E. Balas

Prashant Johri Editors

Data and
Communication
Networks
Proceedings of GUCON 2018
Advances in Intelligent Systems and Computing

Volume 847

Series editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
e-mail: kacprzyk@[Link]

[Link]@[Link]
The series “Advances in Intelligent Systems and Computing” contains publications on theory,
applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually all
disciplines such as engineering, natural sciences, computer and information science, ICT, economics,
business, e-commerce, environment, healthcare, life science are covered. The list of topics spans all the
areas of modern intelligent systems and computing such as: computational intelligence, soft computing
including neural networks, fuzzy systems, evolutionary computing and the fusion of these paradigms,
social intelligence, ambient intelligence, computational neuroscience, artificial life, virtual worlds and
society, cognitive science and systems, Perception and Vision, DNA and immune based systems,
self-organizing and adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics including
human-machine teaming, knowledge-based paradigms, learning paradigms, machine ethics, intelligent
data analysis, knowledge management, intelligent agents, intelligent decision making and support,
intelligent network security, trust management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are primarily proceedings
of important conferences, symposia and congresses. They cover significant recent developments in the
field, both of a foundational and applicable character. An important characteristic feature of the series is
the short publication time and world-wide distribution. This permits a rapid and broad dissemination of
research results.

Advisory Board
Chairman

Nikhil R. Pal, Indian Statistical Institute, Kolkata, India

e-mail: nikhil@[Link]

Members
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing, Universidad Central de Las Villas, Santa
Clara, Cuba
e-mail: rbellop@[Link]
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
e-mail: escorchado@[Link]
Hani Hagras, School of Computer Science & Electronic Engineering, University of Essex, Colchester, UK
e-mail: hani@[Link]
László T. Kóczy, Department of Information Technology, Faculty of Engineering Sciences, Győr, Hungary
e-mail: koczy@[Link]
Vladik Kreinovich, Department of Computer Science, University of Texas at El Paso, El Paso, TX, USA
e-mail: vladik@[Link]
Chin-Teng Lin, Department of Electrical Engineering, National Chiao Tung University, Hsinchu, Taiwan
e-mail: ctlin@[Link]

Jie Lu, Faculty of Engineering and Information, University of Technology Sydney, Sydney, NSW, Australia
e-mail: [Link]@[Link]
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute of Technology, Tijuana, Mexico
e-mail: epmelin@[Link]

Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro, Rio de Janeiro, Brazil
e-mail: nadia@[Link]
Ngoc Thanh Nguyen, Wrocław University of Technology, Wrocław, Poland
e-mail: [Link]@[Link]
Jun Wang, Department of Mechanical and Automation, The Chinese University of Hong Kong, Shatin,
Hong Kong
e-mail: jwang@[Link]

More information about this series at [Link]

[Link]@[Link]
Lakhmi C. Jain Valentina E. Balas
•

Prashant Johri
Editors

Data and Communication

Networks
Proceedings of GUCON 2018

123
[Link]@[Link]
Editors
Lakhmi C. Jain Valentina E. Balas
University of Technology Sydney Aurel Vlaicu University of Arad
Sydney, Australia Arad, Romania

and Prashant Johri

School of Computing Science and
University of Canberra Engineering
Canberra, Australia Galgotias University
Greater Noida, Uttar Pradesh, India
and

Liverpool Hope University

Liverpool, UK

and

KES International
Shoreham-by-Sea, UK

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-981-13-2253-2 ISBN 978-981-13-2254-9 (eBook)
[Link]

Library of Congress Control Number: 2018960207

© Springer Nature Singapore Pte Ltd. 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore

[Link]@[Link]
Preface

The book constitutes selected high-quality papers presented in International

Conference on Computing, Power, and Communication Technologies 2018
(GUCON 2018) organized by Galgotias University, India, in September 2018. It
discusses issues in electrical, computer, and electronics engineering and technolo-
gies. The selected papers are organized into three sections—cloud computing and
computer networks; data mining and big data analysis; and machine learning and
systems. In-depth discussions on various issues under topics provide an interesting
compilation for researchers, engineers, and students.
We are thankful to all the authors that have submitted papers for keeping the
quality of the GUCON 2018 at high levels. The editors of this book would like to
acknowledge all the authors for their contributions and the reviewers. We have
received an invaluable help from the members of the International Program
Committee and the chairs responsible for different aspects of the workshop. We
appreciate also the role of Special Sessions Organizers. Thanks to all of them, we
had been able to collect many papers on interesting topics, and during the con-
ference, we had very interesting presentations and stimulating discussions.
Our special thanks go to Janus Kacprzyk (Editor in Chief, Springer, Advances in
Intelligent Systems and Computing Series) for the opportunity to organize this
guest-edited volume.
We are grateful to Springer, especially to Dr. Thomas Ditzinger (Senior Editor,
Applied Sciences and Engineering, Springer-Verlag), for the excellent collabora-
tion, patience, and help during the evolvement of this volume.
We hope that the volume will provide useful information to professors,
researchers, and graduated students in the area of soft computing techniques and
applications, and all will find this collection of papers inspiring, informative, and
useful. We also hope to see you at a future GUCON event.

Australia/UK Lakhmi C. Jain

Romania Valentina E. Balas
India Prashant Johri
Sep 2018

[Link]@[Link]
GUCON 2018 Organization

Chief Patrons

Mr. Suneel Galgotia, Chancellor, Galgotias University, India

Mr. Dhruv Galgotia, CEO, Galgotias University, India

Patron

Prof. Renu Luthra, Vice-Chancellor, Galgotias University, India

Co-Patrons

Prof. Harish C. Rai, EX Advisor, All India Council for Technical Education, India
Prof. Sibaram Khara, Pro Vice Chancellor, Galgotias University, India

General Chairs

Prof. S. N. Singh, Vice-Chancellor, MMMUT, Gorakhpur, India

Prof. J. Ramkumar, Immediate past Chairman IEEE UP Section, India
Prof. Priestly Shan, SECE, Galgotias University, India
Prof. Prabhakar Tiwari, MMMUT, Gorakhpur, India

Conference Chair and Convener

Prof. Rabindra Nath Shaw, India

vii

[Link]@[Link]
viii GUCON 2018 Organization

Technical Program Committee Chairs

Prof. Lakhmi C. Jain, Australia

Prof. Prashant Johri, India

Publication Chairs

Prof. Valentina E. Balas, Romania

Prof. Jitendra Verma, India

Volume Editors

Prof. Lakhmi C. Jain, Australia

Prof. Valentina E. Balas, Romania
Prof. Prashant Johri, India

Advisory Committee

Prof. Bimal K. Bose, USA

Prof. Muhammad H. Rashid, USA
Prof. Muhammet Koksal, Turkey
Prof. Nasrudin Bin Abd Rahim, Malaysia
Prof. Hamid Ali Abed Al-asadi., Iraq
Prof. Tarek Bouktir, Algeria
Prof. Subramaniam Ganesan, USA
Prof. Claudio Moraga Roco, Spain
Prof. Suganthan P. N., Singapore
Prof. Sanjib Kumar P., Singapore
Prof. Ralph Kennel, Germany
Prof. Udaya Madawala, New Zealand
Prof. Dylan Lu, Australia, India
Prof. Yeap Peik Foong, India
Prof. Terence Karran, India
Prof. B. Chitti Babu, U.K
Prof. Jyh-Horng Chou, Taiwan
Prof. Saad Mekhilef, Malaysia
Prof. Yurii Boreisha, USA

[Link]@[Link]
GUCON 2018 Organization ix

Track Chairs

Cloud Computing, Computer Networks, Automata, Computer

Architecture

Prof. Vincenzo Piuri, Italy

Prof. Anupam Chattopadhyay, Singapore

Data Mining and Big Data Analysis

Prof. George A. Tsihrintzis, Greece

Prof. Ajay K. Datta, USA

Machine Learning Algorithms

Prof. Farzin Deravi, UK

Prof. Maria Virvou, Greece

GUCON 2018 Editorial Board Members

Conference Chair (Covener)

Volume Editors
Technical Chairs
Publication Chairs
Track Chairs

Technical Program Committee and Reviewers

Dr. A. R. Abhyankar
Dr. Aditi Sharan
Dr. Ajay Mittal
Dr. Sudhir Kumar Sharma
Dr. Ajai Jain
Dr. Alok Kushwaha
Dr. Amit Agarwal
Dr. Amalendu Patnaik

[Link]@[Link]
x GUCON 2018 Organization

Dr. Anil K. Ahlawat

Dr. Anil K. Singh
Dr. Anuradha
Dr. Arun Kumar Verma
Dr. Aseem Chandel
Dr. Asheesh K. Singh
Dr. Ashutosh Dixit
Dr. Asif Ekbal
Dr. B. Dushmanta Kumar Patro
Dr. Baij Nath Kaushik
Dr. Bhaskar Biswas
Dr. Bharat Singh Rajpurohit
Dr. C. Patvardhan
Dr. C. Rama Krishna
Dr. C. K. Nagpal
Dr. Chandra Sekaran
Dr. Chiranjeev Kumar
Dr. Chittaranjan Hota
Dr. D. Bhagwan Das
Dr. D. A. Mehta
Dr. D. S. Kushwaha
Dr. D. S. Yadav
Dr. Desh Deepak Sharma
Dr. Dhram Singh
Dr. Dimple J. Gupta
Dr. Diwakar Bhardwaj
Dr. Girish Patnaik
Dr. Jai Govind Singh
Dr. Joy Deep Mitra
Dr. K. V. Arya
Dr. Kiran Kumar Pattanaik
Dr. Kishor K. Bhoyar
Dr. Komal Kumar Bhatia
Dr. Lalit Kumar Awasthi
Dr. M. K. Dutta
Dr. M. P. Singh
Dr. Madhavi Sinha
Dr. Manisha Sharma
Dr. Mohd. Rihan
Dr. Mayank Pandey
Dr. Munesh C. Trivedi
Dr. N. Badal, Knit
Dr. Nanhay Singh
Dr. Narendra Kohli
Dr. Naresh Chauhan

[Link]@[Link]
GUCON 2018 Organization xi

Dr. Naveen Kumar

Dr. Neelam Duhan
Dr. Neeraj Tyagi
Dr. O. P. Verma
Dr. Pooja Jain
Dr. Pooja Pathak
Dr. Prabhat Ranjan
Dr. Prabhakar Tiwari
Dr. Prabin Panigrahi
Dr. Pragya Dwivedi
Dr. Pradeep Sharma
Dr. Pramod Kumar
Dr. Pramod Kumar Singh
Dr. Punam Bedi
Dr. R. K. Singh
Dr. R. S. Yadav
Dr. R. S. Rao
Dr. Rahul Rishi
Dr. Rajesh Prasad
Dr. Reena Dadhich
Dr. Ruchika Malhotra
Dr. S. P. Tripathi
Dr. Sapna Gambhir
Dr. Suneeta Agarwal
Dr. Sujoy Das
Dr. Sukomal Pal
Dr. Sunil Kumar Khatri
Dr. Tanveer Siddiqui
Dr. Tarun Shrimali
Dr. Vasudha Bhatnagar
Dr. Vishal Bhatnagar
Dr. Yashpal Singh
Prof. Herbert H. C. Lu
Dr. Senthilrajan Agni
Dr. Abhineet Anand
Dr. Anurag Baghel
Dr. Balamurugan Balusamy
Dr. Priti Bansal
Dr. Sonia Bansal
Dr. Annappa Basava
Dr. Rohit Beniwal
Dr. Vandana Bhasin
Dr. Rodrigo Bortoletto
Dr. John Moses Cyril
Dr. Pinaki Chakraborty

[Link]@[Link]
xii GUCON 2018 Organization

Dr. Sansar Chauhan

Dr. Rahul Chaurasiya
Dr. Surya Deo Choudhary
Dr. Anurag Dixit
Dr. Nripendra Narayan Das
Dr. Indrani Das
Dr. Aparna Datt
Dr. Parneeta Dhaliwal
Dr. Chandrakant Divate
Dr. Rajesh Dubey
Dr. Arman Faridi
Dr. Ankush Ghosh
Dr. Utkarsh Goel
Dr. Pallavi Goel
Dr. Amit Goel
Dr. Priyanka Goyal
Dr. Deepak Gupta
Dr. Suneet Gupta
Dr. Raza Haidri
Dr. Syed Shabih Hasan
Dr. Manas Hati
Dr. Brijesh Iyer
Dr. Manisha Jailia
Dr. Prashant Johri
Dr. Jegathesh Amalraj Joseph
Dr. Sandeep K. Singh
Dr. Vinay Kumar
Dr. Amita Kapoor
Dr. Sandhya Katiyar
Dr. Anvesha Katti
Dr. Ruqaiya Khanam
Dr. Aanchal Khatri
Dr. Shrawan Kumar
Dr. Devendra Kumar
Dr. Avneesh Kumar
Dr. Arun Kumar
Dr. Sanjeev Kumar
Dr. Vipin Kumar
Dr. Sanjay Kumar
Dr. Bhavnesh Kumar
Dr. Sandeep Kumar
Dr. Neetesh Kumar
Dr. Mohanraj M.
Dr. Ramakrishnan Malaichamy
Dr. Manas Kumar Mishra

[Link]@[Link]
GUCON 2018 Organization xiii

Dr. Baibaswata Mohapatra

Dr. Thiyagarajan Muthunatesan
Dr. Rashid Mahmood
Dr. Yogendra Meena
Dr. Gitanjali Mehta
Dr. A. K. Mishra
Dr. Keshav Niranjan
Dr. Manoj Panda
Dr. Sanjeev Pippal
Dr. V. A. Sankar Ponnapalli
Dr. Shiv Prakash
Dr. Sheetla Prasad
Dr. Mohammed Abdul Qadeer
Dr. R. Gunasundari Ranganathan
Dr. Ranjeet Kumar Ranjan
Dr. Rohit Raja
Dr. Bharti Rana
Dr. Mukesh Rawat
Dr. Navaid Zafar Rizvi
Dr. Pravinth Raja S.
Dr. Anil Kumar Sagar
Dr. Rajeev Sharma
Dr. Birendra Kumar Sharma
Dr. Shrddha Sagar
Dr. Jyoti Sahni
Dr. Mohd. Saifuzzaman
Dr. Kavita Saini
Dr. Kamalesh Sethuramalingam
Dr. Priestly Shan
Dr. Gavaskar Shanmugam
Dr. Dilip Kumar Sharma
Dr. R. P. Sharma
Dr. Mayank Sharma
Dr. Sudhir Sharma
Dr. Lokesh Kumar Sharma
Dr. Vishnusharma
Dr. Jitendra Singh
Dr. Girish Singh
Dr. Karan Singh
Dr. Harikesh Singh
Dr. Prashant Singh
Dr. Neetay Singh
Dr. Ajay Shanker Singh
Dr. Arun Solanki
Dr. Subhranil Som

[Link]@[Link]
xiv GUCON 2018 Organization

Dr. Ritesh Srivastava

Dr. Vijayalakshmi Subramanian
Dr. Hardeo Kumar Thakur
Dr. Pradeep Tomar
Dr. Shashi Kant Verma
Dr. Sohan Kumar Yadav
Dr. Vinod Yadav
Dr. Dileep Yadav
Dr. Chandra Yadav
Dr. Emre Yay
Dr. Aasim Zafar
Dr. Usha Chauhan
Dr. Chetna Dabas
Dr. Sanjoy Das
Dr. Sumithra Gavaskar
Dr. Vimal Kumar

[Link]@[Link]
Contents

A Novel Hash-Based Mutual RFID Tag Authentication Protocol . . . . . . 1

Mansi Saxena, Rabindra Nath Shaw and Jitendra Kumar Verma
Analysis of Binary PSK Modulations Over the Line-of-Sight Plus
Scatter Fading Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Veenu Kansal and Simranjit Singh
Approximate Bit Error Rate of DPSK with Imperfect Phase
Noise in TWDP Fading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Veenu Kansal and Simranjit Singh
Analysis of Workloads for Cloud Infrastructure Capacity Planning . . . 29
Eva Patel and Dharmender Singh Kushwaha
Trust-Based Scheme for Location Finding in VANETs Using
Trustworthiness of Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Sanjoy Das, Indrani Das, Rishi Pal Singh, Prashant Johri
and Ashwini Kumar
Implementation and Analysis of Energy Efficiency of M-ary
Modulation Schemes for Wireless Sensor Network . . . . . . . . . . . . . . . . 57
Samir Ahmad Sheikh and Sindhu Hak Gupta
Protecting Multicast Sessions Against Single-Link Failure in
Survivable WDM Mesh Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
B. Mohapatra and Kunal Sain
A Cross-Layer Routing Protocol for Wireless Sensor Networks . . . . . . . 83
Pallavi Yarde, Sumit Srivastava and Kumkum Garg
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot . . . 93
Shivam Mishra and V. K. Sachan
Detection of RREQ Flooding Attacks in MANETs . . . . . . . . . . . . . . . . . 109
B. Nithya, Aishwarya Nair and A. S. Sreelakshmi

xv

[Link]@[Link]
xvi Contents

Selection of Software Development Model Using TOPSIS

Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Dayanand Gaur and Sakshi Aggarwal
Situation-Aware Conditional Sensing in Disaster-Prone Areas
Using Unmanned Aerial Vehicles in IoT Environment . . . . . . . . . . . . . . 135
J. Sathish Kumar, Mukesh A. Zaveri, Saurabh Kumar and Meghavi Choksi
An Empirical Analysis of Collaborative Filtering Algorithms
for Building a Food Recommender System . . . . . . . . . . . . . . . . . . . . . . 147
Ashique Mohaimin Ornab, Sakia Chowdhury and Seevieta Biswas Toa
Performance Analysis of Iris Recognition System . . . . . . . . . . . . . . . . . 159
Ruqaiya Khanam, Zohreen Haseen, Nighat Rahman and Jugendra Singh
Challenges in Mining Big Data Streams . . . . . . . . . . . . . . . . . . . . . . . . . 173
Veena Tayal and Ritesh Srivastava
Controlling of Non-minimum Phase System Using Harmony Search
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Vivek Kumar Jaiswal, Anurag Singh, Shekhar Yadav
and Shyam Krishna Nagar
Framework for Real-World Event Detection Through Online Social
Networking Sites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Ritesh Srivastava, M. P. S. Bhatia, Veena Tayal and J. K. Verma
Novel Architecture for Internet of Things and Blockchain
Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Chetna Dabas and Aniket Dabas
Homomorphism Between Fuzzy Set-Valued Information Systems . . . . . 219
Waseem Ahmed, M. M. Sufyan Beg and Tanvir Ahmad
A Novel Approach for Predicting the Outcome of Request
in RAOP Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
Amreen Ahmad, Tanvir Ahmad and Abhishek Bhatt
A Soft Computing Methodology for Estimation and Forecasting
of Daily Global Solar Radiation (DGSR) . . . . . . . . . . . . . . . . . . . . . . . . 235
J. Christy Martina and T. Amudha
A Shapley Value Approach for Community Detection in Social
Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Amreen Ahmad, Tanvir Ahmad, Abhishek Bhatt and Sadaf Siddiqui
Multiple Face Detection Using Hybrid Features with SVM
Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Sandeep Kumar, Sukhwinder Singh and Jagdish Kumar

[Link]@[Link]
Contents xvii

A Brief Overview on Generative Adversarial Networks . . . . . . . . . . . . . 267

Raj Patel
Prediction of PhotoVoltaic Power Generation Using Monte
Carlo Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Gautam Seth, K. Abhay Prithvi, Arpit Paruthi, Saksham Jain
and Umang Soni
Storage and Analysis of Synchrophasor Data for Event Detection
in Indian Power System Using Hadoop Ecosystem . . . . . . . . . . . . . . . . . 291
Akhilendra Pratap Singh, G. Hemant Kumar, Subhendu Sekhar Paik
and Diptendu Sinha Roy
Analysing Trends in Student’s Performance Across Maharashtra
Through Non-adaptive and Adaptive Online Assessments Based
on the Underlying Framework of Classical Test and Item
Response Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Chandrani Singh and Ajit Pandey
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

[Link]@[Link]
About the Editors

Dr. Lakhmi C. Jain, Ph.D., M.E., B.E.(Hons) is Fellow (Engineers Australia)

and is with the University of Canberra, Australia, University of Technology
Sydney, Australia and Liverpool Hope University, Liverpool, UK. He founded the
KES International for providing a professional community with the opportunities
for publications, knowledge exchange, cooperation, and teaming. Involving around
5000 researchers drawn from universities and companies worldwide, KES facili-
tates international cooperation and generates synergy in teaching and research. KES
regularly provides networking opportunities for professional community through
one of the largest conferences of its kind in the area of KES. His interests focus on
the artificial intelligence paradigms and their applications in complex systems,
security, e-education, e-health care, unmanned air vehicles, and intelligent agents.

Valentina E. Balas is currently Full Professor in the Department of Automatics

and Applied Software at the Faculty of Engineering, “Aurel Vlaicu” University of
Arad, Romania. She holds a Ph.D. in Applied Electronics and Telecommunications
from Polytechnic University of Timisoara. Dr. Balas is author of more than 280
research papers in refereed journals and international conferences. Her research
interests are in Intelligent Systems, Fuzzy Control, Soft Computing, Smart Sensors,
Information Fusion, Modeling and Simulation. She is the Editor-in Chief to
International Journal of Advanced Intelligence Paradigms (IJAIP) and to
International Journal of Computational Systems Engineering (IJCSysE), member in
Editorial Board member of several national and international journals and is the
director of Intelligent Systems Research Centre in Aurel Vlaicu University of Arad.
She is a member of EUSFLAT, SIAM and a Senior Member IEEE, member in
TC—Fuzzy Systems (IEEE CIS), member in TC—Emergent Technologies (IEEE
CIS), member in TC—Soft Computing (IEEE SMCS).

Dr. Prashant Johri is working as Professor in School of Computing Science and

Engineering, Galgotias University, Greater Noida, India. He completed his M.C.A.
from Aligarh Muslim University and Ph.D. in Computer Science from Jiwaji
University, Gwalior, India. He has also worked as a Professor and Director (M.C.A.),

xix

[Link]@[Link]
xx About the Editors

Noida Institute of Engineering and Technology, (N.I.E.T.). He has served as Chair in

many conferences and affiliated as member of program committee of many confer-
ences. He has supervised 3 Ph.D. students and many [Link]. students for their thesis.
His area of research interests includes big data, data analytics, data retrieval and
predictive analytics, information security, privacy protection, big data open platforms
etc. He is actively publishing in these areas.

[Link]@[Link]
A Novel Hash-Based Mutual RFID Tag
Authentication Protocol

Mansi Saxena , Rabindra Nath Shaw and Jitendra Kumar Verma

Abstract Radio frequency identification (RFID) is an integral part of our life. This
term is coined for short-range radio communication technology. It is used to send
and receive the digital information between stationary location and non-stationary
object or between movable objects. It automates the wireless technology using radio
waves to identify an object. This technology has widespread applications in the
field of security, access control, transportation, etc. In this paper, we analyze an
existing RFID-based protocol and demonstrate that it is insecure against imperson-
ation attack, man-in-middle attack, server-masquerading attack, insider attack, and
denial-of-service. We also propose a novel protocol, namely Encrypted Tag Identity
and Secret Value Protocol, to overcome the shortcomings and loopholes existing in
the surveyed protocols.

Keywords Impersonation attack · Server masquerading · Mutual authentication

Public and private keys · Secret value · Random numbers

1 Introduction

In modern world, radio frequency identification (RFID) has become an important and
integral part of human life for promoting productivity and convenience [1, 2]. RFID is
a wireless technology that uses radio frequency electromagnetic fields to transfer data
from a Tag attached to an object and thus provides non-contact system for the purpose

M. Saxena
AbyM Technology, Noida 201301, Uttar Pradesh, India
R. N. Shaw
School of Electrical Electronics & Communication Engineering,
Galgotias University, Greater Noida 201310, Uttar Pradesh, India
J. K. Verma (B)
School of Computing Science & Engineering,
Galgotias University, Greater Noida 201310, Uttar Pradesh, India
e-mail: [Link]@[Link]

© Springer Nature Singapore Pte Ltd. 2019 1

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
2 M. Saxena et al.

of automatic identification and tracking of moving objects. It has multidimensional

applications like transportation, automated payment system, object traceability, and
access control. RFID refers to a technology where digital data encoded in RFID Tags
are captured by a RFID Reader via radio waves. An RFID system contains three key
components which are as follows.
(i) RFID Tags: It is attached to the object which has to be identified by the help
of carrying identification data.
(ii) RFID Readers: It reads or writes the identification information on Tags using
Radio waves.
(iii) Back-end database: It collect the records related to tagged objects and associate
them with the information that is going to be read by the Reader.
RFID has drawn a significant attention in recent years. Many researchers have con-
tributed a lot of research and products to this technology. In recent years, many RFID-
based protocols are proposed for secure implementation and deployment of this tech-
nology. Srivastava et al. [3] proposed a secure and robust hash-based mutual RFID
Tag authentication protocol in telecare medicine information system. They claimed
that their protocol provides mutual authentication and is secure against eavesdrop-
ping and replay attacks. But after analyzing we demonstrate that the given protocol
is vulnerable to impersonation attack, insider attack, server-masquerading and man-
in-middle attack. Therefore, we introduce a new protocol, namely, Encrypted Tag
Identity and Secret Value Protocol (ETISVP) which fills the loopholes of the exist-
ing protocol and proved to be safe against the impersonation, server-masquerading,
denial-of-service and man-in-middle attacks.
Rest of the paper is organized as follows. Section 2 presents related work, Sect. 3
presents state-of-the-arts of hash-based Tag authentication protocol, Sect. 4 provides
possible security threats from attacks and loopholes of the protocol under observa-
tion followed by proposed protocol and its security analysis along with proposed
comparison of results of existing protocols and proposed protocol in Sects. 5 and 6.
Section 7 concludes the paper.

2 Related Work

Most of the protocols developed for mutual authentication between the Tag and the
Reader in RFID system are based on hash functions. Despite of being not very secure
and containing loopholes, these protocols are easily exploited by the attackers. Many
schemes have been proposed to counter the privacy issues in RFID System. Few of
them are as follows.
Okubhu et al. [4] proposed a hash-based protocol, which was the improved ver-
sion of “Kill Command Feature”-based protocols introduced by Auto-ID Center [5].
However, the high cost of searching tags remained the major flaw of this protocol.
Weis et al. [6] proposed “Hash Lock Technique” which was the improved version of
this protocol. Meanwhile, many protocols have been proposed such as randomized

[Link]@[Link]
A Novel Hash-Based Mutual RFID Tag Authentication Protocol 3

hash lock scheme by MIT, anonymous ID scheme by NTT, external re-encryption

scheme by RSA Laboratory and XOR based onetime pad scheme by RSA Labora-
tory. Tsudik [7] proposed an improved authentication protocol, namely, Yet-Another
Trivial RFID Authentication Protocol (YATRAP), in order to provide tracing resis-
tance Tag authentication through monotonically increasing time-stamps on the Tag.
Furthermore, they proposed a new protocol [8] which was found vulnerable to replay
attack due to lack of Reader’s authenticity.
Similarly, Chatmon et al. proposed an anonymous RFID authentication protocol
in [9]. Sun et al. [10] proposed a RFID technology to guard inpatient medication
safety. In 2009, Huang and Ku [11] proposed the RFID protocol for medication
safety of inpatient, which was found to be vulnerable to denial-of-service and replay
attacks.
Chien et al. [12] proposed improved version of [11] which was still remained vul-
nerable to the impersonation and replay attacks. Peris-Lopez et al. [13] introduced
a new concept of IS-RFID system; however, it is vulnerable to easy manipulation
[14–16]. Chen et al. [17] proposed tamper resistant protocol which fails to guaran-
tee the safety against impersonation, desynchronization attacks and traceability. To
overcome the desynchronization problem, Cho et al. [18] followed by Kim [15, 16]
proposed hash-based RFID Tag mutual authentication protocols. Suja et al. proposed
an RFID authentication protocol based on cyclic redundancy check (CRC) and ham-
ming distance calculation between Reader to Tag. They claimed that their protocol
resists against tracing and cloning attacks in the most efficient way [19].
Considering the great success of RFID technology in Telecare sector, Srivastava
et al. [3] proposed a hash-based mutual RFID Tag authentication protocol in telecare
medicine information system. We have reviewed and cross-examined their protocol
against all possible attacks, and we have shown possible loopholes of this protocol
in subsequent section (Table 1).

Table 1 Preliminaries
Symbol details
Ik Identity of the kth Tag
 Concatenation operation
N Random number
Nr Random number of Reader
Nt Random number of Tag
S Secret value
Sj Secret value used in the jth session
H(·) Hash function
⊕ Bitwise XOR
DB Database

[Link]@[Link]
4 M. Saxena et al.

3 State-of-the-Arts

This section presents state-of-the-arts of hash-based mutual RFID Tag authentication

protocol and its functioning. The underlying work is the basis of our proposed work
in this paper. The hash-based Tag authentication protocol constitutes three phases
that are as follows.

3.1 Pre-phase

Initially, the Tag and Data server share the Tag identity I k , hash function H(·), and the
secret value S j . The Tag and the Reader have their own random number generators.

3.2 Reader’s Request

The Reader generates the random number N r and sends a request to Tag with this
random number.

3.3 Tag’s Response

On receiving request from the Reader, Tag is invoked and Tag generates a random
number N t and subsequently perform the following computations:
• X = H(S j  I k ) ⊕ N t
• Calculate Y  X ⊕ H(I k  N r  N t ), and then
• Z  H(Y ⊕ T 1 ⊕ N t ).
Tag sends the response message (X; Z; T 1 ) to the Reader and Reader sends the
response message (X; Z; T 1 ; N r ) to the Data server after adding N r to it.

3.4 Data Servers Response and Tag Authentication

Data server after extracting the Tag identity (I k ) and shared Secret value (S j ) from
the database performs the following computations:
• If the expected legitimate time interval for transmission delay, T < (T 2 − T 1 )
the server rejects the login request, where T 2 is current time-stamp at server.
• Computes N *t  X ⊕ H(S j  I K ).
• Check Z *  H(X ⊕ H(I k  N r  N *t ) ⊕ N * ⊕ T 1 ) ≈ Z.

[Link]@[Link]
A Novel Hash-Based Mutual RFID Tag Authentication Protocol 5

• Checks until Z * is equal to Z as extracted from response message sent by Reader.

• Computes W  H(X ⊕ H(I k  N r  N t ) ⊕ T 2 ⊕ S j ).
When Z * ≈ Z, the Data server sends the Tag data to the Reader. It sends the
message U  DATA  W to the Reader. DATA is the information of Tag that needs
to be transmitted.

3.5 On Receiving U, Reader’s Response

On receiving U from Data server, Reader extracts the DATA from U and sends the
remaining part to the Tag for further communication.

3.6 Data Server Authentication and Secret Value Updation

If T < (T 3 − T 2 ), where T 3 is the current time-stamp and T is the expected

legitimate time interval for transmission delay then Tag rejects the request.
• Tag computes W *  H(Y ⊕ S j ⊕ T 2 ).
• If W * ≈ W , Tag authenticates the Data server.
• Updates the S j to S j+1  H(S j ⊕ N r ⊕ N t ) on both Tag and server side.

4 Security Analysis against Possible Attacks and Loopholes

We perform security analysis of the protocol which emphasizes that the protocol is
not secure and is vulnerable to various attacks such as any attacker can invoke the
Tag by sending the random number N r . The Tag performs following operations:
• X  H(S j  I k ) ⊕ N t
• Calculates Y  X ⊕ H(I k  N r  N t )
• Z  H(Y ⊕ T 1 ⊕ N t ).
Furthermore, the Tag sends the response message (X; Z; T 1 ) to the Attacker
assuming him to be the correct Reader. The attacker sends the response message (X;
Z; T 1 ; N r ) to the Data server adding N r to it. Now with these values, the attacker can
easily fetch the DATA from the message U  DATA  W sent by the Data Server
assuming the attacker to be the correct Reader.

[Link]@[Link]
6 M. Saxena et al.

4.1 Impersonation Attack

In the first step of protocol, the Reader sends the request to the Tag along with a
random number N r . At Tag’s side, there is no verification mechanism to ensure that
request is made by the correct Reader.

4.2 Server-Masquerading Attack

Since, the Tag identity (I k ) and the Secret value (S j ) are stored as plaintext in the
database. Any privileged insider having access to database can easily get these values
and behave as the Data server by intercepting the response message (X; Z; T 1 ; N r )
send by the Reader. The Attacker with values (I k ; S j ; X; Z; T 1 ; N r ) computes:
• N *t  X ⊕ H(S j  I k )
• W  H(X ⊕ H(I k  N r  N t ) ⊕ T 2 ⊕ S j )
• U  InvalidDATA  W.
In this way, the Attacker can send the invalid data to the Reader. The Reader will
accept the message U from the Attacker, assuming him to be the Data server. Since,
there is no mechanism for the verification of the correct Data server at the Reader’s
side.

4.3 Denial-of-Service Attack

The protocol is not fail-safe against the computation exhaustive attacks. At Data
Server, there is an authentication step to verify the Reader.
• Check Z*  H(X ⊕ H(I k  N r  N *t ) ⊕ N *t ⊕ T 1 ) ≈ Z.
But there is no limit on the number of wrong messages can be sent to the Data
server. The Attacker can send number of fake messages to the server which leads to
the excessive computation on the server side keeping the server busy and unable to
process any request.

4.4 Man-in-Middle Attack

All the messages between the Tag, Reader, and Data server are transmitted as plaintext
on the communication channels. These messages can be intercepted and used by the
Attacker as done in server-masquerading attack. Also the attacker can change these
messages on the communication channel in order to invoke the denial-of-service
attack by sending the fake messages to the Data server.

[Link]@[Link]
A Novel Hash-Based Mutual RFID Tag Authentication Protocol 7

4.5 Insider Attack

Insider attacker is one who is having the administrative access to the server. Tag iden-
tity (I k ) and secret value (S j ) are stored as the plaintext to the server. The privileged
insider, who has direct access to the server, can get these parameters and use the
secret information for personal benefit as discussed in server-masquerading attack.

5 Encrypted Tag Identity and Secret Value Protocol:

Proposed Work

We propose Encrypted Tag Identity and Secret Value Protocol (ETISVP) with the
objective to fill the loopholes of the existing protocol. Our protocol uses the encryp-
tion scheme to safely communicate the messages between the Tag, Reader, and Data
server. Encryption makes it impossible for the Attacker to intercept or modify the
message during the transmission on communication channel. Table 1 shows the ter-
minologies which are frequent in ETISVP.
The Tag identity (I k ) and Secret value (S j ) are stored as the hash function cipher
in the database. So, the any privileged insider or attacker cannot hack and use these
values from the database. The functioning of ETISVP is shown in Table 2.

5.1 Pre-phase

Initially, Tag, Reader, and Data server have their private keys for encryption and
share their public keys with each other for decryption.
Hash function H(·), Tag identity (T ) and secret value (V j ) are shared by Tag and
Server. T and V j are stored as H(T ) and H(V j ) in the database. V j is the secret value
of Tag in jth session.

5.2 Readers Request

The Reader generates the random number N r and performs the following computa-
tions.
• Encrypt N r using its private key, K R[PR] {N r }.
• Sends a request to Tag with K R[PR] {N r } on secure communication channel.

[Link]@[Link]
8 M. Saxena et al.

5.3 Tags Response

Decrypt the message received from Reader using Reader’s public key: K R[PU] {N r }.
Generates a random number N t and performs the following computations:
• X  (H(V j )  H(T )) ⊕ N t
• Calculates Y  X ⊕ H(H(T )  N r  N t ), and then
• Z  H(Y ⊕ T 1 ⊕ N t ) where T 1 is the time-stamp at the Tag.

Table 2 Encrypted tag identity and secret value protocol

[Link]@[Link]
A Novel Hash-Based Mutual RFID Tag Authentication Protocol 9

Tag generates the response message (X, Z, T 1 ) and encrypts it using its private key:
K T[PR] {X, Z, T 1 }. Tag sends K T[PR] {X, Z, T 1 } to the Reader on secure communication
channel.

5.4 Reader’s Side

On receiving response of Tag on secure communication channel, the following oper-

ations take place on Reader’s side.
• Decrypt the message received from Tag using Tag’s public key: K T[PU] {X, Z, T 1 }.
• Add N r and encrypt the message using its private key: K R[PR] {X, Z, T 1 , N r }.
• Reader sends K R[PR] {X, Z, T 1 , N r } to the Data server on secure communication
channel.

5.5 Data Server Response and Tag Authentication

Decrypt message received from Reader using Reader’s public key: K R[PU] {X, Z, T 1 ,
N r }. Extracts H(T) and H(V j ) from database and performs the following computa-
tions:
If the expected legitimate time interval for transmission delay, T < (T 2 − T 1 )
• The server rejects the login request, where T 2 is current time-stamp at server.
• Computes N *t  X ⊕ (H(V j )  H(T )).
• Check Z *  H(X ⊕ H(H(T )  N r  N *t ) ⊕ N t ⊕ T 1 ) ≈ Z.
• Checks until Z * is equal to Z.
• Computes W  H(X ⊕ H(H(T)  N r  N t ) ⊕ T 2 ⊕ V j )
• If Z * ≈ Z the Data server generates the message: DATA  W .
• Encrypt it using its private key: K D[PR] {DATA  W }.

5.6 Reader’s Response

On receiving the encrypted data from Data server, Reader performs the following
operations.
• Decrypt the message using Server’s public key: K D[PU] {DATA  W }.
• Extract DATA from {DATA  W }.
• Encrypt W using its private key: K R[PR] {W }.
And sends K R[PR] {W } to the Tag on secure communication channel.

[Link]@[Link]
10 M. Saxena et al.

5.7 Data Server Authentication and Secret Value Updation

On receiving encrypted data from Tag, the following operations take place on Data
server side.
• Decrypt the message using Reader’s public key: K R[PU] {W }.
• Tag rejects the request if, T < (T 3 T 2 ), where T 3 is the current time-stamp and
T is the expected legitimate time interval for transmission delay.
• Tag computes W *  H(Y ⊕ H(V j ) ⊕ T 2 ).
• If W * ≈ W , Tag authenticates the Data server.
• Updates the V j to V j+1  H(H(S j ) ⊕ N r ⊕ N t ) on both Tag and server side.

6 Security Analysis

Proposed protocol ETISVP overcomes the loopholes of the existing protocol hash-
based Tag authentication protocol, and hence, is highly safe against most com-
mon attacks like man-in-middle attack, eavesdropping attack, impersonation attack,
server-masquerading attack and Insider Attack. The analysis of these attacks is pro-
vided in the following subsections.

6.1 Man-in-Middle Attack

In the proposed ETISVP protocol, during the message exchange, sender encrypts
the message using the private key before transmitting it on communication channel.
This message can only be decrypted by the corresponding public key. Therefore,
any Attacker cannot modify the message without having the public key. Since, the
attacker do not have the public key, he cannot use the message to harm the system.

6.2 Impersonation Attack

The Reader in the first step sends the random number to the Tag by encrypting it with
the private key as K R[PR] {N r }. The Tag can decrypt the message by the corresponding
public key. Therefore, the Attacker cannot send the random number to the Tag without
knowing the private key of Reader to encrypt the message. Hence, the Attacker cannot
fool the Tag by behaving as Reader.

[Link]@[Link]
A Novel Hash-Based Mutual RFID Tag Authentication Protocol 11

Table 3 Performance Evaluation

Protocol Computation cost Computation cost Communication
(Tag) (Reader) rounds
YA-TRAP [13] 2H + 3RNG RNG 4
Suja and Arivarasi 3H + 2MOD + RNG RNG 5
[19]
Existing Protocol [3] 2H + RNG RNG 5
ETISVP 4H + RNG + RNG + Encryption 5
Encryption

6.3 Server-Masquerading Attack

According to the proposed protocol, Tag identity (T ) and secret value (V j ) are stored
in the encrypted form in database. Therefore, any privileged insider cannot fetch
these values from database. Also, the message sent by the Reader to the Data server is
encrypted by private key of Reader. Hence, cannot be decrypt by the Attacker without
the corresponding public key. So the Attacker cannot get the required information
for server masquerading.

6.4 Eavesdropping Attack

The proposed protocol is safe against the eavesdropping attack. Since, all the mes-
sages on the communication channel are in encrypted form. Therefore, any attacker
cannot get the valuable information without knowing the corresponding public key
for decryption. Hence, the proposed protocol meets the required challenges (Table 3).

7 Conclusion

In this paper, we have analyzed the hash-based mutual Tag authentication protocol
based on random numbers and synchronized secret value. The protocol proved to
be insecure against the eavesdropping attack, man-in-middle attack, impersonation
attack, denial-of service attack, and insider attack. The protocol is vulnerable to
these attacks due to the secret value, tag identity, and hash function stored in the
database as plaintext. Also, the protocol requires safe communication mechanism
for transferring intermediate computed values over the communication channel.
Proposed ETISVP protocol filled these loopholes by using encryption for mes-
sage exchange and by storing the Tag identity and secret value in the database in
encrypted form. The protocol introduced is secure against these attacks and hence,

[Link]@[Link]
12 M. Saxena et al.

strengthen the security of RFID technology. The ETISVP provides immunity from
several attacks; hence, it is safe to deploy in the industry. However, encryption of data
delays the attack only rather than making it fool proof. In the modern world, we have
several machines which are capable of holding huge computation power which poses
threat to the proposed protocol for decrypting the encrypted information. Therefore,
we are intended to work upon strengthening ETISVP as a future direction of work.
Apart from this, we are also intended to work upon quantum computing environment
as existing protocol are easily deciphered in quantum computing environment due
to its immense computing power.

References

1. Shepard, S.: RFID: Radio Frequency Identification. McGraw Hill Professional (2005)
2. Landt, J.: The history of RFID. IEEE Potentials 24(4), 8–11 (2005)
3. Srivastava, K., Awasthi, A.K., Kaul, S.D., Mittal, R.: A hash based mutual RFID tag authenti-
cation protocol in telecare medicine information system. J. Med. Syst. 39(1), 153 (2015)
4. Ohkubo, M., Suzuki, K., Kinoshita, S., et al.: Cryptographic approach to privacy-friendly tags.
In: RFID Privacy Workshop, Cambridge, USA, vol. 82 (2003)
5. Center, A.-I.: 860 MHz-960 MHz Class-I radio frequency identification tag radio frequency &
logical communication interface specification proposed recommendation version 1.0. 0 (2002)
6. Weis, S.A., Sarma, S.E., Rivest, R.L., Engels, D.W.: Security and privacy aspects of low-cost
radio frequency identification systems. In: Security in Pervasive Computing, pp. 201–212.
Springer (2004)
7. Tsudik, G.: YA-TRAP: yet another trivial RFID authentication protocol. In: Pervasive Comput-
ing and Communications Workshops, 2006. PerCom Workshops 2006. Fourth Annual IEEE
International Conference on, IEEE, pp. 4–7 (2006)
8. Tsudik, G.: A family of dunces: trivial RFID identification and authentication protocols. In:
Privacy Enhancing Technologies, pp. 45–61. Springer (2007)
9. Chatmon, C., van Le, T., Burmester, M.: Secure anonymous RFID authentication protocols.
Florida State University, Department of Computer Science, Tech. Rep
10. Sun, P.R., Wang, B.H., Wu, F.: A new method to guard inpatient medication safety by the
implementation of RFID. J. Med. Syst. 32(4), 327–332 (2008)
11. Huang, H.-H., Ku, C.-Y.: A RFID grouping proof protocol for medication safety of inpatient.
J. Med. Syst. 33(6), 467 (2009)
12. Chien, H.-Y., Yang, C.-C., Wu, T.-C., Lee, C.-F.: Two RFID-based solutions to enhance inpatient
medication safety. J. Med. Syst. 35(3), 369–375 (2011)
13. Peris-Lopez, P., Orfila, A., Mitrokotsa, A., Van der Lubbe, J.C.: A comprehensive RFID solution
to enhance inpatient medication safety. Int. J. Med. Inform. 80(1), 13–24 (2011)
14. Yen, Y.-C., Lo, N.-W., Wu, T.-C.: Two RFID-based solutions for secure inpatient medication
administration. J. Med. Syst. 36(5), 2769–2778 (2012)
15. Kim, H.: Enhanced hash-based RFID mutual authentication protocol. In: Computer Applica-
tions for Security, Control and System Engineering, pp. 70–77. Springer (2012)
16. Kim, H.: RFID mutual authentication protocol based on synchronized secret. Int. J. Secur.
Appl. 7(4), 37–50 (2013)
17. Chen, Y.-Y., Huang, D.-C., Tsai, M.-L., Jan, J.-K.: A design of tamper resistant prescription
RFID access control system. J. Med. Syst. 36(5), 2795–2801 (2012)
18. Cho, J.-S., Yeo, S.-S., Kim, S.K.: Securing against brute-force attack: a hash-based RFID
mutual authentication protocol using a secret value. Comput. Commun. 34(3), 391–397 (2011)
19. Suja, S., Arivarasi, A.: An RFID Authentication Protocol for Security and Privacy. In: Interna-
tional Conference on Computing and Control Engineering (2012)

[Link]@[Link]
Analysis of Binary PSK Modulations
Over the Line-of-Sight Plus Scatter
Fading Model

Veenu Kansal and Simranjit Singh

Abstract In this paper, the average bit error rate (ABER) performance of binary
coherent and non-coherent modulation is investigated over a new fading channel
which is well known as Beaulieu-Xie fading channel. This fading model consists of
multiple no. of line-of-sight (LOS) components with some diffuse power. Specifi-
cally, the closed-form expressions are derived for the exact ABER of binary phase-
shift keying (BPSK) and differential phase-shift keying (DPSK) by using the prob-
ability density function (PDF) approach. The final exact expression for BPSK is
obtained in the form of infinite series which needs a reasonable number of terms
to converge and the final expression for DPSK is expressed in terms of elementary
functions. It is observed that ABER decreases with increase in specular power and/or
with decrease in fading severity. The obtained results are applied to general scenarios
for various values of fading parameters by adjusting the specular power and fading
severity. The numerical results plotted by using the derived expressions show a close
agreement with the results obtained by Monte Carlo simulation.

Keywords Average BER (ABER) · Binary phase-shift keying (BPSK)

Femtocells · High-speed trains

1 Introduction

In wireless communication system, bit error rate (BER) is the important performance
metric to measure the quality of transmission over the fading channels. Over the past
years, several mathematical models such as Rayleigh, Rician, Nakagami-m, k–μ,
and two-wave with diffuse power (TWDP) [1–7] have been used to analyze the
small-scale fading. However, none of the above-mentioned distributions are suitable

V. Kansal · S. Singh (B)

Department of ECE, Punjabi University Patiala, Patiala, India
e-mail: simranjit@[Link]
V. Kansal
e-mail: veenukansal@[Link]

© Springer Nature Singapore Pte Ltd. 2019 13

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
14 V. Kansal and S. Singh

for modeling the fading channel where multiple dominant specular components are
present in addition to the diffuse scatter power. To model such a scenario, a fading
model was recently proposed by Beaulieu and Jiandong [8]. This fading is based on
non-central chi distribution (just like generalized Rician and k–μ fading); however,
it is much more flexible than aforementioned distributions for modeling practical
scenarios with multiple dominant specular waves. This model can be obtained from
the non-central chi distribution as in similar way by which the Nakagami-m fading
is obtained from the central chi distribution. While the Nakagami-m fading is ide-
ally suitable to model scatter fading wireless channels, the Beaulieu-Xie fading is
uniquely efficient to characterize practical LOS channels. It is reported that this new
fading distribution is especially suitable for modeling high-speed trains and future
femtocells channels. One major difference between typical fading channels and the
fading channels of the train station for the high-speed railway is that the K-factor of
the later is considerably smaller [9]. The existing fading models such as Rayleigh
and Rician can be obtained from the Beaulieu-Xie fading channel by adjusting the
parameters K and m.
In the literature, the ABER analysis of wireless fading channels such as Rayleigh,
Rician, and TWDP, etc. can be found in [1–7]. In [1], the authors provide a generalized
approach to analyze BER performance for different types of modulation techniques,
with or without diversity reception over additive white Gaussian noise (AWGN) and
also, same analysis is presented for generalized fading models. An analysis with
DPSK modulation scheme over Rician fading channel is presented in [2]. Further,
error probability for uncoded BPSK is investigated in TWDP fading in [5]. By using
moment-generating function (MGF) approach, the performance of an average symbol
error rate of quadrature amplitude modulation (QAM) is studied over TWDP fading
by deriving the analytical expressions in [6]. The ABER analysis of M-ary phase-shift
keying (MPSK) in the presence of TWDP fading is provided in [7]. In [10], the second
order statistics for Beaulieu-Xie fading is investigated. Furthermore, the analysis of
this fading channel is done with selection combining and equal combining techniques
in [11, 12]. To the best of our knowledge, the exact ABER analysis of Beaulieu-Xie
fading with BPSK and DPSK modulations is not reported in the literature yet. The
exact BER analysis of BPSK and binary DPSK is presented in this paper with an
aim to gain a better understanding about this fading channel.
The remaining paper is organized as follows: The description of PDF of fading
model is provided in Sect. 2. The exact analytical expressions for ABER of BPSK
and DPSK are derived in Sect. 3. Section 4 presents the discussion of analytical and
simulation results. Section 5 contains the concluding remarks of this paper.

2 System Model

A novel fading model has been proposed by Beaulieu and Jiandong [8] which models
multiple dominant specular components with some diffused power. The PDF of the

[Link]@[Link]
Analysis of Binary PSK Modulations … 15

SNR can be evaluated by applying a square transformation [13] in (1). The resulting
expression of PDF of SNR can be given by:
  γ (m−1)/ 2   
pγ (γ )  ηm+1 exp(−m K − ηγ ) Im−1 2 m K ηγ , (1)
mK

where η  L m−11 (−m K )/γ̄ , K  λ2 /, L αn (·) is the associated nth-order Laguerre
polynomial [11], and γ̄ is the average SNR, Im−1 (·) is the modified Bessel function
of the first kind and (m−1)th order, m is the shape factor,  defines the spread of
the model, and λ represents the height and location of the PDF. The Beaulieu-Xie
fading behaves as Rician fading, when the fading parameter ‘m’ is equal to 1 for any
value of ‘K’. Also, it can be reduced to Rayleigh fading when m  1, K  0.

3 Performance Evaluation

Average Bit Error Rate The exact closed-form expressions for BPSK and DPSK
modulation techniques over Beaulieu-Xie fading are derived in this section. The
ABER for various modulation techniques over a fading model can be derived by
integrating the conditional BER of any modulation techniques and the PDF of a fading
model. The ABER, Pe of a modulation technique can be computed mathematically
as:
∞
Pe  Pe (γ ) pγ (γ )dγ , (2)
0

3.1 ABER of BPSK Modulation Scheme Over

the Beaulieu-Xie Fading Channel

The conditional BER of BPSK over an AWGN channel [5] is given by:
1 √ 
Pe (γ )  erfc γ , (3)
2
To calculate the exact expressions for ABER of BPSK over the Beaulieu-Xie
fading channel, substitute the values of Pe (γ ) and pγ (γ ) from (3) and (1), respectively,
in (2). By using ([14], 3.381.9) in (3), we obtain the expression:

[Link]@[Link]
16 V. Kansal and S. Singh

∞ 
1
Pe  √ (0.5, γ ) (η)m+1 exp(−m K − ηγ )
2 π
0
 γ (m−1)/2   
× Im−1 2 m K ηγ dγ , (4)
mK
By expressing the modified Bessel function of the first kind and (m−1)th order
in the form of an infinite series, and by using identity ([14], 6.455.1), the resulting
expression for unconditional ABER of BPSK is given as:
∞
1 (m + g + 0.5) (m K )g
Pe  √ exp(−m K )  
2 (1 + η)π g0
g! (m + g + 1) 1 + 1 m+g
η

1
2 F1 1, m + g + 0.5, m + g + 1, 1
, (5)
1+ η

where 2 F1 (q; x; y; z) is a Gauss hypergeometric function.

3.2 ABER of DPSK Modulation Over the Beaulieu-Xie

Fading Channel

The ABER for the DPSK modulation can be derived by integrating the conditional
BER of the DPSK modulation with the PDF of the Beaulieu-Xie fading model as in
(2). The conditional BER of DPSK over an AWGN channel [10] is given by:
1
Pe (γ )  exp(−γ ), (6)
2
To obtain the ABER of the DPSK modulation scheme over the Beaulieu-Xie
fading channel, substitute the values of Pe (γ ) and pγ (γ ) from (6) and (1), respectively,
into (2), the expression is obtained as:
∞  γ (m−1)/ 2   
1
Pe  exp(−γ )(η)(m+1)/ 2 exp(−m K − ηγ ) Im−1 2 m K ηγ dγ ,
2 mK
0
(7)

Now by applying the identity ([14], 6.643.2), integral in (7) can be solved as:
m
exp(−m K ) η −m K η −m K η
Pe  exp M−m,0 , (8)
2 m K (1 + η) 2(1 + η) 1+η

[Link]@[Link]
Analysis of Binary PSK Modulations … 17

With
  s+1 the  help of ([14],
 9.220.2) and Mu,s (z) 
exp −z
2
z / 2 M s − u + 1 ; 1 + 2s; z , the expression in (8) can be simplified
2
as:

exp(−m K ) mKη
Pe   m M m, m, , (9)
2 1 + η1 1+η

where M(a, b, c) is a confluent hypergeometric function.

Further, by using the property M(a, a, c)  exp(c) in (9), the final expression for
the unconditional ABER of DPSK as:

ηm mK
Pe  m exp − . (10)
2(1 + η) 1+η

4 Numerical Results and Discussion

The final ABER expressions for BPSK and DPSK in (5) and (10) are calculated
analytically and curves are drawn for various combinations of fading parameter, ‘K’
and ‘m’, respectively. To validate the accuracy of analytical results, they have been
compared with simulation results. The simulation values at very low BER have been
eliminated since their accurate simulation requires a large amount of time.
Figures 1 and 2 show the results for ABER of BPSK over the Beaulieu-Xie fading
channel with constant total power of 6.98 dB. Figure 1 depicts the ABER of BPSK
against SNR per bit, γ̄ for a constant value of K  2 and various values of m. It
can be demonstrated that for a constant K, the performance of system improves as
m increases. It is observed from the results that m  5 gives a SNR advantage of
1 dB over m  3 and m  3 gives a SNR advantage of 13 dB over m  1 case to
achieve an ABER of 10−3 . With increase in the value of ‘m’, the SNR advantage
decreases because of decrease in the fading severity. The above observation comes
from system, because fading severity decreases with increase in m. The numerical
results are closely matched to the simulation results. Figure 2 shows similar results
as Fig. 1 but for constant m  3 and by varying K. On comparing the results for K
 0 with curves for K  2, it is observed that K  2 has a SNR advantage of 2 dB
over K  0 to achieve an ABER of 10−3 . This effect is because of the increase in
specular power with increase in K.
Figure 3 shows the ABER of DPSK for several values of K and m. The simulation
results obtained for the special case of Beaulieu-Xie fading as Rayleigh and Rician
fading are also provided in Fig. 3. The obtained results are in accordance with the
numerical results plotted by using final closed-form expressions of ABER of DPSK.
It is clear that for an ABER of 10−2 , the SNR improvement for curves shown in
Fig. 3, corresponding to m  1, K  2 over curves for m  1, K  0 is approximately

[Link]@[Link]
18 V. Kansal and S. Singh

Fig. 1 ABER of BPSK over

the Beaulieu-Xie fading for
fixed K  2 and various
values of m

Fig. 2 ABER of BPSK over

the Beaulieu-Xie fading for
fixed m  3 and various
values of K

3 dB. The results presented in this section are useful to analyze the effects of fading
parameters ‘K’ and ‘m’ on the ABER performance.

5 Conclusion

This paper presents exact analytical expressions for the ABER of BPSK and DPSK
that are obtained over the Beaulieu-Xie fading. In particular, the final expressions

[Link]@[Link]
Analysis of Binary PSK Modulations … 19

Fig. 3 ABER of DPSK over

the Beaulieu-Xie fading for
various combinations of K,
m with constant total power
of 6.02 dB

for BPSK and DPSK are in the form of infinite series and elementary functions,
respectively. Various scenarios for different combination of ‘K’ and ‘m’ have been
examined and discussed by which it is found that the ABER decreases as specular
power, ‘K’ and/or fading severity parameter, ‘m’ increases. The analytically obtained
expressions in this paper are verified through simulations.

References

1. Simon, M.K., Alouini, M.S.: A unified approach to the probability of error for noncoherent and
differentially coherent modulations over generalized fading channels. IEEE Trans. Commun.
46(12), 1625–1638 (1998)
2. Kam, P.Y., Zhong, C.: Tight bounds on the bit-error-probabilities of 2DPSK and 4DPSK in
nonselective Rician fading. IEEE Trans. Commun. 46(7), 860–862 (1998)
3. Yacoub, M.D.: The k-μ distribution and η-μ distribution. IEEE Antennas Propag. Mag. 49(1),
68–81 (2007)
4. Durgin, G.D., Rappaport, T.S., de Wolf, D.A.: New analytical models and probability density
functions for fading in wireless communications. IEEE Trans. Commun. 50(6), 1005–1015
(2002)
5. Oh, S.H., Li, K.H.: BER performance of BPSK receivers over two-wave with diffuse power
fading channels. IEEE Trans. Wirel. Commun. 4(4), 1448–1454 (2005)
6. Suraweera, H.A., Lee, W.S., Oh, S.H.: Performance analysis of QAM in a two-wave with
diffuse power fading environment. IEEE Commun. Lett. 12(2), 109–111 (2008)
7. Singh, S., Kansal, V.: Performance of M-ary PSK over TWDP fading channels. Int. J. Electron.
Lett. 4(4), 433–437 (2015). Taylor and Francis
8. Beauliu, N.C., Jiandong, X.: A novel fading model for channels with multiple dominant specular
components. IEEE Wirel. Commun. Lett. 4(1), 54–57 (2015)

[Link]@[Link]
20 V. Kansal and S. Singh

9. Guan, K., Zhong, D., Ai, B., Kuerner, T.: Propagation measurements and analysis for train
stations of high-speed railway at 930 MHz. IEEE Trans. Veh. Technol. 63(8), 3499–3516
(2014)
10. Olutayo, A., Ma, H., Cheng, J., Holzman, J.F.: Level crossing rate and average fade duration
for the Beaulieu-Xie fading model. IEEE Wirel. Commun. Lett. 6(3), 326–329 (2017)
11. Olutayo, A., Cheng, J., Holzman, J.F.: Asymptotically tight performance bounds for selection
diversity over Beaulieu-Xie fading channels with arbitrary correlation. In: IEEE International
Conference on Communication Proceedings, Paris, France, pp. 1–6 (2017)
12. Olutayo, A., Cheng, J., Holzman, J.F.: Asymptotically tight performance bounds for equal-
gain combining over a new correlated fading channel. In: IEEE 15th Canadian Workshop on
Information Theory Proceedings, Quebec, Canada, pp. 1–5 (2017)
13. Kansal, V., Singh, S.: Analysis of effective capacity over Beaulieu-Xie fading model. in Proc.
IEEE International Women in Engineering Conference on Electrical and Computer Engineer-
ing, Dehradun, India, pp. 207–210 (2017)
14. Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series and Products, 5th edn. Academic, San
Diego, CA (1994)

[Link]@[Link]
Approximate Bit Error Rate of DPSK
with Imperfect Phase Noise in TWDP
Fading

Veenu Kansal and Simranjit Singh

Abstract This study provides closed-form expressions for average bit error rate
(ABER) of differential phase shift keying (DPSK) with phase error in two-wave
with diffuse power (TWDP) fading. It is considered that the envelope of phase error
is Gaussian distributed. The effect of phase synchronization on wireless system
is studied for different values of TWDP fading parameters and phase error. The
analytical results are evaluated to study the impact of phase error on the system
performance. Also, the results are compared with the case of perfectly synchronized.

Keywords BER · DPSK · TWDP

1 Introduction

The performance of PSK with noisy phase reference has been a topic of interest
for many researchers since the 1960s. Initial work on modeling the phase error was
done by Tikhonov and Viterbi. Since then a lot of work has been done to study the
effect of phase error on different digital modulation schemes over various fading
models such as Rayleigh, Nakagami-m and Rician [1–7]. In all these papers, the
effects of imperfect carrier synchronization have been studied by assuming the phase
distribution either Gaussian or Tikhonov. However, phase synchronization errors can
have a significant effect on the receiver’s capability to make the correct decisions.
The problem gets more severe as data rates increase due to phase error. In 2002, a
fading model was proposed by Durgin et al. which could be used to model fading
scenarios having two specular paths in addition to the diffusely propagating power
[8]. This fading model is known as TWDP and can be used to characterize the fading
scenarios which are more severe than Rayleigh fading. The practical significance

V. Kansal · S. Singh (B)

Punjabi University, Patiala 147002, India
e-mail: simranjit@[Link]
V. Kansal
e-mail: veenukansal@[Link]

© Springer Nature Singapore Pte Ltd. 2019 21

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
22 V. Kansal and S. Singh

of this model comes from modern uses of wireless channels in applications such as
data transmission in dynamic environments such as mobile personal devices (such
as smartphones or smart/connected automobiles) and Internet service in aircraft that
have become pervasive in recent years.
The existing fading models such as Rayleigh and Rician fading are included in
TWDP fading as special cases. The error performance analysis of various digital
modulation schemes in TWDP fading is reported in the literature [9–12]. However,
till date no work is available on the analysis of modulation schemes with phase noise
error in TWDP fading model. In this paper, we derive the approximate closed-form
expression for average BEP of DPSK with noisy phase reference in TWDP fading
channel. In Sect. 2, the PDF of TWDP fading model is presented. In Sect. 3, the
PDF approach is used to derive the ABER of DPSK with noisy phase reference over
TWDP fading model. The numerical results are evaluated and discussed in Sect. 4.
Section 5 holds the conclusion of the paper.

2 TWDP Fading Model

The TWDP fading model is having two specular components with some Rayleigh
distributed diffuse power. The instantaneous signal-to-noise ratio
 (SNR) is defined
as γ  r 2 E b /N0 , and the average SNR is defined as γ̄  E r 2 NEb0 . The PDF of
SNR (γ ) for TWDP fading channel can be evaluated as [9]:

η 
L 1
    
pγ (γ )  ai exp −P2i− j − ηγ I0 2 P2i− j ηγ , (1)
2 i1 j0

   i−1 
where η  K + 1/γ̄ , P2i− j  K 1 + (−1) j  cos π 2L−1 ,  denotes the relative
strength of two specular components, K is defined as ratio of specular power to diffuse
power, and L is the order of PDF. The TWDP fading can be reduced to special cases
[9] of fading such as Rayleigh fading (K = 0) and Rician fading (K > 0,   0).

3 Performance Analysis

The conditional BER of DPSK with noisy phase reference over AWGN channel can
be expressed as [13]:
1  
P(e|γ , φ)  exp −γ cos2 φ , (2)
2
where φ represents the phase error. It is considered that φ is Gaussian distributed
with zero mean and 2σ 2 variance, i.e., C (0, 2σ 2 ). The PDF of φ can be given by:

[Link]@[Link]
Approximate Bit Error Rate of DPSK with Imperfect Phase Noise … 23

1 −φ 2
pφ (φ)  √ exp , (3)
2π σ 2 2σ 2

The ABER of a modulation scheme with phase noise error over a fading channel
can be obtained by performing a twofold integration, first over the PDF of φ and
second over the PDF of fading model [2],
∞ π

Pe  P(e|γ , φ) pφ (φ) pγ (γ )dφ dγ , (4)

0 −π

To derive the ABER of DPSK with Gaussian distributed phase error over TWDP
fading model, first solve the inner integral of (4) as:
π

P(e|γ )  P(e|γ , φ) pφ (φ)dφ, (5)

−π

Substitute P(e|γ , φ) and pφ (φ) from (2) and (3), respectively, into (5), and we
get:
π
1   1 −φ 2
P(e|γ )  exp −γ cos2 φ × √ exp dφ, (6)
2 2π σ 2 2σ 2
−π

By using the small angle approximation, cos(φ) ∼

 1− (φ)
2

2
, (6) can be rewritten
as:
π
exp(−γ ) 1
P(e|γ )  √ exp −φ 2 −γ dφ, (7)
2 2π σ 2 2σ 2
−π

π   √
Now by employing the general formula, i.e., −π exp −x 2 dx  π , the condi-
tional BER of DPSK with Gaussian phase error over AWGN channel can be expressed
as:
exp(−γ )
P(e|γ )   , (8)
2 1 − 2σ 2 γ

Now solving the outer integral of (4) over PDF of TWDP fading, we get
∞

Pe  P(e|γ ) pγ (γ )dγ , (9)

0

By substituting P(e|γ ) and pγ (γ ) from (8) and (1) into (9), we get:

[Link]@[Link]
24 V. Kansal and S. Singh

∞
exp(−γ ) η 
L 1
    
Pe   × ai exp −P2i− j − ηγ I0 2 P2i− j ηγ dγ ,
2 1 − 2σ 2 γ 2 i1 j0
0
(10)

Put σφ2  (Cγ )−1 in (10), and rearrange (10) as:

∞
η 
L 1
  1   
Pe  ai exp −P2i− j  exp(−γ (1 + η))I0 2 P2i− j ηγ dγ ,
4 i1 j0 1− 2
C 0
(11)

By using the formula given in the appendix, the final ABER expression is calcu-
lated as:

η    −1/ 2
L 1
2 1 P2i− j η
Pe  ai exp −P2i− j 1 − × M 1; 1; , C > 2.
4 i1 j0 C 1+η 1+η
(12)

where M(a; b; c) is confluent hypergeometric function.

4 Numerical Results and Discussion

The numerical results are provided for the ABER of partially coherent differential
PSK over TWDP fading model. By using (12), the results are plotted in Figs. 1, 2, and
3. As mentioned in (12), these results are valid only for value of C greater than 2. The
results are taken for different parameters such as C, , and K. The value of imperfect
synchronization factor (C) is taken to be 3, 10, and ∞. When C  ∞, the DPSK
modulation is considered to be having no phase error. For low values of C, the phase
error is more, and as expected, the degradation is severe. As value of C increases, the
degradation becomes less severe and closer to the perfectly synchronization case.
In Fig. 1, results are plotted for a combination of different values of C and  and
for fixed value of K. It is observed that the performance of system deteriorates as 
approaches 1.
Similarly, in Fig. 2, results are plotted for fixed value of  with different values of
C and K. The results demonstrated that the ABER decreases as the value of specular
power (K) increases.
Figure 3 shows the results for ABER versus C for fixed values of , K, and SNR
(in dB). The plot depicts that the performance of system improves as the value of C
increases; i.e., as the phase error decreases, the ABER also decreases.

[Link]@[Link]
Approximate Bit Error Rate of DPSK with Imperfect Phase Noise … 25

Fig. 1 ABER versus SNR 0

10
for DPSK with phase error
-1
with fixed K and varying C 10
and 
-2
10 K=10, Δ=1

-3
10

K=10, Δ= 0.5

BER
-4
10
-5
10
Imperfect Sync (C=3)
-6
10 Imperfect Sync (C=10)
Perfect Sync K=10, Δ= 0
-7
10

-8
10
0 5 10 15 20 25 30 35
Average SNR(dB)

Fig. 2 ABER versus SNR 0

10
for DPSK with phase error
with fixed  and varying C 10
-1

and K
-2 K=2,Δ=0.5
10
-3
10
BER

-4
10
-5
10 K=5,Δ=0.5
Imperfect Sync (C=3)
-6
10 Imperfect Sync (C=10) K=8,Δ=0.5
Perfect Sync
-7
10
-8
10
0 5 10 15 20 25 30 35
Average SNR(dB)

5 Conclusion

This paper analyzed the performance of ABER of partially coherent DPSK in TWDP
fading channels. The closed-form expressions are derived for ABER which is a
convenient tool for error performance analysis. The final expression of unconditional
ABER of DPSK along with Gaussian phase error over TWDP fading is valid only for
value of C greater than 2. The analysis depicted that if the imperfect synchronization
factor, ‘C’ increases; the system performance is improved and is closed to perfectly
synchronization case.

[Link]@[Link]
26 V. Kansal and S. Singh

Fig. 3 ABER versus C (in

-2
dB) for DPSK modulation 10
for fixed K, SNR, and 

BER
-3
10

SNR=10 dB,K=10,Δ=0.5
SNR=15 dB,K=10,Δ=0.5
SNR=20 dB,K=10,Δ=0.5
SNR=25 dB,K=10,Δ=0.5
SNR=30 dB,K=10,Δ=0.5
-4
10
3 4 5 6 7 8 9 10
C (in dB)

Appendix

The expression in (11) can be further solved by using identity given in ([14], [Link])
as:
∞
   1 β (v + 1) β
exp(−αx)I2v 2 βx dx  √ exp M−1/2,v
αβ 2α (2v + 1) α
0

Now, (11) can be simplified as:

η    −1/2
L 1
2 P2i− j η
Pe  ai exp −P2i− j 1 − exp
4 i1 j0 C 2(1 + η)
1 P2i− j η
× M−1/2,0 (13)
(1 + η)P2i− j η 1+η
  s+1/2  
Use Mk,s (z)  exp −z2
z M s − k + 21 ; 1 + 2s; z [14] in (13), and the final
equation can be obtained as given in (12).

References

1. Viterbi, A.J.: Phase-locked loop dynamics in the presence of noise by Fokker-Planck tech-
niques. Proc. IEEE 51, 1737–1753 (1963)
2. Simon, M.K., Alouini, M.S.: Simplified noisy reference loss evaluation for digital communi-
cation in the presence of slow fading and carrier phase error. IEEE Trans. Veh. Technol. 50(2),
480–486 (2001)

[Link]@[Link]
Approximate Bit Error Rate of DPSK with Imperfect Phase Noise … 27

3. Lo, C.M., Lam, W.H.: Error probability of binary phase shift keying in Nakagami-m fading
channel with phase noise. Electron. Lett. 36(21), 1773–1774 (2000)
4. Lindsey, W.C.: Phase-shift-keyed signal detection with noisy reference signals. IEEE Trans.
Aerosp. Electron. Syst. 2, 393–401 (1966)
5. Prabhu, V.K.: PSK performance with imperfect carrier phase recovery. IEEE Trans. Aerosp.
Electron. Syst. 12, 275–285 (1976)
6. Chandra, A., Patra, A., Bose, C.: Performance analysis of PSK systems with phase error in
fading channels: a survey. Phys. Commun. 4, 63–82 (2011)
7. Smadi, M.A., Aljazar, S.O., Ghaeb, J.A.: Simplified bit error rate evaluation of Nakagami-m
PSK systems with phase error recovery. Wirel. Commun. Mob. Comput. 12, 248–256 (2012)
8. Durgin, G.D., Rappaport, T.S., de Wolf, D.A.: New analytical models and probability density
functions for fading in wireless communications. IEEE Trans. Commun. 50(6), 1005–1015
(2002)
9. Subadar, R., Singh, A.D.: Performance of SC receiver over TWDP fading channels. IEEE
Wirel. Commun. Lett. 2(3), 267–270 (2013)
10. Singh, S., Kansal, V.: Performance of M-ary PSK over TWDP fading channels. Int. J. Electron.
Lett. 4(4), 433–437 (2015)
11. Das, P., Subadar, R.: Performance of M-EGC receiver over TWDP fading channels. IET Com-
mun. 11(12), 1853–1856 (2017)
12. Singh, S., Sharma, S.: Performance analysis of spectrum sensing techniques over TWDP fading
channels for CR based IoTs. Int. J. Electron. Commun. 80, 210–217 (2017)
13. Lindsey, W.C., Simon, M.K.: Telecommunication Systems Engineering. Prentice-Hall, Engle-
wood Cliffs, NJ (1973)
14. Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series, and Products, 7th edn. Academic
Press (2000)

[Link]@[Link]
Analysis of Workloads for Cloud
Infrastructure Capacity Planning

Eva Patel and Dharmender Singh Kushwaha

Abstract Workload analysis and characterization are the first steps toward effec-
tive cloud infrastructure capacity planning. Identifying workload patterns based on
resource utilization not only enables informed decisions about mapping of current
request to available capacity, but also serves as a meaningful indicator for future
resource requirements. Of paramount concern is the optimal utilization of data cen-
ter server capacity, i.e., the CPU, I/O, and memory. The compute capacity of modern
servers can be further harnessed by optimal utilization of individual CPU cores. A
precise CPU core-level usage monitoring and provisioning can lead to cumulative
benefits of optimal CPU utilization, efficient VM placement, reduced VM migra-
tions, and energy efficiency through lower power consumption. In this paper, we
make a preliminary analysis of usage patterns of CPU cores in the case of CPU-
and memory-intensive workloads on an experimental cloud setup in our laboratory.
The aim is to make a comparative analysis of the utilization of individual CPU
cores with that of aggregated CPU usage to explore the feasibility of incorporat-
ing a fine-grained usage detail for resource scheduling and VM provisioning. Initial
experiments reveal observable differences between the utilization of individual CPU
cores and that reported by aggregate CPU usage. Usage difference ranges from 1 to
29% below and between 4 and 20% above the aggregate. Incorporating such finer
details can leverage the vast compute capacity of multicore servers and effective
power usage.

Keywords Workload analysis · Workload characterization · Capacity planning

Server capacity planning · Workload intensity · Job inter-arrival patterns
Multicore scheduling

E. Patel (B) · D. S. Kushwaha

Department of Computer Science and Engineering, MNNIT Allahabad, Allahabad,
UP 211004, India
e-mail: evapatel08@[Link]
D. S. Kushwaha
e-mail: dsk@[Link]

© Springer Nature Singapore Pte Ltd. 2019 29

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
30 E. Patel and D. S. Kushwaha

1 Introduction

Scalability, agility, guaranteed quality of service, the illusion of infinite computa-

tional resources, and access to high-end computing infrastructure on a pay-as-per-
usage basis, without the painstaking exercise of setting up a private IT infrastructure,
have led to the relentless adoption of cloud computing solutions for a range of
computing requirements—business intelligence, engineering design, scientific com-
puting, social networking, and content delivery. These services are hosted on data
centers that house hundreds of thousands of servers and supporting infrastructure.
For instance, Amazon, which made the IT infrastructure available via the Internet in
2006, a technology termed cloud computing [1], has around 450,000 servers spread
across seven different data centers around the world [2] that provide the infrastructure
services—compute, storage, and networking. Google data center caters to the diverse
requests of its cloud users through 900,000 plus servers hosted in data centers in 16
different geographical locations [2]. The envisioned growth in computing and storage
requirements, diversity of services, and emergence of innovative server technologies
that also entail growing power consumption and increasing carbon footprints calls
for adoption of capacity planning of cloud data centers as a continual process.
Data center capacity planning aims to determine a system configuration for given
service requests that comply with the service-level agreement (SLA) [3] with min-
imal over- or under-provisioning of resources. The first step to effective capacity
planning is the analysis of resource usage patterns of cloud workloads to identify
associated challenges and their implications on resource provisioning and overall sys-
tem performance. A vast body of literature records design of scheduling algorithms
with CPU utilization as the primary performance metric [4–10]. With several CPU
cores crammed into a single socket, and multisocket servers with massive compute
capacity, the use of individual core usage as decision parameter for VM scheduling
is one possible approach to effective utilization of server compute capacity.
Driven by the observation that aggregate CPU usage is not always reflective of
the usage of individual CPU cores, we conduct a preliminary analysis of utilization
of CPU cores in case of CPU-intensive and memory-intensive workloads. Through
this study, the aim is to achieve following objectives:
1. Effective utilization of compute capacity of all the cores of a multicore server by
considering individual CPU core capacity rather than the aggregate CPU usage.
2. Identify and segregate cores that are underutilized, to schedule to incoming work-
load thereby reducing energy wastage.
3. Mitigate situations of overloaded CPU cores to maintain desired performance
and sustainable energy consumption levels.
The rest of this paper is organized as follows. Section 2 gives a brief overview of
related work on workload analysis and characterization. Section 3 discusses some
background concepts and findings from past research that reveal suboptimal use
of CPU cores. Experimental setup and results are discussed in Sect. 4. Section 5
concludes with directions for future work.

[Link]@[Link]
Analysis of Workloads for Cloud Infrastructure … 31

2 Related Work

Resource requirements of applications deployed on the cloud depend on their pro-

cessing needs. Consequently, each workload has different challenges which guide
the analysis and characterization process.
Based on behavioral characteristics, Mahambre et al. [4] identify five cloud
workload patterns, i.e., periodicity, threshold, relationship, variability, and image
similarity, for capacity planning decisions such as migration, re-provisioning,
load balancing, and initial placement of the workloads. Moreno et al. [5] present a
comprehensive analysis of workload characteristics to study heterogeneity due to user
behavior and task resource usage. Their study shows that users are responsible the
most for introducing heterogeneity in the cloud as compared to task diversity which
is a consequence of diverse service requests due to the illusion of infinite resources
to the users. Authors in [6] characterize the interactive behavior of Web applications
in terms of number of instructions executed per second, and CPU, memory and disk
utilization. Peng et al. [7] develop a classification scheme based on computational
needs of the workloads for compute-intensive, I/O-intensive, and network-intensive
applications. Zhang et al. [8] consider both the heterogeneity of machine hardware
and workload diversity for dynamic capacity provisioning. Cloud infrastructure is
heterogeneous due to the presence of machines from multiple generations, heteroge-
neous processor architectures and speed, and different memory and disk capacities.
Singh and Chana [9] characterize workloads based on QoS requirements for dif-
ferent workload types such as scientific computing, online transaction processing,
performance testing, and storage and backup services. Authors in [10] classify cloud
workloads based on their functional characteristics, into six categories—scientific
processing, Big Data application, OLTP, caching, streaming, and Web serving. The
aim is to customize the resource requirements with the actual utilization of a specific
workload for effective cloud monitoring.
From our study of the above work and the conclusions drawn by Lozi et al. [11], we
infer that compute capacity of data center servers can be harnessed more effectively
by considering per-core CPU usage rather than the aggregated utilization.

3 Related Concepts

3.1 Capacity Planning

Capacity planning is the process of determining computing infrastructure (hardware,

software, and connection interface) and the floor area to house these components to
cater to services for a future time period [12]. Data center capacity planning involves
server capacity planning as well as network capacity planning. This work focusses
on server capacity planning in which an IT department determines the amount of
server hardware resources required to provide the desired levels of service for a given

[Link]@[Link]
32 E. Patel and D. S. Kushwaha

workload mix for the least cost [12]. The central objective is to provision resources
in compliance with SLA.
The notion of capacity planning as perceived by cloud actors depends on their
corresponding role. National Institute of Standards and Technology (NIST) [13]
defines five cloud actors: consumer, provider, auditor, broker, and carrier. Of interest
for server capacity planning are the roles of cloud consumer and cloud provider.
The main aim of a cloud provider is to provision resources in a way that maximizes
returns on investment, whereas a cloud user is concerned with quality of service
that would justify rental cost. These benefits can be fully realized through optimal
utilization of compute capacity of server.

3.2 Workload Analysis and Characterization

Workload analysis involves a detailed investigation of workload features of interest, to

identify their behavioral characteristics, intra- and inter-dependencies, and impact on
system performance. Workload characterization is the process of precisely describing
the system’s global workload in terms of its main components. Workload analysis
measures cloud services along different dimensions including infrastructure capacity
planning, energy efficiency, performance, reliability, and security. Two fundamental
issues with workload analysis specific to infrastructure capacity planning [3] are:
1. Workload arrival pattern
2. Workload intensity
Workload arrival pattern measures the rate at which jobs arrive for service. Work-
load intensity refers to the amount of work done over a unit of time. Job arrival
patterns are highly unpredictable and exhibit different behavioral patterns—diur-
nal, seasonal, and flash crowd [14]. Additionally, workloads may have either steady
resource demands or may manifest temporal patterns such as periodic, bursting,
growing, and on-and-off. An effective server capacity planning scheme should ensure
resource availability on demand and provisioning for required time duration.
Job Arrival Patterns as Non-homogeneous Poisson Process. Non-
homogeneous Poisson process is one way to model real-world job arrival patterns.
Formally, given the occurrence of events at a constant rate λ, over a period T , the job
inter-arrival time can be modeled as a Poisson process with exponential distribution
which is governed by Eq. (1) below:

(λt)k −λt
P(N (t)  k)  e , k  0, 1, 2, . . . (1)
k!
where the random variable N(t) denotes the number of events that occur in t time units,
P is the probability of occurrence of k events, and λ, known as the rate parameter,
measures the number of events that occur per unit of time. Variability in job arrival

[Link]@[Link]
Analysis of Workloads for Cloud Infrastructure … 33

can be captured using non-homogeneous Poisson process in which the rate parameter
λ changes with time.

3.3 Multicore Scheduling

Multicore processors are the consequence of unsustainable power consumption and

heat dissipation resulting from increased CPU clock frequency for performance gains
[15]. The layout of a typical quad-core processor chip is shown in Fig. 1. Each core
has a separate Level 1 (L1) cache for data and instructions and a Level 2 (L2) cache
that holds both instruction and data, which are private. L2 cache is either united
or distributed. L2 cache is usually physically distributed with a united Level 3 (L3)
cache. Some multicores have an off-chip Level 4 (L4) cache. Two central issues when
scheduling on multicore servers are contentions due to access to shared resources
and efficient utilization of individual cores. The vast processing capability of modern
multicore processors can be tapped, by keeping all the cores busy to the maximum.
This has been achieved by adapting the operating system scheduler for single-core
systems to that of scheduling large number of cores.
The Linux Scheduler. Completely Fair Scheduler (CFS), introduced in Octo-
ber 2007, is the default process scheduler in Linux since kernel version 2.6.23 and
supersedes the O(n) scheduler and O(1) scheduler. O(n) scheduler was used in ker-
nel versions from 2.4 to 2.6 and was replaced by O(1) scheduler in 2001, due to
non-scalability issues. The O(1) scheduler maintains a constant scheduling time
irrespective of the number of jobs in the system and uses average sleep time of a
process to distinguish between interactive and non-interactive jobs. However, this

Fig. 1 A multicore chip

[Link]@[Link]
34 E. Patel and D. S. Kushwaha

heuristic makes the scheduler complex with incidental cases of non-interactive jobs
being identified as interactive jobs. CFS uses the concept of execution time and alle-
viates such miscalculations by organizing the jobs as red–black tree. The job with
lowest execution time is the leftmost node of the tree and is picked up by CFS when
invoked to schedule a process. CFS is a work-conserving scheduler which implies
that the cores are prevented from being idle if there are processes ready for execution.
However, experimental studies conducted by Lozi et al. [11] reveal that the Linux
scheduler leaves cores idle while threads are waiting in run queues. This is revealed
by performance degradation.

4 Results and Discussion

To make a comparative analysis of individual core usage with that of aggregate CPU
utilization, we monitor per-core CPU utilization for workload types given below:
• A CPU-intensive workload is one that involves massive computation such as in
financial modeling and scientific applications.
• Memory-intensive workloads involve high memory activity such as pushing of
large volumes of data into and out of memory and frequent and long durations
of memory read and write operations. Such workloads need not necessarily bloat
memory usage and involve minimal CPU processing.

4.1 Experimental Cloud Setup

Usage of CPU cores is collected on a cloud setup in the laboratory. The experimental
cloud consists of a controller node, a Network File System (NFS) server node, and
two compute nodes. The configuration and software environment details are given
in Table 1.

Table 1 Node configuration

Node Configuration Environment
Server and compute Intel core i7 4790 3.60 GHz Linux Kernel Version
quad-core CPU, 16-GB 4.4.0-116-generic,
1600 MHz DDR3 RAM QEMU-KVM Version 2.5.0
Controller Intel core i7 4770 3.40 GHz Linux Kernel Version
quad-core CPU, 16-GB 4.4.0-116-generic
1600 MHz DDR3 RAM

[Link]@[Link]
Analysis of Workloads for Cloud Infrastructure … 35

4.2 Design of Experiments

Applications used to generate different workload types are listed in Table 2, and dif-
ferent test cases are given in Table 3. Experiments are conducted on VMs running on
a single host that uses full virtualization with Kernel-based Virtual Machine (KVM).
The core virtualization infrastructure is implemented as a loadable kernel module
([Link]). The modules, [Link] and [Link], correspond to processors
specific to Intel and AMD, respectively. Virtualization with KVM leverages basic
mechanisms of CPU scheduling, memory management, and I/O access, provided by
the Linux kernel.
To emulate real-world service request scenario, we generate variable job arrival
times using a non-homogeneous Poisson process. Workloads are executed through a
Bash script according to the generated job arrival times. Resource utilization log is
collected for 1 hour at 3 seconds interval using System Activity Report (sar) system
monitoring tool for both the guest and the host. Usage statistics extracted from the
collected log are analyzed using R programming environment.
Baseline Resource Usage. To obtain CPU utilization for running basic system
services and virtualization overheads, we run two VMs with no other workload on
the guest or the host. A plot of the aggregated CPU usage is shown in Fig. 2, and
usage statistics are given in Table 4. Spikes in the graph indicate some high CPU
activity for a short duration. Mean value indicates that only 1–3% of the core usage
accounts for running operating system as well as virtualization software.
CPU-Intensive Workload on Single VM. Per-CPU core usage of host and guest
machines when running CPU-intensive workload on a single machine is depicted in
Fig. 3. Summary statistics for host and guest are given in Tables 5 and 6, respectively.

Table 2 Workload generation

Workload type Real-world application
CPU intensive Prime factorization, Fibonacci series
generation, factorial calculation
Memory intensive Swap two integers, array copy

Table 3 Test cases for usage monitoring

Test case Measurement objective
Baseline resource usage Overhead due to basic system services and
virtualization
CPU-intensive workload on single VM Individual CPU core usage
Memory-intensive workload on single VM Individual CPU core usage
CPU-intensive workload on one VM, CPU core usage for workloads with
memory-intensive workload on second VM complementary computational requirements
CPU-intensive workloads on three VMs CPU core usage for workloads having similar
computational requirements

[Link]@[Link]
36 E. Patel and D. S. Kushwaha

Fig. 2 Baseline aggregate

host usage

Table 4 Summary statistics of baseline host CPU core usage

CPU usage Min. Median Mean Max. Standard
deviation
Aggregate 0.15 2.05 2.15 48.34 2.53
Core 0 0.00 2.40 2.35 39.28 2.55
Core 1 0.00 2.20 1.99 64.87 3.13
Core 2 0.00 2.40 1.80 80.49 3.91
Core 3 0.00 1.00 2.24 44.42 2.51

Table 5 Summary statistics for host CPU core usage with CPU-intensive workload
CPU usage Min. Median Mean Max. Standard
deviation
Aggregate 1.96 75.48 70.60 80.53 13.01
Core 0 0.00 04.00 41.57 100.00 44.97
Core 1 0.34 100.00 80.73 100.00 33.48
Core 2 0.00 100.00 90.54 100.00 20.21
Core 3 0.34 100.00 68.12 100.00 41.44

Mean utilization of the cores lies between 41 and 90%, whereas aggregate usage is
70% with core 2 being highest utilized and core 0 being the least utilized than what
is given by the aggregate usage. High values for standard deviation show the widely
varying usage levels of individual cores. Core usage by the VM as given in Table 8
reveals near equal usage of all the vCPUs for executing the workload.

[Link]@[Link]
Analysis of Workloads for Cloud Infrastructure … 37

(a) (b)

(c)
Fig. 3 CPU utilization for CPU-intensive workload. a Host aggregate, b per-core host, and c per-
core guest

Memory-Intensive Workloads. Figure 4 shows a plot of per-core CPU usage

for host and guest, running memory-intensive workloads on a single VM. The plots
reveal that CPU activity is low with memory-intensive workloads as compared to
CPU-intensive workloads, as the case should be. The cores either are mostly idle or
have very low utilization as the median values in Table 7 reveal.
In this case, utilization of core 1 is 2% lesser than the aggregate and utilization of
core 2 is 5% higher as given in Table 6.
Memory- and CPU-Intensive Workloads. The usage plots for this case are
shown in Fig. 5, and Table 8 contains the usage statistics. While mean aggregated CPU

[Link]@[Link]
38 E. Patel and D. S. Kushwaha

Table 6 Summary statistics for guest CPU core usage with CPU-intensive workload
CPU usage Min. Median Mean Max. Standard
deviation
Aggregate 2.11 100.00 92.11 100.00 17.28
Core 0 3.01 100.00 92.22 100.00 17.06
Core 1 1.67 100.00 92.09 100.00 17.32
Core 2 1.67 100.00 92.11 100.00 17.27
Core 3 1.33 100.00 92.08 100.00 17.34

(a) (b)

Fig. 4 Memory-intensive workloads a Aggregate CPU usage, b per-CPU core host usage, and c
per-core guest usage

usage is close to 21%, core 0 and core 1 have higher mean usage. Such a usage detail
can be detrimental when scheduling pCPUs to vCPUs in a virtualized environment.
Also, since computational requirements of CPU- and memory-intensive workloads

[Link]@[Link]
Analysis of Workloads for Cloud Infrastructure … 39

Table 7 Summary statistics of host and guest CPU core usage for memory-intensive workload
CPU usage Min. Median Mean Max. Standard
deviation
Host usage
Aggregate 1.40 11.59 11.84 28.85 03.76
Core 0 0.00 1.97 13.135 96.20 20.81
Core 1 0.00 0.60 08.389 98.00 21.59
Core 2 0.00 1.60 15.485 100.00 22.10
Core 3 0.00 0.200 10.446 77.110 08.22
VM usage
Aggregate 5.51 45.05 45.76 100.00 14.74
Core 0 4.19 45.80 45.92 100.00 16.33
Core 1 1.60 45.20 45.59 100.00 16.23

(a) (b)
Fig. 5 Compute- and memory-intensive workloads: a host aggregate usage and b host per-core
utilization

are complementary, the workloads do not contend for compute capacity of cores as
can be seen from Fig. 6.
Compute-Intensive Workloads on three VMs. To study CPU core usage with
workloads having similar computational requirements, we run two VMs, each with
compute-intensive workloads, on a single host. Figure 7 contains the usage plots,
and utilization values are given in Table 9.
From Table 9, we make similar observations as in the previous cases that in the
case of multiple VMs running workloads in terms of similar resource requirements,
core utilization is not uniform. Mean utilization of Core 1 is higher than that reported
by the aggregate usage while that of Core 3 is below the aggregate value.

[Link]@[Link]
40 E. Patel and D. S. Kushwaha

Table 8 Summary statistics of host CPU Core usage for memory- and CPU-intensive workloads
CPU usage Min. Median Mean Max. Standard
deviation
Aggregate 3.23 20.62 20.95 53.79 6.23
Core 0 0.00 17.14 24.63 100.00 22.34
Core 1 0.00 18.30 25.02 100.00 24.93
Core 2 0.00 25.27 19.95 100.00 24.43
Core 3 0.00 13.21 19.17 100.00 21.94

(a) (b)

Fig. 6 Guest per-core utilization. a Memory-intensive workload and b CPU-intensive workload

Table 9 Summary statistics of host with CPU-intensive workloads on three VMs

CPU usage Min. Median Mean Max. Standard
deviation
Aggregate 53.36 73.85 75.08 100.00 10.51
Core 0 07.12 94.67 74.70 100.00 25.18
Core 1 07.38 72.65 82.02 100.00 26.16
Core 2 08.45 96.67 75.52 100.00 23.61
Core 3 10.03 94.64 67.95 100.00 22.37

Table 10 summarizes the percentage difference in utilization of individual cores

from that of the aggregated usage. With compute-intensive workloads on single VM
(Case II), a core usage of 29% below and 20% above the aggregate usage is recorded.
When running only memory-intensive workloads (Case III), individual core usage is
3% above and below the aggregate usage. Similarly, in case of CPU- and memory-
intensive workloads (Case IV) and CPU-intensive workloads on multiple VMs (Case
V), the core usage varies between 1% below and 4% above the aggregate usage
for Case IV and 7% above and below the aggregate usage in Case V. Knowledge

[Link]@[Link]
Analysis of Workloads for Cloud Infrastructure … 41

(a) (b)

Fig. 7 CPU-intensive workloads on three VMs. a Host aggregate usage and b per-core host usage

Table 10 Usage comparison of CPU cores with aggregate usage

Aggregate Core 0 Core 1 Core 2 Core 3
Case II 70.60 41.57 80.73 90.54 68.12
Case III 11.84 13.14 8.40 15.48 10.45
Case IV 20.95 24.63 25.02 19.95 19.17
Case V 75.08 74.70 82.02 75.52 67.95

of individual core usage can be instrumental in effective scheduling and resource

utilization.

5 Conclusion and Future Scope

In this work, we conducted an empirical study of utilization of CPU cores of a

multicore server to analyze core-level CPU usage to verify our observations that
aggregated CPU usage is not reflective of utilization at the core level. We collected
CPU usage for five test cases—baseline usage and CPU core usage in case of CPU-
intensive workload on single VM, memory-intensive workload on single VM, CPU-
intensive workload on one VM and memory-intensive workload on the other, and
CPU-intensive workloads on two VMs. We observe that basic system services and
virtualization software incur some overhead in terms of CPU usage, as the case should
be. The remaining four cases clearly bring out differences between aggregated CPU
usage and individual cores which ranges from 1 to 29% below and between 4 and 20%
above the aggregate usage. A precise knowledge of individual core usage will help in

[Link]@[Link]
42 E. Patel and D. S. Kushwaha

identifying cores that are underutilized and making informed scheduling decisions.
Additionally, it will also be a vital input for reducing the load of overloaded cores
and devising policies to prevent aggravating of the overload situation.
As future research, we would study core usage for I/O- and network-intensive
workloads. A model based on the usage measurements of the four workload type-
s—CPU, memory, I/O, and network intensive—would then be developed for work-
load characterization and prediction at the level of individual CPU cores. We would
delve deeper into the varied usage levels of the CPU cores to study their criticality
and impact on overall performance on different workload types. Scheduling policies
that consider usage of CPU cores and account for inferences derived from the investi-
gations would be developed, the impact of such a fine-grained scheduling on overall
system performance would be analyzed, and approaches to mitigate performance
loss would be investigated.

References

1. Varia, J., Mathew, S.: Overview of Amazon Web Services. Amaz. Web Serv. (2014)
2. [Link]
Last Accessed 2018/06/10
3. Feitelson, D.G.: Workload modeling for computer systems performance evaluation. Cambridge
University Press, Cambridge (2015)
4. Mahambre, S., Kulkarni, P., Bellur, U., Chafle, G., Deshpande, D.: Workload characterization
for capacity planning and performance management in IaaS cloud. In: 2012 IEEE International
Conference on Cloud Computing in Emerging Markets (CCEM), pp. 1–7 (2012)
5. Moreno, I.S., Garraghan, P., Townend, P., Xu, J.: Analysis, modeling and simulation of work-
load patterns in a large-scale utility cloud. IEEE Trans. Cloud Comput. 2, 208–221 (2014)
6. Magalhães, D., Calheiros, R.N., Buyya, R., Gomes, D.G.: Workload modeling for resource
usage analysis and simulation in cloud computing. Comput. Electr. Eng. 47, 69–81 (2015)
7. Peng, J., Chen, J., Zhi, X., Qiu, M., Xie, X.: Research on application classification method in
cloud computing environment. J. Supercomput. 73, 3488–3507 (2017)
8. Zhang, Q., Zhani, M.F., Boutaba, R., Hellerstein, J.L.: Dynamic heterogeneity-aware resource
provisioning in the cloud. IEEE Trans. Cloud Comput. 2, 14–28 (2014)
9. Singh, S., Chana, I.: QRSF: QoS-aware resource scheduling framework in cloud computing.
J. Supercomput. 71, 241–292 (2015)
10. Orzechowski, P., Proficz, J., Krawczyk, H., Szymański, J.: Categorization of cloud workload
types with clustering. In: Proceedings of the International Conference on Signal, Networks,
Computing, and Systems, pp. 303–313 (2017)
11. Lozi, J.P., Lepers, B., Funston, J., Gaud, F., Quéma, V., Fedorova, A.: The Linux scheduler: a
decade of wasted cores. In: Proceedings of the Eleventh European Conference on Computer
Systems, p. 1 (2016)
12. Capacity Planning—Discipline for Data center decisions, [Link]
6814/2049/9759/[Link]. Last Accessed 2018/06/10
13. Sokol, A.W., Hogan, M.D.: NIST Cloud Computing Standards Roadmap (2013)
14. Calzarossa, M.C., Della Vedova, M.L., Massari, L., Petcu, D., Tabash, M.I.M., Tessera, D.:
Workloads in the clouds. In: Principles of Performance and Reliability Modeling and Evalua-
tion, pp. 525–550. Springer, Berlin (2016)
15. Vajda, A.: Programming many-core chips. Springer Science & Business Media, Berlin (2011)

[Link]@[Link]
Trust-Based Scheme for Location
Finding in VANETs Using
Trustworthiness of Node
Sanjoy Das, Indrani Das, Rishi Pal Singh, Prashant Johri
and Ashwini Kumar

Abstract The purpose of vehicular ad hoc networks is to improve the traffic condi-
tions and mitigate the problems occurred during ply. Any wrong decision taken by
the vehicle may lead to hazardous problems and may create various serious issues.
Thus, they need to be handled very carefully. Here, we have used trust-based method-
ology to find location with the help of trustworthiness of the node. In our proposed
trust-based scheme, firstly calculation of trustworthiness of each vehicle exists in
the network. After using trustworthiness, the location information of a vehicle is
calculated. The purpose of calculating trustworthy is that higher the value of trust-
worthy of a node, higher the chances of providing the most accurate response. We
have implemented the scheme in Java. The result shows that proposed scheme solves
problem in the previously existing scheme.

Keywords Trust · Trustworthiness · Location · RSU

S. Das (B)
Indira Gandhi National Tribal University, Regional Campus Manipur, Manipur, India
e-mail: [Link]@[Link]
I. Das
Assam University, Silchar, India
e-mail: indranidas2000@[Link]
R. P. Singh
Guru Jambheshwar University of Science and Technology, Hisar, India
e-mail: pal_rishi@[Link]
P. Johri · A. Kumar
Galgotias University, Greater Noida, India
e-mail: [Link]@[Link]
A. Kumar
e-mail: ashwinipaul@[Link]

© Springer Nature Singapore Pte Ltd. 2019 43

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
44 S. Das et al.

1 Introduction

The vehicular ad hoc networks (VANETs) is used to provide information that is

critical in the decision-making process of vehicles at the time of hazardous situa-
tions like congestion, earthquake, floods, and jams to the users. This network helps
minimize road accidents, fuel wastage, time, and provide much safer, economical,
and more fruitful environment for driving [1]. In VANET, there are various appli-
cations that will provide security to information that is being exchanged between
the vehicles. This helps in overcoming problems related to accidents, traffic condi-
tion, etc. The trust management scheme is a very important and basic requirement to
maintain a safe and trustworthy exchange of communication in VANET. This helps
maintain a reliable and secured communication among vehicles during their journey
[2]. The reputation and trust issues are very important aspect in this environment for
the decision-making process. The reputation is basically the opinion of vehicles on
each others [3]. We have introduced the concept of evaluation the trustworthiness of
the vehicle on the basis of their responses. Trust is basically the confidence on the
messages that are passed to establish communication in VANETs [4].
The paper is organized as follows: In Sect. 2, we have discussed the work-related
trustworthiness proposed by researchers and academicians. In Sect. 3, various algo-
rithms for computing the trust percentage and trusted location using trustworthi-
ness are discussed. In Sect. 4, computation of trustworthiness of node is presented.
Section 5 discusses the simulation environments and result analysis. Finally, we have
concluded the paper with future direction for further research in Sect. 6.

2 Related Work

The geographical constraint restricts the movement of the vehicles in VANETs.

Furthermore, the vehicles, due to dependent velocities, are spatially dependent on
each other. As a result, the quality of connectivity largely depends upon the quality
of communication link [5–8]. When a new node (i.e., vehicle) joins the network, its
validation is essential to maintain security of the network. A trust-based evaluation is
proposed in [6, 9, 10] based on security authentication mechanisms. The authors have
classified them as direct authentication and indirect authentication. The direct method
is based on the historical security event records. Whenever nodes and authority unit
communicate each other, the events relating to security are recorded in their database.
The authentication units (AUs) that belong to a particular organization can easily
share the same database. The recorded data is very much useful in determining
credibility of a newly joined node. When a group of nodes are communicating,
it solely depends on their decision to accept a new node or not. The acceptance
of a new node depends on trust value and recommendation given by other nodes.
If selfish nodes exist in the network, it is possible that they may deny the new
vehicle node. This condition of the network is undesirable, and this degrades the

[Link]@[Link]
Trust-Based Scheme for Location Finding in VANETs … 45

overall performance of the network. So, the presence of malicious nodes needs to
be identified before computing the indirect trust value in the indirect method. In
this scheme, correlation coefficient of the recommendation trust value is used for
distinguishing malicious nodes. This helps in discarding the recommendation of
trust values received from malicious nodes. Trust evaluation based on the historical
security events and recommendation is discussed in [11].
There is a huge scope of work in trust management for point-to-point networks,
wireless sensor networks (WSNs), VANETs, mobile ad hoc networks (MANETs),
etc. The trust establishments for VANETs discussed in [12–14] are very less. In [12],
authors discussed a trust computation method based on Perron–Frobenius theorem
for VANETs. The trust management based on AHP [13] provides a secure alternative
to replace the current security mechanism. A trust and reputational infrastructure-
aided approach TRIP for VANETs is discussed in [14]. The trust value is computed
with the help of reputation score given by other vehicles. In [15], authors discussed
the intrusion detection system by using watchdog scheme for evaluating them. In
[16], authors devise two techniques that improve throughput for ad hoc network. In
[5] introduce an intrusion detection scheme, which produces a more realistic system,
this system is more reliable in handling malicious nodes. The currently available
trust management schemes and works focus on evaluating and protecting the privacy
of messages.

3 Methodology

We have proposed the following schemes for finding the trusted location of vehicle.
Firstly, trust percentage of information is computed by using the responses received
from vehicles. Based on this, we have calculated the trust percentage of the informa-
tion on the basis of the number of request and number of positive responses. In this
particular case, the trust percentage of the information is calculated. Each vehicle
giving the positive response about the information will be rewarded as points for its
true information providing, thus calculating the trustworthiness of each node present
in the network. In the second case when the trust is below 50%, then instead of going
to RSU or TA the vehicle will ask the trustworthiness of the each node in the network
and take the response of the most trustworthy node.

3.1 Computation of Trust Percentage of Vehicles

In this case, vehicle wants to find that whether the location information is present
or not. It could be about anything like nearby hospitals and hotels. So the source
vehicle broadcast the request (REQ) message to its neighbor vehicles. The neigh-
boring vehicles of the source vehicle reply to REQ with reply (REP) messages. The
source node received REP messages from its neighbour nodes, depending upon true

[Link]@[Link]
46 S. Das et al.

responses vehicle will be rewarded a (+1) point and its number of true response will
be updated. The trust value is computed by source node based on REP messages.
In our work, we have proposed four cases for calculating the trust percentage of
whether the location is present or not.
Case 1 (Computation of trust percentage of location present using vehicle): The
source node received REP messages from its neighbour nodes and compute ratio of
trusted information about a location. If source vehicle receives REP messages from
10 nodes and 6 nodes replied that the location information is present and 4 nodes
replied with no information is available i.e. false reply. Every true response received,
and the vehicle will be rewarded one point thus increasing its trustworthiness. In this
scenario, the ratio is [(6/10) * 100]  60% of the trusted information about location.
The source vehicle may trust broadcast location till it is above 50%; otherwise, a
vehicle can compute location using case 2.
Case 2 (Computation of trusted location using trustworthiness of the most trust-
worthy node): In this case, we have assumed that at every true response of the vehicle,
we will count the total number of true responses of the vehicles and the total number
of the responses of the vehicle. Thus, we will calculate the trustworthiness of each
node present in the network by dividing the total number of true responses by the
total number of responses multiplied by 100. Now, thus, when the trust percentage
of the information is below 50%, source node will directly ask the most trustworthy
node and the response from the most trustworthy node will be more reliable.
Case 3 (Computation of trusted location using RSU): In sparsely populated
network scenario, sometimes it is possible that source vehicle is alone plying. In this
scenario, the particular vehicle collects information from Road-Side Units (RSUs).
The location information received from RSU is fully trusted. In case there is no
neighboring node and RSU available to the source node, then the location is computed
with the help of Case 4.
Case 4 (Computation of Trusted Location Using Trust Authority (TA)): The
source vehicle asks for the location information from trusted authority (TA) (e.g.,
police vans, ambulance, and post office vehicles).

4 Computation of Trustworthiness of Node

For the calculation of the trustworthiness of the node, a buffer is maintained at every
node. This buffer will maintain the information regarding the trustworthiness of the
node. On the basis of the buffer information, the source vehicle will choose the
most trustworthy node among the various vehicles in the range of the source vehicle.
Source vehicle, when found the trust percentage of the information and ask the most
trustworthy node and the reply from the most trustworthy node, will be final. The
proposed working of the scheme involves these steps.

[Link]@[Link]
Trust-Based Scheme for Location Finding in VANETs … 47

Table 1 Trustworthiness of nodes

Total number of responses by Total number of true responses Trustworthiness of the
the vehicle by the vehicles vehicles in percentage
50 50 100
50 45 90
50 40 80
50 35 70
50 30 60

Step 1

Source vehicle broadcast REQ message to its neighboring vehicles and vehicles
responding with positive reply will be awarded one point. A buffer is used for stor-
ing these awarded points. Also, two different counters are used to keep record of
responses received from vehicle. One of the counters stored the total number of true
responses received from the vehicle, and the other keeps the record of total number
of responses received.
The trustworthiness of a node computed according to Eq. 1
 
Total Number of True Response
Trustworthiness  × 100 (1)
Total Number of Response

Step 2
At each response, there is always update among the counters of the nodes regarding
their total true response and total responses, thus updating the trustworthiness of each
node.
Step 3
If at a certain point of time trust percentage of information is below 50%, then the
vehicle will ask the trustworthiness of the nodes within the communication range,
thus sharing the buffer data stored at each node among all the nodes present in the
communication range (Table 1).

Table 2 Simulation Parameter Value

parameters
Number of nodes 10, 20, 30, 40, 50
Simulated area 1000 × 1000
Antenna Omnidirectional
Transmission range of node 150 m
Simulation time (in seconds) 1000
IDE) NetBeans 8.0.2 J2SE

[Link]@[Link]
48 S. Das et al.

Table 3 Number of node exists within the transmission range of different source node
Node density Number of node within the range
Source node: 2 Source node: 3 Source node: 7
10 6 9 9
20 13 19 19
30 20 29 29
40 27 39 39
50 34 47 46

Table 4 Trust percentage of source nodes with variable node density

Number of nodes Trust percentage source nodes
2 3 7
10 33.33 67 55
20 46.15 57 42
30 40.00 51 41
40 44.44 43 48
50 44.11 38 43

5 Simulation and Result Analysis

In this paper, we are implementing our scheme on NetBeans IDE 8.0.2 J2SE. VANET
scenario composed of 10–50 vehicles distributed in an area of 1000 × 1000 units, and
the maximum range of vehicles is 150 m. Vehicles are considered to be at a traffic
point situation. Here, we have arbitrarily considered nodes 2, 3, and 7 as source
nodes and with a different number of nodes present in the range. The network size
gradually increases from 10, 20, 30, 40, and 50. The source nodes (i.e. node 2, 3, 7)
are broadcast messages to know whether a location is present or not. The number of
nodes within the transmission range of source node for a particular time shown in
Table 3. The trust percentage of the information computed after getting the response
from other nodes fall within the transmission range. In different node densities, the
trust of source nodes 2, 3, and 7 is computed as shown in Table 4. In Table 5, we have
computed the trustworthiness of the each node present in the simulation scenario
with various source nodes as 2, 3, and 7 (Table 2).

5.1 Scenario 1 (Source Node Is 2)

Figure 1 shows the node deployment of various nodes considered for simulation.
Figure 2 shows the trust percentage of source node 2 under different node densities.
The trust percentage obtain is always below 50% in different node density. For source
node 2 in Table 5 and Fig. 9, nine nodes present in the network whose trustworthiness

[Link]@[Link]
Trust-Based Scheme for Location Finding in VANETs … 49

Table 5 Trustworthiness of each node

Node ID Source node Node ID Source node
2 3 7 2 3 7
1 41 47 66 26 30 66 77
2 0 47 50 27 40 46 55
3 41 0 41 28 0 40 80
4 66 47 58 29 0 46 77
5 66 52 33 30 0 60 55
6 50 41 58 31 44 57 37
7 33 58 0 32 55 42 62
8 0 41 58 33 55 60 62
9 0 64 58 34 44 66 50
10 0 29 41 35 40 50 75
11 36 62 63 36 66 42 87
12 69 56 45 37 44 64 25
13 63 56 45 38 0 71 50
14 63 43 63 39 0 57 50
15 45 43 45 40 0 35 62
16 36 43 45 41 75 46 100
17 45 37 18 42 37 38 50
18 0 51 81 43 37 69 50
19 0 68 45 44 66 46 0
20 0 31 45 45 37 38 50
21 40 46 66 46 37 53 100
22 30 68 55 47 37 53 0
23 30 66 55 48 0 46 0
24 30 60 66 49 0 0 50
25 50 60 33 50 0 0 50

is above 50%. It can ask the node 41 directly with the trustworthiness of 75%. The
answer by this node would be more reliable in comparison to the first technique
which was just asking to the nodes for the information without any prior knowledge
of the nodes participating for the trust calculation (Fig. 3).

5.2 Scenario 2 (Source Node Is 3)

Figure 4 shows the deployment of nodes for source node 3. Now, in Fig. 5, it observed
that twice trust percentage goes below 50. To overcome such situation, Table 5 and
Fig. 6, there are 23 nodes present in the network whose trustworthiness is above 50%
and one being 71% as the highest trustworthy node. Thus in such a situation, we

[Link]@[Link]
50 S. Das et al.

have a chance to test the information on a better scale and have a better results. It can
ask the node 38 directly with the trustworthiness of 71%. The answer by this node
would be more reliable in comparison to the first technique which was just asking to
the nodes for the information without any prior knowledge of the nodes participating
for the trust calculation. In this particular scenario, we can also notice that time to
take decision will be much less if we have to take request and response from one
vehicle.

5.3 Scenario 3 (Source Node Is 7)

Figure 7 shows the node deployment for source node 7. In Fig. 8, trust percentage is
above 50% only when the number of node in the network is 10 and for other node
density the values goes below 50. To overcome this problem, from Table 5 and Fig. 9,
we can see that there are 24 nodes present in the network whose trustworthiness is
above 50% and two being 100% as the highest trustworthy nodes. Thus in such a
situation, we have a chance to test the information on a better scale and have better
results. It can ask the node 41 or 46 directly with the trustworthiness of 100%. The
answer provided by this node would be more reliable in comparison to the first
technique which was just asking to the nodes for the information without any prior
knowledge of the nodes participating for the trust calculation. Here we can notice
that the source node 7 has an option of two nodes in the range that too with 100% of
trustworthiness. Thus in this particular case would be like it is better to go for such
vehicle which has its trustworthiness 100% instead of trust percentage below 50%.

Fig. 1 Node placement

[Link]@[Link]
Trust-Based Scheme for Location Finding in VANETs … 51

45

40

35
Trust Percentage

30

25

20

15

10

0
10 20 30 40 50
Number of Node

Fig. 2 Trust percentage of node 2 with varying node density

80

70

60
Trustworthiness

50

40

30

20

10

0
0 5 10 15 20 25 30 35 40 45 50
Node ID

Fig. 3 Trustworthiness of each node w.r.t node 2

6 Conclusion and Future Direction

We can see the variation in trust percentage of information when we increase the
number of nodes and thus increasing the number of responses. Now we can use the
trustworthiness of the each node and can find the most trustworthy node. Request
the most trustworthy node, and the response of the most trustworthy node will be
more reliable when the trust percentage of the information is below 50%. Thus we

[Link]@[Link]
52 S. Das et al.

Fig. 4 Node placement during simulation (source node  3)

70

65

60
Trust Percentage

55

50

45

40

35
10 15 20 25 30 35 40 45 50
Number of Node

Fig. 5 Trust percentage of node 3 with varying node density

found that our proposed scheme is more reliable in finding the genuiness of requested
information and gives a more satisfactory response to the situation. Thus, we find
that the above technique is more worthy for the current situation proposed in this
paper.
There is a lot to emphasize on in this area such as reducing the response time
of the vehicles in serious conditions. We must focus on bringing those schemes
which could handle the openness in the environment of the VANETs. There could be
various ways of handling the situation if we can come up with it may result in much

[Link]@[Link]
Trust-Based Scheme for Location Finding in VANETs … 53

80

70

Trustworthiness 60

50

40

30

20

10

0
0 5 10 15 20 25 30 35 40 45 50
Node ID

Fig. 6 Trustworthiness of each node (source node is 3)

Fig. 7 Node placement during simulation (source node  7)

more accuracy. There are many issues in which VANET scenario could be handled
in better ways. The whole concept revolves around if we could find a way to check
whether the response is trustworthy or not. So considering such scenario, we must
try to progress our work for further improvement.

[Link]@[Link]
54 S. Das et al.

55
50
45
Trust Percentage 40
35
30
25
20
15
10
5
0
10 15 20 25 30 35 40 45 50
Number of Node

Fig. 8 Trust percentage of node 7 with varying node density

100

90

80

70
Trustworthiness

60

50

40

30

20

10

0
0 5 10 15 20 25 30 35 40 45 50
Node ID

Fig. 9 Trustworthiness of each node (source node is 7)

References

1. Toor, Y., Muhlethaler, P., Laouti, A.: Vehicle ad hoc networks: applications and related technical
issues. IEEE Commun. Survey Tutor. 10(3), 74–88 (2008)
2. Sharma, K., Soni, S., Chaurasia, B.K.: Reputation and trust computation in VANETs. In: Pro-
ceedings of the IEMCON-14 Conference on Electronics Engineering and Computer Science,
pp. 118–122 (2014)
3. Ding, Q., Jiang, M., Li, X., Zhou, XH.: Reputation-based trust model in vehicular ad hoc
networks. In: International Conference on Wireless Communications and Signal Processing
(WCSP), pp. 1–6 (2010)

[Link]@[Link]
Trust-Based Scheme for Location Finding in VANETs … 55

4. Sumra, I.A., Hasbullah, H., Lail, J., Rehman, M.: Trust and trusted computing in VANET.
Comput. Sci. J. 1(1), 29–51 (2011)
5. Hasswa, A., Zulkernine, M., Hassanein, H.: Routeguard: an intrusion detection and response
system for mobile ad hoc networks. In: Proceedings of the IEEE International Conference on
Wireless and Mobile Computing, Networking and Communications, vol. 3, pp. 336–343(2005)
6. Tyagi, P., Dembla, D.: Investigating the security threats in vehicular ad hoc networks (VANETs):
towards security engineering for safer on-road transportation. In: International Conference on
Advances in Computing, Communications and Informatics, pp. 2084–2090 (2014)
7. Zhang, J.: A survey on trust management for VANET. In: International Conference on Advanced
Information Networking and Applications, IEEE Computer Society, pp. 105–112 (2011)
8. Yang, X., Liu, L., Vaidya, N.: A vehicle-to-vehicle communication protocol for cooperative
collision warning. In: 1st Annual International Conference on Mobile and Ubiquitous Systems:
Networking & Services, pp. 114–123 (2004)
9. Shuai, Z., Jun, T., Yijian, Y.: Anonymous authentication—oriented vehicular privacy protection
technology research in VANET. In: IEEE International Conference on Electrical and Control
Engineering, pp. 4365–4368 (2011)
10. Zhang, C.: An efficient message authentication scheme for vehicular communication. IEEE
Trans. Veh. Technol. 57(6), 3357–3368 (2008)
11. Tong, Z.: P2DAP—Sybil attacks detection in vehicular ad hoc networks. IEEE J. Sel. Areas
Commun. 29(3), 582–594 (2011)
12. Chaurasia, B.K., Verma, S., Tomar, G.S.: Trust computation in VANETs. In: Proceedings of
the International Conference on Communication Systems and Network Technologies (CSNT),
pp. 468–471 (2013)
13. Saraswat, D., Chaurasia, B.K.: AHP based trust model in VANETs. In: Proceedings of the IEEE
5th International Conference on Computational Intelligence and Communication Networks
(CICN), pp. 391–393 (2013)
14. Marmol, F.G., Perez, G.M.: TRIP a trust and reputation infrastructure based proposal for
vehicular ad hoc networks. J. Netw. Comput. Appl. 35(3), 934–941 (2012)
15. Obimbo, C., Arboleda, L.M., Chen, Y.: A watchdog enhancement to IDS in MANET. In:
Proceedings of the IASTED Conference on Wireless Networks, pp. 510–515 (2006)
16. Marti, S., Giuli, T.J., Lai, K., Baker, M.: Mitigating routing misbehavior in mobile ad hoc net-
works. In: Proceeding of the 6th ACM Annual International Conference on Mobile Computing
and Networking, pp. 255–265(2000)

[Link]@[Link]
Implementation and Analysis of Energy
Efficiency of M-ary Modulation Schemes
for Wireless Sensor Network

Samir Ahmad Sheikh and Sindhu Hak Gupta

Abstract In a wireless sensor network because of the path loss, there is a consid-
erable difference in the signal strength from the transmitted node to receiving node.
Path loss is a crucial factor for a WSN, and it may be evaluated using stochastic,
deterministic, or empirical methods. For a WSN, it is challenging task to optimize the
transmission power, reliability, and data rate in the presence of path loss. Choice of
appropriate modulation scheme is the most important aspect of the physical layer. As
optimum modulation scheme is capable of minimizing the error and making WSN
more reliable. In this paper, a new approach is considered to relate path loss of the
WSN to M-ary modulation schemes. Critical comparative analysis for M-ary FSK
and M-ary PSK is done for the case scenario. Performance is analyzed for free space
earth model and plane earth model. Results indicate that PSK is more reliable than
FSK, whereas the data transmission rate for FSK is greater than PSK. Power required
to transmit the data in FSK is less than less.

Keywords Path loss · Data rate · Power transmitted · PSK · FSK · Reliability

1 Introduction

A wireless sensor network (WSN) is the combination of autonomous devices. These

autonomous devices are spatially scattered and are used for monitoring and recording
physical parameters like pressure, temperature, humidity, EEG, and ECG for envi-
ronment or living body, respectively [1]. Based on this fact, WSN has got various
applications in upcoming fields like wireless body area network (WBAN), emer-
gency relief, and precision agriculture (PA). Wireless sensor networks are used in

S. A. Sheikh (B) · S. H. Gupta

Department of Electronic and Communication Engineering, ASET,
Amity University, 125, Noida, UP, India
e-mail: samirkashmir@[Link]
S. H. Gupta
e-mail: shak@[Link]

© Springer Nature Singapore Pte Ltd. 2019 57

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
58 S. A. Sheikh and S. H. Gupta

different fields to gather the information, and thus collecting the information, the
sensor nodes communicate with each other and pass the information further to clus-
ter head. Cluster heads transmit it to the base station and from that it is processed and
examined. But while communicating with each other they face attenuation, fading,
noise and path loss [2, 3]. The prime focus is energy efficiency between the sensor
nodes, as low battery power and limited computation capability are inherent features
of sensors. This may cause data loss evaluation of transmitting signal. This leads for
advanced design methodologies to address the energy content in any wireless sensor
network. For this research paper, main focus would be path loss, power transmitted,
data rate, and reliability between two sensor networks.
Path loss models are used to abstract the actual characteristic of the sensor nodes
in the wireless sensor network. The appropriate modeling of the propagation and
path loss is of predominant in the WSN system analysis and design [4]. Most impor-
tant parameters of any WSN are energy dissipation, reliability, route optimization,
decision support system, data rate, power transmitted, reliability, and connectivity.
The sensor nodes which are used in the precision agriculture communicate with each
other and have low power efficiency. Power to the nodes is usually provided by the
batteries. But batteries have a finite energy and need replacement when depleted,
which increase the maintenance cost. Path loss is one of the important parameters
used to evaluate the performance of WSN. This factor depicts the difference between
which exists between the transmitted power and the received power. There are dif-
ferent types of path loss models like free space path loss model, plane earth path loss
model, and COST 231 [2, 3]. These all models are all in terms of the frequency. But
new models correlate path loss to different modulation schemes which are developed
for a wireless sensor network.
In every WSN modulation scheme, it plays an important role in transmitting
the information in an efficient way. There are different types of M-ary modulation
schemes that exist in the literature which govern the transmission of data because
of the three fundamental parameters as frequency, time, and phase. These digital
modulation schemes include basic schemes such as M-ary frequency shift keying,
M-ary phase shifting keying, and amplitude shifting keying (ASK). But our focus
is on M-ary phase shift keying and M-ary frequency keying. The choice of efficient
modulation schemes is important for reliable communication in wireless sensor net-
work [5]. Using M-ary modulation schemes in transmission of data can reduce the
transmission time by sending multiple bits per symbol, and it results in the tangled
circuitry and increased radio power consumption. The trade-off parameters are for-
mulated in and it is concluded that under start-up power most influential conditions,
the binary modulation scheme is more efficient. So in any wireless sensor network,
crucial point is the choice of modulation scheme by which the life time, efficiency,
and reliability of the sensor nodes can be increased. Several factors like data rate
and symbol rate, reliability, power transmitted, data rate, and path loss have been
analyzed in order to determine the efficiency of different M-ary modulation schemes
[5]. This will lead to improve the performance, life time, and reliability of the sensors
in a wireless sensor network. The main contribution of this paper is:

[Link]@[Link]
Implementation and Analysis of Energy Efficiency … 59

• To develop a mathematical model that correlates path loss to M-ary modulation

schemes;
• With the help of simulation results analyze the performance of M-ary PSK and
M-ary FSK for considered schemes.

2 System Model

The system model is defined for a wireless sensor network model which is linear and
the nodes are linearly spaced and are equidistant to each other with a fixed distance
of 10 m. The total numbers of nodes are fixed, and each node is homogenous. In this
scenario, receivers are assumed to be synchronized for simulation. The communica-
tion between the sensor and the sink is wireless. To evaluate the path loss plane earth
path loss (PE) and free space path loss model (FS) can be used.
Free space path loss model is derived from path loss equation and which is
expressed in dBm [2, 3]. The application of the model can be conceivably applied
to development as the minimum level for the computation of the path loss as this
Eq. (1) represents free space plane path loss model.

PL(dBm)  −27.56 + 20 log10 ( f ) + 20 log10 (d) (1)

where d is the distance between the sender and the receiver node in meters, and f is
the frequency in Mhz.
For frequency of 2.4 Ghz, Eq. (1) becomes:

PL(dBm)  −27.56 + 20 log10 (2400) + 20 log10 (d)

PL(dBm)  40.40 + 20 log10 (d)
20 log10 (d)  PL − 40.40
d  10(PL−40.40)/20 . (2)

Plane earth path loss model is derived from path loss equation and which is
expressed in dBm [2, 3]. The estimated path loss is expressed as follows:

PL(dBm)  40 log10 (d) − 20 log10 (h t ) − 20log10 (h r ) (3)

where h t and h r are height of transmitter and receiver antenna in cms, and d is the
distance between the transmitter and the receiver node in meters. For h t  h r 
13 cm, Eq. (3) can be written as follows:

[Link]@[Link]
60 S. A. Sheikh and S. H. Gupta

PL  40 log10 (d) − 20 log10 (0.13) − 20log10 (0.13)

PL  40 log10 (d) + 35.44
40 log10 (d)  PL − 35.44
d  10(PL−35.44)/40 (4)

In a WSN, it has been stated in [6] that energy per bit E b in M-ary modulation
schemes may be represented as given in Eq. (5)
Eb 1 Prevd 1
 SNR ·  ·
No R No R
 
1 Pt · G r · G t · λ2 do γ
 (5)
No · R (4π )2 · doy · L d

where G t · G r and γ are gain of the transmitter, gain of receiver antenna, and path
loss exponent, respectively, and R is data rate, Pt is the power transmitted, and NEoa
is the ratio between the energy per bit and the noise level which is fix for different
M-ary modulation schemes.
A novel mathematical relation is proposed using Eqs. (2) and (4) in Eq. (5) which
correlates path loss to different modulation schemes and is developed for a WSN.
Path loss can be related to the power transmitted and the data rate by keeping one as
the constant. For M-ary PSK and FSK modulation schemes using Free Space Path
Loss model which correlates power transmitted, data rate with the Path Loss:

Eb N0 · R · (4π )2 · L γ (PL−40.04)
Pt  × × 10 20 (6)
No Gt · Gr · λ 2

For M-ary PSK and FSKmodulation schemes using Plane Earth Path Loss model
which correlates power transmitted, data rate with the Path Loss:

Eb No · R · (4π )2 · L γ (PL−35.44)
Pt  × × 10 40 (7)
No Gt · Gr · λ 2

where γ is path loss exponent  3.5, do  1m,G r  G t  L  1, R (data rate) and

No (noise level)  −180 dB.
Using Eq. (6) and (7) in Eq. (8), reliability [1] can be expressed in terms of path
loss for both free space path loss model and earth space path loss model.
 
Pt ∗ |h|2
Reliablity  W log2 1 + (8)
No ∗ W

where W is the bandwidth, and since in this scenario ISM band is assumed, the value
of W will be 2.4 GHz.

[Link]@[Link]
Implementation and Analysis of Energy Efficiency … 61

Eb
Table 1 No [dB] required for M-ary FSK and M-ary PSK [6]
m 2 4 8 16 32 64
M-ary PSK: 10.5 11 14 18.5 23.4 28.5
Eb
No
M-ary FSK: 13.5 10.8 9.3 8.2 7.5 6.9
Eb
No

From the data base, different M-ary modulation schemes can be analyzed. The
table below shows the variation of NEbo with changing the value of M for PSK and
FSK modulation schemes [1] (Table 1).
The proposed mathematical system model operates over M-ary modulation under
different conditions. These conditions include increasing number of modulation lev-
els and NEbo which is different for different modulation schemes [6, 4].

3 Results and Discussion

In order to determine the performance of the system, we compare different parameters

like power transmitted, data rate, and reliability with the path loss in a wireless sensor
network using free space earth path loss model and plane earth path loss model.
Figures 1 and 2 illustrate the effect of path loss on the power transmitted for
M-ary modulation schemes (PSK and FSK) using free space path loss model. From
the above figures, it is observed that as the path loss increases, there is an increase
in the power transmitted. With increase in the number of bits, the power required to
transmitted increases but for M-ary FSK less power to transmit the data as compared
to the M-ary PSK.

Fig. 1 Power transmitted 10 -9

4
versus path loss using free
2-FSK
space earth path loss model 3.5 4-FSK
8-FSK
16-FSK
3
Power Transmitted

32-FSK
64-FSK
2.5

1.5

0.5

0
0 5 10 15 20 25 30
Path Loss

[Link]@[Link]
62 S. A. Sheikh and S. H. Gupta

Fig. 2 Power transmitted 10-7

1.2
versus path loss using free
space earth path loss model 2-PSK
1 4-PSK
8-PSK

Power Transmitted
16-PSK
0.8 32-PSK
64-PSK

0.6

0.4

0.2

0
0 5 10 15 20 25 30
Path Loss

Figures 3 and 4 illustrate the effect of path loss on the data rate for M-ary modula-
tion schemes (PSK and FSK) using free space path loss model. It is observed as the
path loss increases, the data rate decreases, and with increasing, the number of bits
the data rate to transmitted information increases but for M-ary FSK shows better
data rate as compared to the M-ary PSK.
Figures 5 and 6 illustrate the effect of path loss on the reliability for M-ary mod-
ulation schemes (PSK and FSK) using free space path loss model. It is observed
that with increasing the number of bits, the reliability to transmitted information
increases but for M-ary PSK shows better reliability as compared to the M-ary FSK.

Figures 7 and 8 illustrate the effect of path loss on the power transmitted for
M-ary modulation schemes (PSK and FSK) using plane earth path loss model. It is
observed that with increasing the number of bits, the power required to transmitted
information increases but for M-ary FSK less power to transmit the data as compared
to the M-ary PSK.

Fig. 3 Data rate versus path 10 16

loss using free space earth 18
path loss model 16 2-PSK
4-PSK
14 8-PSK
16-PSK
32-PSK
12 64-PSK
data rate

10
8
6
4
2
0
0 5 10 15
Path Loss

[Link]@[Link]
Implementation and Analysis of Energy Efficiency … 63

Fig. 4 Data rate versus path 10

17

loss using free space earth 4.5

path loss model 4 2-FSK
4-FSK

3.5 8-FSK
16-FSK
32-FSK
3 64-FSK

data rate
2.5
2
1.5
1
0.5
0
0 5 10 15
Path Loss

Fig. 5 Reliability versus 10 9

path loss using free space 3.5
2-FSK
earth path loss model 3 4-FSK
8-FSK
16-FSK
2.5 32-FSK
64-FSK
Reliability

1.5

0.5

0
0 5 10 15 20 25 30
Path Loss

Fig. 6 Reliability versus 10 9

path loss using free space 14
2-PSK
earth path loss model 4-PSK
12 8-PSK
16-PSK
10 32-PSK
64-PSK
Reliability

0
0 5 10 15 20 25 30
Path Loss

Figures 9 and 10 illustrate the effect of path loss on the data rate for M-ary
modulation schemes (PSK and FSK) using plane earth path loss model. It is observed

[Link]@[Link]
64 S. A. Sheikh and S. H. Gupta

Fig. 7 Power transmitted 10

-8
8
versus path loss using plane
earth path loss model 7 2-FSK
4-FSK
8-FSK
6 16-FSK

Power Transmitted
32-FSK
64-FSK
5

0
0 5 10 15 20 25 30
Path Loss

Fig. 8 Power transmitted 10 -6

2.5
versus path loss using plane
earth path loss model 2-PSK
4-PSK
2 8-PSK
Power Transmitted

16-PSK
32-PSK
64-PSK
1.5

0.5

0
0 5 10 15 20 25 30
Path Loss

Fig. 9 Data rate versus path 10

13

loss using plane earth path 7

loss model 6 2-FSK
4-FSK
8-FSK
5 16-FSK
32-FSK
64-FSK
data rate

0
0 5 10 15
Path Loss

that with increasing the number of bits, the data rate to transmitted information
increases but for M-ary FSK shows better data rate as compared to the M-ary PSK.

[Link]@[Link]
Implementation and Analysis of Energy Efficiency … 65

Fig. 10 Data rate versus 10

13

path loss using plane earth 3

path loss model 2-PSK
2.5 4-PSK
8-PSK
16-PSK
32-PSK
2 64-PSK

data rate
1.5

0.5

0
0 5 10 15
Path Loss

Figures 11 and 12 illustrate the effect of path loss on the reliability for M-ary
modulation schemes (PSK and FSK) using plane earth path loss model. It is observed
that with increasing the number of bits, the reliability to transmitted information
increases but for M-ary PSK shows better reliability as compared to M-ary FSK.
Tables 2 and 3 illustrate the variation in the power transmitted, data rate, and
reliability at path loss of 10 dBm. Clearly, it satisfies the results which are above
mentioned. But when comparing the results of free space path loss model and the
plane earth path loss model, the plane earth path loss model illustrates the better
results in terms of power transmitted, reliability as compared to the free space earth
path loss model. Data rate is less in plane earth path loss model as compared to free
space earth path loss model.

Fig. 11 Reliability versus 10 9

path loss using plane earth 14
path loss model 12
2-FSK
4-FSK
8-FSK
16-FSK
10 32-FSK
64-FSK
Reliability

0
0 5 10 15 20 25 30
Path Loss

[Link]@[Link]
66 S. A. Sheikh and S. H. Gupta

Table 2 Effect of path loss on M-ary modulation schemes in wireless sensor network at PL 
10 dBm using free space path loss model
M-ary modulation Power Data rate (kbps) Reliability
transmitted (mW)
2-FSK 12.50 × 10−10 2.26 × 1015 18.10 × 105
4-FSK 6.71 × 10−10 4.39 × 1015 9.66 × 105
8-FSK 4.75 × 10−10 6.20 × 1015 6.85 × 105
16-FSK 3.69 × 10−10 7.99 × 1015 5.32 × 105
32-FSK 3.14 × 10−10 9.39 × 1015 4.53 × 105
64-FSK 2.73 × 10−10 10.7 × 1015 3.94 × 105
2-PSK 6.28 × 10−10 4.70 × 1015 9.04 × 105
4-PSK 7.03 × 10−10 4.19 × 1015 1.01 × 106
8-PSK 1.40 × 10−9 2.10 × 1015 2.02 × 106
16-PSK 3.95 × 10−8 7.63 × 1014 2.71 × 106
32-PSK 1.22 × 10−8 2.36 × 1014 1.75 × 107
64-PSK 3.95 × 10−8 7.60 × 1013 5.65 × 107

Table 3 Effect of path loss on M-ary modulation schemes in WSN at PL  10 dBm using plane
earth path loss model
M-ary modulation Power Data rate (kbps) Reliability
transmitted (mW)
2-FSK 1.34 × 10−6 2.23 × 1012 15.39 × 108
4-FSK 7.21 × 10−7 4.15 × 1012 9.09 × 108
8-FSK 5.10 × 10−7 5.87 × 1012 6.67 × 108
16-FSK 3.96 × 10−7 7.87 × 1012 5.29 × 108
32-FSK 3.37 × 10−7 8.89 × 1012 4.55 × 108
64-FSK 2.93 × 10−7 10.29 × 1012 4.00 × 108
2-PSK 6.73 × 10−7 4.45 × 1012 8.56 × 108
4-PSK 7.53 × 10−7 3.97 × 1012 9.47 × 108
8-PSK 1.50 × 10−6 1.99 × 1012 1.68 × 109
16-PSK 4.24 × 10−6 7.06 × 1011 3.52 × 109
32-PSK 1.31 × 10−5 2.23 × 1011 6.46 × 109
64-PSK 4.24 × 10−5 7.06 × 1010 1.01 × 1010

[Link]@[Link]
Implementation and Analysis of Energy Efficiency … 67

Fig. 12 Reliability versus 10 10

2.5
path loss using plane earth
2-PSK
path loss model 4-PSK
2 8-PSK
16-PSK
32-PSK
64-PSK

Reliability
1.5

0.5

0
0 5 10 15 20 25 30
Path Loss

4 Conclusion

This paper analyzes the performance of different M-ary modulation schemes like PSK
and FSK for wireless sensor networks. The aim is to analyze the energy efficiency,
reliability, data rate, and lifetime of the sensor nodes. For the considered scenario,
better modulation strategy to minimize the total power transmitted and data rate
is required to send a given number of bits. Also evaluate reliability of free space
earth model and plane earth model is analyzed, and the results are validated. Results
indicate that PSK is more reliable than FSK, whereas the data transmission rate for
FSK is greater than PSK. Power required to transmit the data in FSK is less than PSK.
Plane earth path loss model illustrates the better results in terms of power transmitted
and reliability as compared to the free space earth path loss model. Data rate is less
in plane earth path loss model as compared to free space earth path loss model.
With increasing the number of bits the power required to transmit the data increases
but for M-ary PSK less power to transmit the data as compared to the M-ary FSK.
The obtained results can be used as a design approach for modulation techniques in
wireless sensor network.

[Link]@[Link]
68 S. A. Sheikh and S. H. Gupta

References

1. Proakis, J.G.: Companders. In: Wiley Encyclopedia of Telecommunications (2003)

2. Vougioukas, S., Regen, C., Anastasiuamd, H., Zude, M.: Comparison of radio path loss model
for wireless sensor networks in orchard environments. In: The CIGR-AgEng International
Conference of Agricultural Engineering, Valencia, Spain, 2012
3. Song Meng, Y., Chong Ng, B., Hui Lee, Y.: Empirical near ground path loss modelling in forest
at VHF and UHF bands. IEEE Trans. Antennas Propag. 57(5), 1461–1468 (2009)
4. Al-Turjman, F.: Cognitive-node architecture and development strategy for the future WSNs.
Mob. Netw. Appl. 1–19 (2017)
5. Shivaprkasha, K.S., Patkar, R., Kulkarni, M.: Performance analysis of energy efficient modu-
lation and coding schemes for wireless sensor networks. Int. J. Parallel Emergent Distrib. Syst.
28(6), 576–589 (2013)
6. Karl, H., Willing, A.: Protocols and Architecture for Wireless Sensor Networks. Wiley. ISBN
0-470-09510-5
7. Kulkami, R.V., Forster, A., Venayaganmoorthy, G.K.: Computational intelligence in wireless
sensor networks: a survey. Commun. Surv. Tutor. 13(1), 68–96 (2011)
8. Srbinovska, M., Dimcev, V., Borozan, V., Gavrovski, C.: Environmental parameters monitoring
in precision agriculture using wireless sensor networks. J. Clean. Prod. 88, 297–307 (2015)
9. Yuan, X., et al.: A genetic algorithm-based dynamic clustering method towards improved WSN
longevity. J. Netw. Syst. Manage. 25(1), 21–46 (2017)
10. Mafuta, M., et al.: Successful development of a wireless sensor network for precision agriculture
in Malawi. Int. J. Distrib. Sens. Netw. 9(5), 150703 (2013)

[Link]@[Link]
Protecting Multicast Sessions Against
Single-Link Failure in Survivable WDM
Mesh Networks

B. Mohapatra and Kunal Sain

Abstract Optical fiber network happens to be dominant transport medium because

of its high data-carrying capacity. A single fiber failure in such networks can disrupt
the information disseminated to several destination nodes, resulting in huge loss
of data. Thus, it is imperative that multicast sessions in WDM networks should
be protected from fiber failures. In this paper, we propose three novel protection
schemes to protect multicast sessions against single fiber cut. The proposed schemes
reduce the amount of network resource used by a session by reducing the number of
branching in the multicast tree. As it is observed that more the number of leaf nodes in
the multicast tree, more is the amount of network resource required to protect it. The
performance of the proposed schemes is evaluated and compared with the reported
state-of-the-art algorithm in the presence of dynamic traffic. An improvement of
7–15% is achieved using blocking probability as a performance metric.

Keywords Multicasting · Protection · Survivability

Wavelength-Division multiplexing

1 Introduction

Recent development in the field of optical networking, especially in WDM technol-

ogy, high-bandwidth applications (both unicast and multicast) such as video confer-
encing, multiplayer gaming, high-definition television, movie broadcast, and other
bandwidth-intensive real-time applications, has become very popular. Optical WDM
network will perhaps also be used as the backbone for tomorrow’s Internet. An effi-
cient method of achieving multicast is through establishment of a multicast tree, with

B. Mohapatra (B)
Department of Electronics and Communication Engineering, School of Electrical, Electronics and
Communication Engineering, Galgotias University, Greater Noida 201310, India
e-mail: writetobm@[Link]
K. Sain
Oracle India Pvt. Ltd., Bangalore 560025, India

© Springer Nature Singapore Pte Ltd. 2019 69

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
70 B. Mohapatra and K. Sain

the destination nodes as the leaf or intermediate nodes and the source as the root node.
In case of optical networks, such a multicast tree is often known as a light-tree.
In order to support multicasting, in optical WDM networks, the network nodes
must be equipped with multicast-capable wavelength-routing switches (MWRSs)
[1] or optical cross-connects (MC-OXC) [2]. Multicast-capable wavelength-routing
switches (MWRSs) replicate a bit stream from one input port to multiple output ports.
MWRS use opaque cross-connects to perform optical–electronic–optical (O/E/O)
conversion. A signal arriving on the input fiber link is replicated into multiple copies
in the electronic domain. One copy is dropped locally at the node, and the remaining
copies are switched to different channels on outgoing fibers. Full wavelength conver-
sion is inherent in such cross-connects and, hence, no explicit wavelength convertors
are needed.
Fiber cuts in an optical network lead to huge loss of data, (generally in the order
of terabytes). In the case of a single fiber failure in a unicast session, only one single
destination is affected. However, even a single-link failure in a multicast session can
be devastating as it causes communication failure to multiple destinations. Thus,
provision of adequate backup is essential in optical network to provide survivable
multicast-based applications. The problem of designing survivable WDM network
to support multicast-based applications is more challenging than to provide unicast-
based applications. The problem gets increased when the number of destinations gets
increased (especially when capacity sharing is taken into consideration).
A. Definition
In this section, some of the terms that are defined have been used throughout the
paper.

Light-path: In an optical network, it is a point-to-point path of light through which

data is transferred.
Light-tree: In case of multicast sessions, the source and destinations are linked
by a collection of light-paths. For a particular multicast session, the collection of
light-paths connecting the source and the destinations form a light-tree.
Link-disjoint: Two paths/trees are said to be link-disjoint if they do not have any
common link among them.
Spanning tree: A spanning tree T of a connected, graph G is a tree composed of all
the vertices and some (or all) of the edges of G.

B. Related work
Significant work has been done to protect and restore light-paths for point-to-
point communication [3–6]. Several schemes, such as path protection, sub-path
protection and link protection with or without resource sharing, have been pro-
posed. Since a single fiber cut may cause serious data loss in case of multicas-
ting than an individual light-path, hence, protection of multicast session from a
single-link failure has been considered in recent literature [7–11].

One of the simplest solutions to the problem of protecting multicast tree is by

computing a link-disjoint backup tree. Two link-disjoint trees together can provide

[Link]@[Link]
Protecting Multicast Sessions Against Single-Link … 71

dedicated 1 + 1 protection to the multicast session. However, this scheme leads to

usage of excessive network resources. Link protection scheme is proposed in [12].
In this scheme, for each link on the light-tree, a backup route is derived. This scheme
suffers from the obvious disadvantage of enormous network resource usages. In [10],
the authors proposed SDS (Shared Disjoint Segment), a segment protection scheme
and OPP_SDP (Optimal Path Pair-based Shared Disjoint Paths), a path protection
scheme. A segment in a multicast tree is a sequence of nodes from the root or a
splitting point to a leaf or another downstream splitting point. In SDS, each segment
in the primary light-tree is protected using link-disjoint backup path. In OPP_SDP,
a link-disjoint path pair is found for each destination. There is also provision for
sharing of a common link between different path pairs (i.e., between primary and
backup; as well as, backup and backup of the same session). However, both the above
algorithms (i.e., SDS and OPP_SDP) do not take into account the concept of spare
capacity sharing among different multicast trees.
The author proposed two new schemes in [13] based on ADALINE (average
number of destination nodes that are affected by a single-link failure). Assuming
a multicast session has a source ‘S’ and four destinations ‘A’, ‘B’, ‘C’, ‘E’. Costs
of the links are shown adjacent to the links. Then the ADALINE is calculated as
follows: failure of the links S–A, A–B, A–C and B–F affects 4, 2, 1 and 1 destination,
respectively. Hence, the ADALINE is calculated as 8/4  2. On an average, a single-
link failure will prevent 2 destination nodes.

Example Network

A modified version of the state-of-the-art scheme OPP_SDP and named as mul-

ticast protection through spanning paths (MPSP) is presented in [14]. A spanning
path is defined as a path in a multicast tree which extends from a leaf node to
other leaf nodes. Here, the authors consider three kinds of backup path sharing,
such as self-sharing, intra-request sharing, and inter-request sharing. Self-sharing
scheme suggests sharing of path among primary and backup path in a primary light-
tree. Intra-request sharing scheme suggests sharing of path among different backup
paths to different destinations in a primary light-tree. Whereas, inter-request sharing
schemes considers, sharing of backup paths of two different primary light-tree under

[Link]@[Link]
72 B. Mohapatra and K. Sain

single-link failure assumption. The result shows significant improvement in terms

of average cost and blocking probability over the OPP_SDP scheme. The author
assumes that wavelengths reserved in a fiber can be used in forward and in back-
ward directions by reconfiguring the switches. However, this assumption cannot be
achieved practically.
Cost and blocking probability can be reduced by removing the residual links from
the backup paths of protected tree are presented in [15]. On the basis of the above
theory, the author presented two algorithms, named as, OPP-RRL (Optimal Path Pair-
based Removing Residual Links) and SLP-ARL (Source Leaf Path-based Avoiding
Residual Links). The basic idea is presented here, using the example network. First
derive the backup paths for all the individual links of the primary light-tree. S–C–A
for S–A, S–C for A–C, C–D–B for A–B and B–D–F for B–F, respectively, are the
backup paths for individual links. Instead of the above, one backup path may be
derived as S–C–D–E to provide connection between source and all the destinations.
Using this technique, the results claim marginal improvements over OPP_SDP in
terms of wavelength link cost and blocking probability.
Segment-based protection scheme is proposed in [16]. A segment is a link which
originates from the source or from intermediate nodes. Failure of a segment (link)
adjacent to the source node disrupts many destinations; hence, a huge of revenue loss
will occur. To overcome this problem, the author proposed segment-based protection
tree (SPT) scheme and claims a significant reduction of cost over OPP_SDP scheme.
A light-tree is constructed for a multicast session. For protection, another multicast
tree is constructed. Each protection tree is originated from the source and reaches
to every destination of the multicast session. Each segment of the primary tree is
protected by the protection tree; also, the protection tree can share wavelength links
with primary tree as well as another protection tree. Hence, it is a case of different
sharing approaches, such as self-sharing, intra-request sharing, and inter-request
sharing, already discussed.
P-cycle is another technology, which also applied to protect multicast sessions
is presented in [17–19]. This technique aims to constructs a cycle which traverse
through the source and all the destinations. Though this offers fast recovery time
(which is an important criterion for restoration), however, it is less capacity efficient.
Recently, elastic optical network (EON) has attracted the attention among the
researchers. To further increase the efficiency and data rate offered by the optical
network, coherent optical OFDM technique is a viable option. Based on this, the
author proposed both mathematical and heuristic models to protect multicast sessions
in [20]. Here, a shared-based protection strategy is adopted. Link-disjoint backup path
is constructed for each source–destination pairs in the primary tree. The backup path
can be shared to protect several light trees under single-link failure assumption. As
the OPP_SDP is a state-of-the-art and widely accepted scheme, our objective is to
compare the proposed algorithm with the same. The future scope of this work is to
implement the proposed algorithms in EON for better results may be obtained in
terms of resource utilization.

[Link]@[Link]
Protecting Multicast Sessions Against Single-Link … 73

C. Our Contribution
In this paper, we propose three routing algorithms for protection of multicast con-
nection in WDM network. The Limited Splitting Spanning Tree (LSST ) Scheme
aims at minimizing the number of branching in the multicast tree with minimum
increase in resource utilization. This is because more the number of branching
more is the number of destination nodes that appear as the leaf nodes in the mul-
ticast tree. Thus, more network resource is required to protect the multicast tree.
Then it aims at linking each leaf node to the source node by a path which is link-
disjoint to the primary light-tree. Our second algorithm is an enhanced version
of LSST Scheme, called enhanced LSST scheme (ELSST ) which uses the same
concept as the LSST Scheme in order to derive the primary light-tree. However, it
aims at maximizing self-sharing between the backup path and the primary light-
tree. Our third algorithm, the Circular Path Session Protection (CPSP) Scheme
tries to put all the participating nodes of a particular session (source and destina-
tions) into a cycle, thus providing effective use of backup resources to establish
a protected multicast session. In each of the above-mentioned algorithms, we
use two different concepts for evaluating the usage cost of a link:

• Length of the link or other physical costs (i.e., hop count, etc.) of getting data from
one end to the other.
• Usage of the link. This means that the more the number of wavelengths in a link
is used, the more costly it is.

We also evaluate the performance of an existing algorithm, OPP_SDP [10], which

we believe to be a state-of-the-art algorithm, based on the above two cost parameters.
However in [10], the authors have evaluated their algorithm only on the basis of
physical link cost.
The organization of the paper is as follows. Section 2 describes the assumptions
and calculation of link usage cost for the schemes. Section 3 illustrates the proposed
algorithms in details. Results and discussions are presented in Sect. 4. Finally, Sect. 5
concludes the paper with future scope of the work.

2 Problem Formulation and Assumptions

A. Problem Definition
The problem of protecting multicast sessions in WDM Networks can be stated
as follows:

• The network is given as G  (V , E) where G is a weighted undirected graph 1.

V is the set of vertices or network nodes. E is the set of edges or optical fiber
links, each link has a capacity of carrying a maximum of W wavelengths.
Each link x ∈ E has a positive cost assigned to it representing the cost of
moving traffic from one end to the other.

[Link]@[Link]
74 B. Mohapatra and K. Sain

• We also make the assumption that each node is equipped with switches capable
of performing full wavelength conversion (MWRSs) and light splitters.
• We represent a multicast session as R(s, D, k) where s is the source, D is the
set of destinations, and k is the group size (including the source), i.e. (no of
Destinations served + 1). The connections are considered to be unidirectional
from the source to the destination. If the routing algorithm is able to discover
sufficient resources to establish the protected multicast session, only then the
session is established else it is considered to be a blocked session.

Our objective is to maximize the number of protected multicast sessions with

minimum possible consumption of network resources.
B. Link Usage Cost
As mentioned earlier, we calculate the link usage cost in two different ways:

• Taking into consideration the physical length of the link or other physical costs
(such as hop count, etc.) for getting the data from a node on one end of the
link to the node on the other end.
• Taking into consideration, the fraction of the number of available wavelengths
used up on the link by existing sessions.

While taking physical cost of the link as the link usage cost, the user might
consider the actual physical cost of data transmission through the link. However,
hop count is considered as the measure of distance between any two nodes in this
work. Thus, we have taken the weight of every edge to be one. While evaluating
the link usage cost on the basis of the number of used-up bandwidths of a link,
we calculate the link usage cost as follows: We have used the above measure
of link cost in order to stop the quick use-up of available wavelengths in a busy
link. Even though this method leads to slightly costlier paths, it helps to spread
the load almost evenly over the whole network. An optimal link cost measure
would be to calculate the link cost based on both physical cost and availability
of wavelengths on the link.
Wij − Aij
Cij  × 100
Wij

where C ij is the link usage cost of the link (i, j). W ij is the capacity of the link (i,
j) in terms of number of wavelengths. Aij is the number of available wavelengths
on the link (i, j).

3 Proposed Schemes

The LSST (Limited Splitting Spanning Tree) Scheme

When a multicast request arrives, apart from allocating resource for establishing

[Link]@[Link]
Protecting Multicast Sessions Against Single-Link … 75

the session, some extra resources have to be reserved to protect it against single-link
failure. The approach to the problem is achieved in two major steps. Firstly, establish a
multicast tree while minimizing the number of leaves. Reducing the number of leaves
generally tends to increase the cost of the tree. We try to keep this as minimum as
possible. Then, the scheme protects the multicast tree against single-link failure by
allocating some extra resources to the respective session.
Step 1 Allocating Resources for the Primary Light-tree—The objective while cal-
culating the primary light-tree is to reduce the number of branching while
ensuring that the excessive network resource is not used. In order to ensure
this, we use a modified form of Dijkstra’s algorithm for single source single
destination shortest path. Here we modify the algorithm to find the shortest
path between two nodes, one belonging to a set of source nodes and the
other belonging to a set of destination nodes. However if there are two sets
of path having equal length, the one with the source that was more recently
found is considered. Once a path is found between two nodes a and b, where
a ∈ S (set of sources) and b ∈ D (set of destinations), b is removed from
the set D and added to the set S. By repetitively applying this, the primary
light-tree is derived.
Step 2 Allocating Resources for the Backup Light-tree—In order to allocate
resources for the backup light-tree, we use a traditional approach. Simply
by joining the leaf nodes to the source node using paths that are link-disjoint
to the Primary Light-tree.

The LSST Scheme

Finding Primary Light-tree

(1) Q ← Sour ce
(2) D ← Destinations
(3) While (D  ∅ AN D test  1)
(a) ∀x ∈ Q, set d[x] ← 0
(b) P ← ∅
(c) while (u ∈/ D AN D P  V )
(i) u ← E xtract_Min(d)
(ii) P ← P ∪ {u}
(iii) ∀y ∈ Ad j(u) AN D y ∈/ P
If d[y] > (d[u] + w(u, y))
(A) d[y] ← (d[u] + w(u, y))
(B) Pr ed[y] ← u
(d) If (u ∈ D)
(i) D ← D − {u}
(ii) Q ← Q ∪ {u}
(e) If (u ∈
/ D), set test  1

[Link]@[Link]
76 B. Mohapatra and K. Sain

Finding Back-path
findBackupTree_LSST(AdjMatrix, PrimaryTree, Source)
(1) D ← lea f N odes(Primar y T r ee)
(2) ∀x ∈ D

(a) Path ← graphshor test path(Primar yT r ee, Sour ce, x)

(b) ad j ← Ad j Matri x
(c) ∀{y, z} ∈ Path
(i) ad j[y, z] ← 0
(ii) ad j[z, y] ← 0
(d) B_Path ← cstgraphshor test path(ad j, Sour ce, x)

The Enhanced LSST (ELLST) Scheme

The ELLST Scheme deals with the problem of protecting multicast sessions in almost
the same way as in the LSST Scheme. However this time we not only try to maximize
the amount of self-sharing, we also try to find backup paths for the leaf nodes that
use less network resource. We protect the session by connecting leaf nodes not only
to the source, but to “already protected” leaf nodes too. This algorithm too uses a
two-step approach to protect a multicast session.
Step 1 Allocating Resources for the Primary Light-tree—In order to allocate
resources for the primary light-tree, the algorithm use the same approach
that is used in the case of LSST scheme.
Step 2 Allocating Resources for the Backup Light-tree—In order to allocate reserve
the resources for the backup light-tree, a leaf node l1 is randomly chosen
from the set of leaf nodes L. Then the shortest path between the source s
and the leaf node l1 along the primary light-tree P1 is calculated. All the
links on the primary light-tree that are not in P1 are set to negligible cost
(in order to encourage self -sharing while guaranteeing that uselessly long
paths are not formed). A backup path which is link-disjoint to P1 joining s
and l1 is found. l 1 is removed from the set L, while the source node s and l1
is added to the set S. Again a leaf node is randomly chosen from the set L,
say l 2 . Again the shortest path between the source s and this node l2 along
the primary light-path P2 is calculated. Now the shortest path, link-disjoint
to P2 between l 2 and any node in the set S is calculated. Then l2 is removed
from the set L and added to the set S. This is continued until backup path for
all leaf nodes are found.

The ELSST Scheme

Finding Primary Light-tree

1. Q ← Sour ce
2. D ← Destinations
3. While (D  ∅ AN D test  1)

[Link]@[Link]
Protecting Multicast Sessions Against Single-Link … 77

(a) ∀x ∈ Q, set d[x] ← 0

(b) P ← ∅
(c) while (u ∈/ D AN D P  V )
(i) u ← E xtract_Min(d)
(ii) P ← P ∪ {u}
(iii) ∀y ∈ Ad j(u) AN D y ∈/ P
If d[y] > (d[u] + w(u, y))
(a) d[y] ← (d[u] + w(u, y))
(b) Pr ed[y] ← u
(d) If (u ∈ D)
(i) D ← D − {u}
(ii) If Q ← Q ∪ {u}
(e) If (u ∈
/ D), set test  1

Finding Back-path
findBackupTree_ELSST(AdjMatrix, PrimaryTree, Source)

1. D ← lea f N odes(Primar yT r ee)

2. S ← Sour ce
3. ∀x ∈ D

(a) Path ← graphshor test path(Primar yT r ee, Sour ce, x)

(b) ad j ← Ad j Matri x
(c) ∀link ∈ Primar yT r ee, set linkCost  ε
(d) ∀{y, z} ∈ Path
(i) ad j[y, z] ← 0
(ii) ad j[z, y] ← 0
(e) B_Path ← cstgraphshor test path(ad j, Sour ce, x)
(f) S ← S ∪ {x}
(g) D ← D − {x}

cstgraphshortestpth(adj, Sources, Destinations)

(1) Q ← Sour ce
(2) D ← Destinations
(3) ∀x ∈ Q, set d[x] ← 0
(4) P←∅
(5) while (u  D)

(a) u ← E xtract_Min(d)
(b) P ← P ∪ {u}
(c) ∀y ∈ Ad j(u) AN D y ∈ / P
(i) If d[y] > (d[u] + w(u, y))
(A) d[y] ← (d[u] + w(u, y))
(B) Pr ed[y] ← u

[Link]@[Link]
78 B. Mohapatra and K. Sain

The CPSP (Circular Path Session Protection) Scheme

The CPSP Scheme deals with the problem protecting multicast sessions by cre-
ating a cycle containing the source and all the destination nodes. However, an exact
method of creating such a cycle is an NP-complete problem. Hence, a heuristic
approach is proposed to create such a cycle.
The Algorithm
Initially, the algorithm starts with finding the shortest distance between the source
node s and the destination node d 1 from the set of destination nodes participating in
the multicast session. The nodes s and d 1 are added to a set S and are linked by the
shortest path. The edges on this path are removed from the network. Then we try to
find the node in the remaining set of destination nodes that is the shortest distance
away from any of the two nodes in S. Let this destination node be d 2 .
Then d 2 is added to the set S and the node in S that d 2 is found to be shortest
distance away, is to be removed from the set S. Hence, a single path is derived by
joining an unlinked destination node to any of the nodes at either end of the current
path. Also new paths found must be link-disjoint to the currently existing path. Once
all the destination nodes and the source node are linked by one path, the nodes at
the two ends are linked, which completes the cycle. This guarantee that even with a
single-link failure, data will continue to flow from the source to all the destination
nodes.

The CPSP Scheme

FindSingleCycle(AdjArr, Source, Destinations)

(1) Q ← Sour ce
(2) D ← Destinations
(3) while (D  ∅ AN D test  1)
(a) ∀x ∈ Q, set d[x] ← 0
(b) P ← ∅
(c) while (u ∈ / D AN D P  V )
(i) u ← E xtract_Min(d)
(ii) P ← P ∪ {u}
(iii) ∀y ∈ Ad j(u) AN D y ∈/ P
If d[y] > (d[u] + w(u, y))
(A) d[y] ← (d[u] + w(u, y))
(B) Pr ed[y] ← u
(d) if (u ∈ D)
(i) D ← D − {u}
(ii) Q ← Q ∪ {u}
(iii) If (length(Q) > 2)then Q ← Q − {Pr ed[u]}
(iv) Remove the path from Pr ed[u]to u f r om the Ad j Arr

(4) Find the shortest path connection Q [1] and Q[2]

[Link]@[Link]
Protecting Multicast Sessions Against Single-Link … 79

4 Results and Discussions

In this section, we present a series of simulation results to evaluate the performance

of the proposed schemes (LSST , ELSST, and CPSP). We also compare the results
against the performance of a standard state-of-the-art algorithm OPP_SDP. The
above schemes are simulated using two widely used networks such as USNet and
NSFNet. It is assumed that all the nodes in the networks are capable of wavelength
conversion and light splitting. Each link in the network is considered to have a
maximum capacity of W wavelengths. Multicasts requests of size k are assumed to
arrive with Poisson’s distribution with arrival rate λ. Each request is assumed to have
unit holding time. Thus, the average load of the network is λ. Source and destination
nodes are randomly distributed across the whole network. The simulation results are
as shown in below given figures. Figures 1, 2 show the results obtained for NSF Net
with k  4 and k  6; and value of λ varied from 10 to 60 for link capacity (W ) 
16; and value of λ varied from 10 to 100 for link capacity (W )  32.
As is seen from all the above results, our algorithm performs quite better as
compared to OPP_SDP in all the cases, especially under heavily loaded conditions.
CPSP performs the best under heavily loaded conditions (though it gives slightly
high blocking probability under very lightly loaded conditions). It is also seen that
the blocking probability is slightly less.
While using Link Usage as the metric while allocating the light-paths as compared
to while using hop count as the metric. Figures 3, 4 show the results obtained for
USNet with k  6 and k  8; and value of λ varied from 10 to 60 for link capacity
(W )  16; and value of λ varied from 10 to 100 for link capacity (W )  32. The results
obtained for USNet are similar to those obtained for NSFNet.

Fig. 1 a Session size k  4 with link capacity W = 16 for NSFNet, b Session size k  6 with link
capacity W = 16 for NSFNet

[Link]@[Link]
80 B. Mohapatra and K. Sain

Fig. 2 a Session size k  4 with link capacity W = 32 for NSFNet, b Session size k  6 with link
capacity W = 32 for NSFNet

Fig. 3 a Session size k  6 with link capacity W = 16 for USNet, b Session size k  8 with link
capacity W = 16 for USNet

5 Conclusions

In this paper, three schemes have been proposed to protect multicast session against
single-link failure. The first scheme, LSST tries to reduce the network resource usages
and in turn reduce blocking probability by reducing the number of splitting points
in the primary light-tree that is to be protected. The second scheme, ELSST is an
enhancement of the LSST scheme which tries to increase the degree of self -sharing
among the primary light-tree and the backup paths for each of the failed link. The
third scheme, CPSP approaches the problem of protecting multicast session using a
single cyclic light-path. The results obtained from simulations performed on NSFNet
and USNet indicate that the proposed algorithms provide significant improvement
over a state-of-the-art existing algorithm OPP_SDP in terms of reduction in blocking
probability.

[Link]@[Link]
Protecting Multicast Sessions Against Single-Link … 81

Fig. 4 a Session size k  6 with link capacity W = 32 for USNet, b Session size k  8 with link
capacity W = 32 for USNet

In the case of NSFNet, when the link capacity W = 16, the improvement offered
by the above schemes are 6–7, 8–9, and 10–11%, respectively. Whereas when the
link capacity W = 32, the improvement appeared to be 8–10, 10–12, and 13–14%,
respectively. On the other hand, in the case of USNet, when the link capacity W = 16,
the improvement offered by the above schemes are 10–12, 12–15, and 15–16%,
respectively. Whereas when the link capacity W = 32, the improvement appeared to be
11–13, 13–16, and 16–17%, respectively. It is observed that the performance of CPSP
scheme is best scheme among the others and it is observed that this improvement is
achieved at the cost of higher working capacity or to higher primary light-tree cost.

References

1. Sahasrabuddhe, L., Mukherjee, B.: Light trees: optical multicasting for improved performance
in wavelength routed networks. Commun. Mag. IEEE 37, 67–73 (1999)
2. Hu, W., Zeng, Q.: Multicasting optical cross connects employing splitter-and-delivery switch.
Photonics Technol. Lett. IEEE 10, 970–972 (1998)
3. Ramamurthy, S., Sahasrabuddhe, L., Mukherjee, B.: Survivable wdm mesh networks. J. Light-
wave Technol. 21, 870 (2003)
4. Ellinas, G., Hailemariam, A., Stern, T.: Protection cycles in mesh wdm networks. IEEE J. Sel.
Areas Commun. 18, 1924–1937 (2000)
5. Van Caenegem, B., Van Parys, W., De Turck, F., Demeester, P.: Dimensioning of survivable
wdm networks. IEEE J. Sel. Areas Commun. 16, 1146–1157 (1998)
6. Crochat, O., Le Boudec, J.-Y.: Design protection for wdm optical networks. IEEE J. Sel. Areas
Commun. 16, 1158–1165 (1998)
7. Singhal, N.K., Mukherjee, B.: Protecting multicast sessions in wdm optical mesh networks. J.
Lightwave Technol. 21, 884 (2003)
8. Rahman T., Ellinas, G.: Protection of multicast sessions in wdm mesh optical networks. In:
Optical Fiber Communication Conference, 2005. Technical Digest. OFC/NFOEC, vol. 2, 3 pp,
2005

[Link]@[Link]
82 B. Mohapatra and K. Sain

9. Khalil, A., Hadjiantonis, A., Ellinas, G., Ali, M.: Dynamic provisioning of survivable hetero-
geneous multicast and unicast traffic in wdm networks. In: IEEE International Conference on
Communications, 2006. ICC ’06, vol. 6, pp. 2465–2470, 2006
10. Singhal, N.K., Sahasrabuddhe, L.H., Mukherjee, B.: Provisioning of survivable multicast ses-
sions against single-link failures in optical wdm mesh networks. J. Lightwave Technol. 21,
2587 (2003)
11. Singhal, N.K., Ou, C., Mukherjee, B.: Cross-sharing versus self-sharing trees for protecting
multicast sessions in mesh networks. Comput. Netw. 50(2), 200–206 (2006). (Optical Net-
works)
12. Wu, C.-S., Lee, S.-W., Hou, Y.-T.: Backup vp preplanning strategies for survivable multicast
atm networks. In: 1997 IEEE International Conference on Communications, ICC 97 Montreal,
‘Towards the Knowledge Millennium’, vol. 1, pp. 267–271, 1997
13. Luo, H., Li, L.: A new consideration for provisioning survivable multicast sessions in WDM
mesh networks. In: Proceedings of International Conference on Communications, Circuits and
Systems, Kokura, Japan, pp. 544-548, 2007
14. Luo, H., Li, L., Yu, H., Wang, S.: Achieving shared protection for dynamic multicast sessions
in survivable mesh WDM networks. IEEE J. Sel. Areas Commun. 25(9), 83–95 (2007)
15. Xiong Wang · Sheng Wang · Lemin Li · Yunji Song, “Achieving resource reduction for pro-
tecting multicast sessions in WDM mesh networks”, Photon Netw Communication, Vol. 15,
pp. 131–140, 2008
16. Long, L., Kamal, A.E.: Tree-based protection of multicast services in WDM mesh networks.
In: Proceedings of 28th IEEE conference on Global telecommunications, GLOBECOM’09,
pp. 1529–1534, 2009
17. Feng, T., Ruan, L., Zhang, W.: Intelligent p-cycle protection for multicast sessions in WDM net-
works. In: Proceedings of International Conference on Communications (ICC), pp. 5165–5169,
2008
18. Feng, T., Ruan, L., Zhang, W.: Intelligent p-cycle protection for dynamic multicast sessions in
WDM networks. J. Opt. Commun. Netw. 2(7), 389–399 (2010)
19. Frikha, A., Cousin, B., Lahoud, S.: Extending node protection concept of p-cycles for an
efficient resource utilization in multicast traffic. In: Proceedings of IEEE 36th Conference on
Local Computer Networks (LCN), Bonn, Germany, pp. 175–178, 2011
20. Cai, A., et al.: Multicast routing and distance-adaptive spectrum allocation in elastic optical
networks with shared protection. J. Light. Technol. 34(17), 4076–4088 (2016)

[Link]@[Link]
A Cross-Layer Routing Protocol
for Wireless Sensor Networks

Pallavi Yarde, Sumit Srivastava and Kumkum Garg

Abstract In the era of Internet of things (IoT), the sensors play an important role
and also face a challenge of energy consumption. Sensors in wireless sensor net-
works (WSNs) deals with accumulation and processing of data and forward that to
the remote locations, generally considered as cloud. Generally, communication is
done between the nodes which are placed at a far locations in the field. Hence,
the energy consumption required to communicate the nodes plays an important
role. In this paper, the proposed algorithm is based on low-energy adaptive clus-
tering hierarchical (LEACH) routing algorithm named as multi-hop cluster LEACH
(MC LEACH) algorithm. The proposed protocol is a cross-layer routing protocol that
deals with physical, MAC, and network layers for the analysis of energy consumption
at individual node as well as in whole network.

Keywords Wireless sensor networks · Cross-layer optimization · LEACH

MC LEACH

1 Introduction

There is a big resolution occurred with Internet of things (IoT) in the field of wireless
sensor networks (WSNs). The aim of IoT is to communicate the data gathered into
sensors network to remote location with one of wireless technologies. The data
from network should reach to remote location, generally called sink with minimum
delay and energy consumption. Delay and energy consumption are very important
quality of service (QoS) measuring metrics. There are numerous delay measuring
protocols available for WSN [1–4]. The proposed work focuses on optimizing energy
dissipation in whole network. Each layer of communication network model plays a

P. Yarde (B) · S. Srivastava

Manipal University Jaipur, Jaipur, Rajasthan, India
e-mail: [Link]@[Link]
K. Garg
Bhartiya Skill Development University, Jaipur, Rajasthan, India

© Springer Nature Singapore Pte Ltd. 2019 83

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
84 P. Yarde et al.

vital role. When these different layers work in collaboration then it is considered as
cross-layer optimization. The proposed protocol also implemented using cross-layer
approach. Here, three layers, network, media access control (MAC), and physical
layers, work in coordination to analyze energy consumption, number of dead nodes,
and number of alive nodes. The proposed protocol is based on low-energy adaptive
clustering hierarchy (LEACH) protocol with some add-ons. The function of the
proposed protocol refers to the probability as well as distance for selecting cluster
head. The cluster head selection process involves probability, distance among the
nodes and sink, and also energy of nodes. The number of cluster heads makes a chain
to propagate the processed data to remote sink, hence called multi-hop protocol. The
proposed protocol MC LEACH involves functions like selection process for cluster
heads, gathering data, processing of data, and designing a chain, i.e., choosing the
best path to reach to the sink. These functions involves role of network layer, MAC
layer, and physical layer in collaboration to as a cross-layer function.

2 Existing Work

There exist many protocols which are energy efficient and also focus on cross-layer
optimization. These protocols use different QoS metrics as measures of performance
like delay [5–8], throughput, network lifetime, residual energy [9–14]. Some of the
protocols also implemented with cross-layer optimization. The proposed protocol
is based on LEACH protocols with some modifications. Most of sensor nodes in
WSN protocols directly transmit accumulated data to the sink instead; this protocol
supports selecting a cluster head to each designed clusters in rotation. The proposed
protocol also supports cross-layer optimization.
Packet size: The amount of data (raw data and supporting information) in bits
used to travel between source and destination is considered as a packet. The size of
packet may vary or of fixed size.
Traffic: Highly dense network has too many packets transmission and hence some
of packets may have a chance to drop. This is because of collision or stop-and-wait
situation at the node. This introduces a delay in communication.
Scalability: It is an ability to enhance number of nodes in the network. The number
of nodes should be enhanced up to optimization level only, so that maximum delay
cause can be avoided there.
An energy-efficient differentiated directed diffusion (EDDD) protocol was pro-
posed in 2006 [15]. The data in network may be destined to the node from its sur-
rounding or from a node to the other node. The first type of data is considered as
real-time (RT) data, while other one is considered as best-effort (BE) data. This pro-
tocol has a capability to differentiate between these two data and work accordingly
to avoid maximum delay.
A cross-layer priority-based congestion control protocol (PCCP) was proposed by
Wang et al. [16]. It is a routing protocol. Here, there are two data queues generated:
one at MAC layer and other at network layer. The node calculates the time consumed
by a packet to wait and communicate and estimates the probability of congestion and

[Link]@[Link]
A Cross-Layer Routing Protocol for Wireless Sensor Networks 85

actual congestion happened. Based on this, it notifies other nodes about congestion
and avoids further delay in communication.

3 Proposed Work

The proposed protocol, multi-hop clusters LEACH (MC LEACH), is a cross-layer

optimized routing protocol which focuses on media access control) (MAC) and net-
work layers. The proposed protocol is a cross-layer routing protocol, which empha-
sizes minimized energy consumption for communication.

3.1 Network Structure

The different energy-level sensor nodes are considered to be randomly spread in

bounded area and considered as heterogeneous network. Whole network nodes are
divided into normal and advanced nodes. The complete network is virtually divided
into number of clusters with each has a radius R. A sink is also placed at the center of
the network. Sink received sensed data from network. Sending sensed data directly
to the sink may exhaust sensor nodes easily. The proposed protocol follows LEACH,
in which each cluster is appointed to of the node as a head called cluster head
(CH). Here, the responsibility of CH is to gather the data from surrounding sensor
nodes of its own cluster and to process that data and finally send it to the sink
[17]. In MC LEACH protocol, each sensor node calculates the distance between
all the remaining nodes and sinks. When a node is elected as CH, it already knows
the distance between itself and the sink as well as between all available CHs of the
network. The distance is calculated using the Euclidian distance formula.
In Fig. 1, each cluster head calculates the distance between all remaining cluster
heads and the sink. Here, we have four clusters C1, C2, C3, and C4. Each cluster has
a cluster head (black circles). The distances between all cluster heads are d23, d12,
d13, d24, d34, and d14. The distances between clusters C1, C2, C3, C4 and the sink
are d1, d2, d3, and d4, respectively.

3.2 Cluster Setup and Selection of Cluster Head

Each sensor node is initiated with a specific amount of energy. The selection of
cluster head in a cluster depends on the probability like LEACH protocol and on
energy of individual nodes; i.e., the node with highest probability and energy among
the cluster nodes will be selected as CH.
Each node manages a routing table. The routing table contains information of its
own and neighboring nodes of its cluster. The parameters stored with routing table are

[Link]@[Link]
86 P. Yarde et al.

Fig. 1 Architecture of wireless sensor networks for MC LEACH protocol

probability to be cluster head, residual energy, and distance between the nodes. Like
LEACH protocol, the node becomes cluster head on the basis of probability and also
with an additional parameter residual energy. The node with maximum probability
and residual energy among the nodes in a cluster will be selected as cluster head.
Each node selects a random number where 0 < r < 1. If, r < threshold T (n), then one
of the criteria for selecting that node as cluster head fulfills. The threshold calculation
is given in Eq. 1.
⎧
⎨ P  n ∈ G
1 −P * r mod 1P
T (n) = (1)
⎩
0 Otherwise

where n is nth node for which threshold is calculated, G is the group of sensor nodes,
and P is the probability of node to become cluster head.
Once the criteria of threshold fulfills, the energy of sorted nodes is compared
and the node with maximum energy and threshold greater than value r is selected as
cluster head for a round. This process keeps repeating for each round; i.e., for each
round, number of clusters and cluster head are created.

[Link]@[Link]
A Cross-Layer Routing Protocol for Wireless Sensor Networks 87

3.3 Cluster Setup

Once the cluster is formed, the responsibility of sensor nodes is to introduce them-
selves with the other nodes of cluster. The nodes broadcast message signal consisting
of node details like their node id and residual energy. Based on receive signal strength
indicator (RSSI) [18], the node calculates the distance between the message sending
node and itself.
Each cluster of MC LEACH protocol has a cluster head and sensor nodes. The
cluster setup details for MC LEACH are shown in Fig. 2. Here, each node is initiated
with a specific energy level. When a cluster is formed, each sensor node broadcasts
about its presence in the network and calculates distance between itself and sink. The
calculated distance is Euclidian distance. The setup phase further works in twofolds,
intra-cluster and inter-cluster.
Intra-cluster Setup Phase Once the cluster is formed and CH is selected, the
remaining nodes of cluster must know about the CH. Hence, CH broadcasts a mes-
sage in cluster about its identification. When this message is received by remaining
nodes in cluster, they calculate the distance between themselves and CH with the
parameter receive signal strength indicator (RSSI).

Pt
Pr (d)  C (2)
d∝
where Pr (d) is the received power by receiver at distance d expressed in dBm, C
is proportionality constant whose value depends on communication model, Pt is
the power used by transmitter to send the message, and α is called distance–power

Fig. 2 Cluster setup in MC LEACH protocol

[Link]@[Link]
88 P. Yarde et al.

gradient. The broadcast signal also consists of the information of residual energy of
sensor node. Hence, a routing table is formed at each node consisting of node id,
residual energy, and distance between the nodes and itself. A cluster head is selected
for a round on the basis of Eq. 1.

Inter-cluster Setup Phase Each CH transmits their initial route to sink to check
whether any other CH is in its path to the sink. If it observes so, it updates its routing
table for designing the path to the sink. Once it is done, each sends a message to
all remaining CHs of different clusters of the network consisting of node ID and
residual energy. Again with RSSI parameter, each CH designs a routing tree keeping
itself as a root node and sink as an end node. This sorting is done by binary search
algorithm (BSA). This routing tree results in the optimized path between each CH
and the sink, and a chain is also formed to transmit the data from different CHs to
the sink.

3.4 Steady-State Phase

The proposed algorithm is based on cross-layer optimization. Three layers, physical

(PHY), media access control (MAC), and network (routing) layers, work in collabora-
tive manner to fulfill the desired criteria. Once routing tree is formed, communication
among the nodes starts.
All sensor nodes of cluster forward their sensed data to CH. CH then processes
gathered data, and finally send to the sink. Instead, all CHs of network design
the chain according to level of residual energy and distance to sink. The CH, whose
residual energy and distance to the sink are least among all CHs available in network,
will be the last node of the chain, i.e., nearest node to the sink and vice versa. If
the energy level of two nodes is equal, then second parameter, i.e., distance, is the
decision parameter. Here, to design an optimized chain route, PHY and MAC layers
play their roles. This routing table as well as routing tree will also update in parallel
with this calculation of designing the optimized chain (path) to the sink.

E T  k ∗ E cluster (3)

4 Network Model and Result

Whole experimental setup is done in MATLAB R2017a. Here, a data-gathering WSN

is considered, where N consistent sensors are randomly deployed in an area with the
sink located at the center [19]. The network is experimented with a number of sensor
nodes between 200 and 1000 to handle a thin sensor network and a very heavy one.
The simulation is done for 3000 rounds. The remaining simulation parameters are
summarized in Table 1. The path between the cluster head and sink is shown in Fig. 3.

[Link]@[Link]
A Cross-Layer Routing Protocol for Wireless Sensor Networks 89

Table 1 List of parameters considered for MC LEACH protocol

Parameters Value
Sink location for network X bs  Y bs  50
Percentage of cluster heads (P) in network 0.1
E 0 (J) 0.5
E Tx1 , E Rx1 (mJ) 1
E Tx2 , E Rx2 (mJ) 0.1
E DA (nJ/bit) 5
Maximum number of rounds 3000
E elec (nJ/bit) 50
2amp (nJ/bit) 50
2fs (nJ/bit) 10

Fig. 3 Path between CH and sink for MC LEACH protocol

The analysis of this protocol is compared with low-energy adaptive clustering hier-
archical (LEACH) protocol for residual energy and communication delay parameters
(Fig. 4).

5 Conclusion and Future Scope

MC LEACH protocol has been simulated on MATLAB and analyzed. The output
results of his protocol show that energy consumption gets affected in a positive
manner if the number of hops toward the sink can be changed rather than using a sin-
gle hop. The introduction of multi-hops communication has substantially improved
LEACH by discovering the best path between the source CH and the destination.
Simulation results in terms of dead nodes and residual energy. The proposed work
can also be analyzed for the delay. Hence, the communication can be faster and can
be implemented in delay-conscious applications.

[Link]@[Link]
90 P. Yarde et al.

Fig. 4 Comparative analysis of LEACH and MC LEACH protocol for residual energy in the net-
work

References

1. Wu, S., et al.: Delay-aware energy-efficient routing towards a path-fixed mobile sink in indus-
trial wireless sensor networks. Sensors 3, 899 (2018)
2. Jabbar, S., et al.: Analysis of factors affecting energy aware routing in wireless sensor network.
Wirel. Commun. Mob. Comput. 1, 1–21 (2018)
3. Jin, Z., et al.: A Q-learning-based delay-aware routing algorithm to extend the lifetime of
underwater sensor networks. Sensors 17(1660), 1–15 (2017)
4. Kang, M.W., Chung, Y.W.: A novel energy-aware routing protocol in intermittently connected
delay-tolerant wireless sensor networks. Int. J. Distrib. Sens. Netw. 13(7) (2015)
5. Al-Anbagi, I., et al.: A survey on cross-layer quality of service approaches in WSNs for delay
and reliability-aware applications. IEEE Commun. Surv. Tutorials 18(1), 525–552 (2014)
6. Khiati, M., Djenouri, D.: BOD-LEACH: broadcasting over duty-cycled radio using LEACH
clustering for delay/power efficient dissimilation in wireless sensor networks. Int. J. Commun.
Syst. 28, 296–308 (2015)
7. Rao, Y., Deng, C., Zhao, G., Qiao, Y., Fu, L., Shao, X., Wang, R.: Self-adaptive implicit
contention window adjustment mechanism for QoS optimization in wireless sensor networks.
Elsevier J. Netw. Comput. Appl. 109, 36–52 (2018)
8. Qaisar, S., Bilal, R.M., Iqbal, W., Naureen, M., Lee, S.: Compressive sensing: from theory to
applications, a survey. J. Commun. Netw. 15(1), 443–456 (2013)
9. Braman, A., Umapathi, G.R.: A comparative study on advances in LEACH routing protocol for
wireless sensor networks: a survey. Int. J, Adv. Res. Comput. Commun. Eng. 3(2), 5683–5690
(2014)

[Link]@[Link]
A Cross-Layer Routing Protocol for Wireless Sensor Networks 91

10. Khadivi, A., Shiva, M.: FTPASC: a fault tolerant power aware protocol with static clustering for
wireless sensor networks. IEEE International Conference on Wireless and Mobile Computing,
Networking and Communications, IEEE, 2006
11. Yarde, P., Srivastava, S., Garg, K.: A modified energy efficient protocol for optimization of
dead nodes and energy consumption in wireless sensor networks. In: IEEE, 11th International
Conference on Sensing Technology (ICST-2017), pp. 31–36, 2017
12. Akkaya, K., Younis, M.: A survey on routing protocols for wireless sensor networks. Ad hoc
Netw. 3(3), 325–349 (2005)
13. Sohraby, K., et al.: Protocols for self-organization of a wireless sensor network. IEEE Pers.
Commun. 7(5), 16–27 (2000)
14. Younis, M., Youssef, M., Arisha, K.: Energy-aware routing in cluster-based sensor networks.
In: Proceedings of 10th IEEE/ACM International Symposium on Modeling, Analysis and
Simulation of Computer and Telecommunication Systems (MASCOTS-2002), 2002
15. Yuan, Y., He, Z., Chen, M.: Virtual MIMO-based cross-layer design for wireless sensor net-
works. IEEE Trans. Veh. Technol. 55(3), 856–864 (2006)
16. Wang, C., Li, B., Sohraby, K., Daneshmand, M., Hu, Y.: Upstream congestion control in
wireless sensor networks through cross-layer optimization. IEEE J. Sel. Areas Commun. 25(4)
(2007)
17. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: An application-specific protocol archi-
tecture for wireless microsensor networks. IEEE Trans. Wirel. Commun. 1(4), 660–670 (2002)
18. Zheng, J., Liu, Y., Fan, X., Li, F.: The study of RSSI in wireless sensor networks. In: 2nd
International Conference on Artificial Intelligence and Industrial Engineering (AIIE2016),
Advances in Intelligent Systems Research, vol. 133, pp. 207–209, 2016
19. Neto, J.H.D., Rego, A.D., Cardoso, A.R., Colestino, J.: MH-LEACH: a distributed algorithm
for multi-hop communication in wireless sensor networks. In: ICN 2014: The Thirteenth Inter-
national Conference on Networks, IARIA, pp. 55–61, 2014

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive
Badminton Playing Robot

Shivam Mishra and V. K. Sachan

Abstract There are many researchers and engineers who aim to design such robots
that can outplay the human players in the field of sports. A design of badminton
playing robot is presented by this paper which can be controlled by the PS-2 (play
station) controller, and it can also serve shuttlecock in the badminton court similar to a
human player. It is done so as to incorporate advancement in the field of sports. It aims
at designing a configuration which can assist the players in practicing badminton.
A shuttlecock holding system is designed using pneumatics which can be preloaded
with three shuttlecocks and can further be increased according to the convenience. A
hitting mechanism was designed so that a standard badminton racket can be swung
and can be used to hit the shuttlecocks. The problems and all the related solutions that
were involved in the designing of this robot are discussed. The practical application
of the design indicates that this robot is able to hold three shuttlecocks and can serve
with high accuracy and is efficient enough to hit the returning shuttlecocks with high
accuracy. This means that a robust robot is designed which can be controlled by any
human being.

Keywords Badminton · Motor control · Mobile robot · PS-2 controller

Serving shuttlecock · Microcontroller

1 Introduction

Badminton is a racquet sport played using racquets to hit a shuttlecock across a net.
Badminton is one of the most popular sports in the world even it is given priority in
almost all the countries. Some of the researchers are currently working on several
experiments and theories to make advancements in this field [1–4]. Bringing more
and more technology in the field of sports will definitely increase the accuracy and
will open up a lot of opportunities for the engineers and the researchers.

S. Mishra (B) · V. K. Sachan

KIET Group of Institution, Ghaziabad, India
e-mail: mishra.shiva999@[Link]

© Springer Nature Singapore Pte Ltd. 2019 93

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
94 S. Mishra and V. K. Sachan

A design is presented for the PS-2 controlled badminton playing robot. Several
subsystems are integrated together to make the complete robot. The robot can play
badminton similar to the human player; it can serve the shuttlecock with high accu-
racy and can also intercept the oncoming shuttlecocks from the opponent. Many
robotic competitions like RoboCup is organized so as to motivate the engineers and
researchers as well as to acknowledge their contribution and efforts in the field of
robotics. Also, the sports robots can also be designed for the learning and teaching
purposes [5]. A lot of experiments are going on in developing the robots which can
play entertainment games [6].
Badminton is the most dynamic game as the maximum speed of the smash ever
recorded is 332 kph in the 2005 Sudirman Cup in Beijing. A player needs very
quick and accurate instincts to counter back such high-speed shuttlecocks. The robot
should be able to predict the exact trajectory of the shuttlecock and should anticipate
according to the prediction accurately so as to achieve the performance equivalent
to that of human players.
Modern technologies provide many alternatives in the selection of drive system. It
depends on the application whether the fixed speed drives are used or variable speed
drive system is used. Here, variable speed drive system has been used to drive the
robot efficiently according to the oncoming shuttlecock from the opponent player.
The integration of several subsystems of various engineering sciences to make
the robust badminton playing robot is one of the challenges that are prominent. The
most important part of the robot is the integration of mechanical design and hardware
subsystems with the software subsystems with the help of embedded systems on
various computing platforms. These prominent challenges provide opportunities to
the researchers to do several experiments to enhance the technology [7].
The paper investigates all the parameters involved in completing this robot. Firstly,
the construction of the mechanical part is defined and then the driving electronics
circuitry and finally, the implementation of the programming part are defined [8]. As
the acceleration and the velocities of the robot and the badminton rackets are very
high, so the safety issues are also concerned. So, the preference is given to the safe
technology materials as well as parameters. Also, it has future scopes and one can
replace the traditional model with some advanced configurations.

2 Methodology

See Fig. 1.

2.1 Mechanical Structure

The structure is made up of aluminum rods with the support of nylon fiber. The
structure stands on a four wheel base drive which is mounted with the DC motor.

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot 95

Fig. 1 Block diagram

The four wheel base drive is the combination of Omni wheels and aluminum rods.
The effect is that the wheel can be driven with full force, but will also slide laterally
with great ease. The linear actuation to the badminton rackets is provided with the
help of pneumatic actuators. A PVC (Polyvinyl Chloride) pipe is used to make the
shuttlecock holding mechanism.

2.1.1 Serving Mechanism

The serving mechanism consists of the shuttlecock holder and a racket for serving the
shuttlecock. The shuttlecock holder is made with the help of PVC pipe. Three slots
are made in the PVC pipe and small pneumatic actuators are mounted so as to hold
the shuttlecocks in position. When the signal is provided from the transmitter then the
lowermost pneumatic actuator releases the Shuttlecock and the serving racket hits the
shuttlecock with the required force. The time lapse between releasing the shuttlecock
and reaching to the serving racket is calculated and is implemented accordingly so
that the racket does not miss the shuttlecock while serving (Fig. 2).

2.1.2 Four Wheel Base Drive

The base is made up of aluminum rods and Omni wheels. The Omni wheels are
mounted with DC (Direct Current) motors of 800 rpm (rotation per minute) so as to
move the robot throughout the standard badminton court effectively. The structure of
the base has the strength to hold 60 kg of weight efficiently. The effect of the Omni
wheels is such that the robot can be driven with full force and can also slide laterally
with great ease (Fig. 3).

2.1.3 Hitting Mechanism

The hitting mechanism is made up of the pneumatic actuators. There are three pneu-
matic actuators which are mounted with the badminton rackets on the opposite side
of the serving mechanism. These are used so as to provide the linear actuation to
the badminton rackets. Using this linear actuation and the force of the pneumatic
cylinder the badminton racket hits the shuttlecock. There are two rackets at the lower

[Link]@[Link]
96 S. Mishra and V. K. Sachan

Fig. 2 Serving mechanism

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot 97

Fig. 3 Four wheel base drive mounted with DC motors

side and one to make the upper hit. The robot has to be adjusted according to the
oncoming shuttle by the controller.
The pneumatic actuators are those kinds of actuators which work using air pres-
sure. The air pressure is filled in the black colored bottles placed on both sides of
the robot. It is prescribed to fill the bottles up with 8.5 bars of air pressure using the
air compressor. The bottles and the pneumatic actuators are connected using 4 mm
pipes. Also, a pressure gauge is mounted so as to keep an eye on the air pressure
(Fig. 4).

[Link]@[Link]
98 S. Mishra and V. K. Sachan

Fig. 4 Hitting mechanism with pneumatic actuators

2.1.4 CAD Design and Actual Robot

See Fig. 5.

2.2 Electronic Circuits

The electronic circuitry is the most important part of the robot which is used to drive
the robot. The different parts of the electronics circuitry used in the robot are as
follows:

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot 99

Fig. 5 CAD design and actual robot

2.2.1 Motor Driver

The motor drivers are used to drive the DC motor. These motor drivers are based on
the parallel MOSFET H-bridge configurations. It is used in parallel configuration so
as to increase the current rating of the motor driver. These motor drivers are designed
by us. There are four motor driver circuits used, one for each DC motor. When the
signal is provided to the motor driver from the microcontroller, it incorporates motion
in the robot and it can easily be controlled by the PS-2 controller (Fig. 6).

2.2.2 Solenoid to Switch Pneumatic Actuators

To direct the airflow into and out of the pneumatic actuators, solenoid valves are used.
The pneumatic actuator position is remotely controlled by the solenoids according
to the electrical signals provided to them. They have two states energized and de-
energized. The position in which the pneumatic actuators will be in, is controlled by
the different airflow patterns. Here, the solenoids are switched by providing 24 V dc
supply and the switching is done with the help of MOSFET switching circuit and
the signal is provided by the microcontroller (Fig. 7).

[Link]@[Link]
100 S. Mishra and V. K. Sachan

Fig. 6 Motor driver circuit

Fig. 7 Solenoid and it’s switching circuit

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot 101

2.3 Programming

Controlling the robot with the help of the PS-2 controller needs software interface.
A PS-2 controller is interfaced with the Arduino microcontroller and the signals are
transmitted to robot’s receiver section wirelessly with the use of Bluetooth module
which is also connected to the Arduino microcontroller both in the transmitter section
to transmit the signal and to the receiver section so as to receive the signal. The
programming section is basically divided into two sections as:

2.3.1 Transmitter Section

The transmitter section basically includes a PS-2 controller which is interfaced with
the Arduino microcontroller, and the signal is transmitted wirelessly using the Blue-
tooth module. In the transmitter section, the Bluetooth module acts as the master and
the microcontroller is also the master; i.e., it sends the signal to the robot. There are
nine pins in the PS-2 as follows:
1. DATA: This is the data line from controller to PS2.
This is an open collector output and requires a pull-up resistor (1–10 k). (A pull-
up resistor is needed because the controller can only connect this line to ground;
it cannot actually put voltage on the line).
2. COMMAND: This is the data line from PS2 to controller.
3. VIBRATION MOTOR POWER.
4. GND: Ground.
5. VCC: VCC can vary from 5 down to 3 V.
6. ATT: To get the attention of the controller, ATT is used. This line must be pulled
low before each group of bytes is sent/received and then set high again.
7. CLK: 500 kH/z, normally high on. SPI bus communication appears.
8. Not Connected.
9. ACK: Acknowledge signal from Controller to PS2 [9] (Fig. 8).

2.3.2 Receiver Section

The receiver section includes the Arduino microcontroller which is also connected
to the Bluetooth module, and this Bluetooth module acts as the slave. The receiver
section receives the signal coming from the transmitter and incorporates the func-
tionality into the robot accordingly (Fig. 9).

3 Block Diagram of Complete Operation

See Fig. 10.

[Link]@[Link]
102 S. Mishra and V. K. Sachan

Fig. 8 PS-2 interfaced with Arduino

Fig. 9 Receiver section

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot 103

Fig. 10 Operation of the

robot

[Link]@[Link]
104 S. Mishra and V. K. Sachan

4 Result

The challenge of holding the shuttlecock in the position is done with the help of
pneumatic actuators and three pneumatic actuators are used to hold the three shuttle-
cocks. When the serving signal is provided to the receiver section then the lowermost
pneumatic actuator retracts and the lowermost shuttlecock comes down and the bad-
minton racket hits the shuttlecock to serve. Then after a delay of 3 s, the second
shuttlecock comes to the lowermost position and the upper shuttlecock comes to the
middle position and the topmost position gets empty so that the shuttlecock can be
loaded again.
As the air pressure in the bottle keeps on decreasing with the number of hits, so
the distance of the shuttlecock is also affected as shown in Table 1 which depicts
the accuracy of the serving distance. The accuracy of serving is very high as, till the
sixteenth hit the distance of shuttlecock is still more than the distance of the foul line.
But after the sixteenth hit, the distance of the serving comes under the foul line and
again air pressure is required to be filled into the bottles to again gain the accuracy
to that level.

Table 1 Serving distance with the ascending hit number

Ascending hit number Distance of shuttlecock (cm) Foul distance (cm)
1 768 400
2 761 400
3 750 400
4 732 400
5 719 400
6 691 400
7 674 400
8 649 400
9 623 400
10 598 400
11 571 400
12 534 400
13 495 400
14 473 400
15 456 400
16 431 400
17 397 400

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot 105

Fig. 11 Success rate of serving and distance versus ascending hit number

4.1 Success Rate of Serving and Hitting

Also, the distance of the shuttlecock varies on hitting it with varying forces which is
very important during the game as the controller of the robot can throw shuttlecocks
at different positions of the standard badminton court. So Table 2 depicts the distance
of the shuttlecock with the varying forces (Figs. 11 and 12).

Table 2 Distance of shuttlecock with the varying force and the hit number
Force (lbs) Distance of shuttlecock (cm) Hit number
81.95 768 1
78.33 756 2
77.01 731 3
75.2 710 4
73.78 680 5
71.25 669 6
68.01 651 7
57.26 633 8
50.93 620 9
41.37 573 10
35.89 550 11
26.3 500 12
20.04 460 13
15.73 409 14

[Link]@[Link]
106 S. Mishra and V. K. Sachan

Fig. 12 Hit number versus force of hitting

5 Conclusion

The serving performance and the design of the robot are discussed and analyzed in
this paper. The calculation of the time lapse between releasing the shuttlecock from
the shuttlecock holder and reaching down to the serving badminton racket is crucial.
In this design, the robot is able to serve three shuttlecocks in one loading with a
very high accuracy. During service, the distance of the shuttlecock keeps decreasing
due to decrement in the air pressure but up to sixteen hits it can easily serve to a
distance more than the distance of the foul line, i.e., a valid serve in crosscourt. The
maximum and the average distances are 768 and 620.3 cm, respectively. The other
three badminton rackets are used to hit the oncoming shuttlecocks from the opponent
to the opposite court. It depends on the controller that how effectively the controller
can control the robot and position it to hit the oncoming shuttlecock. The DC motors
used in the base are of 800 rpm, and it makes the motions of the robot very fast. The
structure of the robot is very stable and it will not topple if stopped suddenly.

6 Future Work

More advancement can be done in this robot by incorporating vision system which
can trace the trajectory of the shuttlecock as well as the speed of the oncoming
shuttlecock. Also, the 3D view of the standard badminton court can be created so as
to trace the shuttlecock. For this purpose, we can use LIDAR (Light Detection and
Ranging). This can enhance the accuracy of the robot and can also help in making this
robot autonomous. Also, the adaptive and fuzzy control system can be incorporated
to enhance the performance of the robot.

[Link]@[Link]
PS-2 Controlled Four Wheel Base Drive Badminton Playing Robot 107

References

1. Arenas, M., Ruiz-del-Solar, J., Norambuena, S., Cubillos, S.: A robot referee for robot soccer.
In: Iocchi, L., Matsubara, H., Weitzenfeld, A., Zhou, C. (eds.) RoboCup 2008: Robot Soccer
World Cup XII, vol. 5399, pp. 426–438. Springer, Berlin (2009)
2. Chen, X., Tian, Y., Huang, Q., Zhang, W., Yu, Z.: Dynamic model based ball trajectory prediction
for a robot ping-pong player. In: IEEE International Conference on Paper presented at the
Robotics and Biomimetics (ROBIO), 2010
3. Yang, H.: Robot football goalkeeper speed control research based on BP neural network algo-
rithm. J. Chem. Pharm. Res. 5(12) (2013)
4. Zitzewitz, J.V., Rauter, G., Steiner, R., Brunschweiler, A., Riener, R.: A versatile wire robot
concept as a haptic interface for sport simulation. In: IEEE International Conference Paper
presented at the Robotics and Automation ICRA ‘09, 2009
5. Zhang, J.: Autonomous mobile robot relay race competition. Adv. Sport Sci. Comput. Sci. 57,
41 (2014)
6. Laue, T., Birbach, O., Hammer, T., Frese, U.: An entertainment robot for playing interactive ball
games. In: Behnke, S., Veloso, M., Visser, A., Xiong, R. (eds.) RoboCup 2013: Robot World
Cup XVII, vol. 8371, pp. 171–182. Springer, Berlin (2014)
7. Stoev, J., Bartic, A., Gillijns, S., Symens, W.: Badminton playing robot-a multidisciplinary test
case in mechatronics. Paper presented at the Mechatronic Systems (2010)
8. Chia, K.S., Yap, X.Y., Low, E.S.: A badminton robot-serving operation design. ARPN J. Eng.
Appl. Sci. 11(6), 3968–3974 (2016)
9. Bill Porter: [Link]
0/ (2010)

[Link]@[Link]
Detection of RREQ Flooding Attacks
in MANETs

B. Nithya, Aishwarya Nair and A. S. Sreelakshmi

Abstract Mobile Ad hoc Networks (MANETs) are prone to vulnerabilities as a

result of lack of centralized management, dynamic topologies, and predefined bound-
ary. There are many attacks that affect the performance of MANETs. Flooding attack
is a Denial of Service attack, also known as DoS attack that brings down the network
by flooding with a huge amount of route request (RREQ) packets, routed to a desti-
nation that does not exist. This paper proposes a robust detection system; a Fuzzy-
based Flooding Attack Detection System (FFADS) using a first-order Mamdani-type
fuzzy inference system to detect RREQ flooding attacks. This proposed system uses
network-specific parameters rather than node-specific factors. Comprehensive sim-
ulations are carried out to observe the variations of input parameters with respect
to the flooding level of the system. The simulation results reveal that any deviation
from the normal behavior of nodes is immediately detected by the proposed system.

Keywords MANET · Flooding attacks · Fuzzy logic · RREQ flooding · AODV

Network traffic

1 Introduction

Distributed DoS attack is an attack which makes a service unobtainable to users,

usually by interrupting or suspending the services temporarily. In an RREQ flooding
attack, a malevolent node floods the network by transmitting fake RREQs to non-
existent nodes. This leads to congestion, thus resulting in a Denial of Service. It is very

B. Nithya (B) · A. Nair · A. S. Sreelakshmi

Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, Tiruchirappalli, India
e-mail: nithya@[Link]
A. Nair
e-mail: aishwarya05nair@[Link]
A. S. Sreelakshmi
e-mail: sreelakshmias97@[Link]

© Springer Nature Singapore Pte Ltd. 2019 109

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
110 B. Nithya et al.

important to catch and prevent flooding attacks as these attacks result in congestion
of the network. Overflow of routing table in the intermediate nodes is caused by
RREQ flooding attacks which disable the nodes from receiving new RREQ packets,
resulting in Denial of Service attack. Furthermore, forwarding these bogus RREQ
packets consumes valuable node resources like energy and bandwidth.
Many solutions have been put forward to detect and avert flooding attacks in
a network including behavioral approach which uses machine learning to classify
nodes as malicious, statistical approach to prevent flooding in AODV networks, game
theory-based approach, dynamic-based profile approach, and fuzzy logic. The main
drawback faced by many of these approaches is that they restrict the number of
attacker nodes and detect the maliciousness of the node, fix other parameters in the
network, or work effectively only for smaller networks. Many surveys have also been
conducted that analyze the effect of this attack on the different network parameters.
Most of this research focuses on the nodes of the network and on classifying them
as malicious and non-malicious, rather than on how the communication within the
network is affected. Also, the dynamic nature of MANETs is often ignored in the
development of solutions to detect the attacks on the network.
This paper introduces a system called Fuzzy-based Flooding Attack Detection
System (FFADS), to catch RREQ flooding attacks especially in AODV-based net-
works using the first-order Mamdani-type fuzzy inference system. The solution
focuses on the use of parameters that are specific to the network rather than the
nodes. To detect the level of flooding attack in the system, the proposed system
uses the dynamic input parameters, which are the properties of the network such
as throughput, packet loss ratio, and routing overhead. While detecting the level of
flooding attack, the proposed FFADS also detects abnormalities in communication
within the network which may or may not be due to RREQ flooding. By stopping
the malicious node from sending packets, the network can be restored to its previous
state with increased throughput and reduced congestion.
This paper is organized into five sections. Section 2 deals with related works
besides their detailed comparison. Section 3 deals with our proposed system for
detection of flooding attacks. This includes an analysis of the different input param-
eters, the structure of the fuzzy inference system, and the conversion of the output
value of the fuzzy system to the level of flooding in the network. Section 4 deals
with performance analysis. It explains the simulation setup, the metrics involved in
calculations, and graphs depicting the variations of input parameters with the number
of flooding nodes. Section 5 presents the conclusion of this paper.

2 Related Works

There are no fixed topologies in a MANET because the topologies are dynamic in
nature. The number of nodes and their relative positions vary greatly, and hence,
detection of the number of malicious nodes becomes difficult. Various methods are

[Link]@[Link]
Detection of RREQ Flooding Attacks in MANETs 111

suggested to detect the presence of flooding nodes in MANET. Some of them are
discussed in the following paragraphs.
A Fuzzy-Based Trust Model (FBTM) is proposed [1] to disclose selfish nodes in
MANETs. The fuzzy-based analyzer is joined with a trust model to identify the non-
cooperative behavior of attacker nodes from the non-attacker nodes. In this scheme,
every node in the network continuously checks its nearby nodes for their activities.
Every node calculates a value called the trust value of its neighbors, and these values
are then passed as input to a fuzzy function. The fuzzy function calculates the net
trust value, and the attacker nodes are then detected based on this value.
A Novel Intrusion Detection System (NIDS) for ad hoc network is suggested in
[2] using fuzzy logic. The first-order Sugeno-type fuzzy inference system (FIS) is
adopted for intrusion detection in a mobile ad hoc network. In this system, different
parameters are calculated from the network and fed to the FIS. The FIS returns the
verity level for each node. This method is accurate in terms of a high true positive
rate and low false positive rate, but requires the number of attacker and non-attacker
nodes to remain constant.
The behavioral approach is adopted in [3] to detect malicious attacks. In this
system, SVM classifier is used to classify nodes as malicious and non-malicious.
The statistical approach is used in [4] to defend against RREQ flooding attacks in
MANETs. The RREQ packets are monitored in real time, and the nodes are compared
against their neighbors to check if they exceed default route request limit.
Alleviating route request flooding attack in MANET is suggested in [5] using node
reputation scheme to abide by the impact of flooding attack in MANET. This scheme
checks the status of a node intermittently and restricts its route request transmitting
rate accordingly.
The detection of SYN flooding attack in AODV protocol is performed in [6] to
disclose the presence of the SYN flooding attack. The scheme proposed uses game
theoretical approach to form a game between the attacker node and the multimedia
server node. The performance of the detection algorithm is measured by examining
the several qualities of the parameters.
Ad hoc flooding attack is analyzed, and a Flooding Attack Prevention (FAP)
mechanism is developed in [7]. The proposed scheme uses a trust function to record
the number of route request packets and calculates the trust value. If the calculated
trust value exceeds the limit, the network drops these route request packets. In [8],
throughput, packet delivery ratio, and round-trip delay are compared with normal
network (without attacker nodes) and a network with few attacker nodes. The per-
formance of the network is compared in all the three scenarios.
Dynamic Profile-Based Technique (DPBT) is proposed [9] to detect the flood-
ing attacks in MANET. The proposed scheme defines a profile value based on the
performance of MANET. It recognizes the attack and tries to stop it every time the
node attempts to exceed the threshold value. This value changes with respect to the
request placed by the network.
A flow-based discovery mechanism against flooding attacks is developed in [10].
Two flow-based detection features are designed, and the algorithm used on them
precisely detects flooding attacks.

[Link]@[Link]
112 B. Nithya et al.

The above-discussed methods use a number of methods to detect the presence

of flooding nodes in the system. Most researches concentrate exclusively on either
detecting the presence of the malicious nodes or on being aware of the presence of
these nodes and mitigating the effects of the attack. The main difficulties faced are
due to the unpredictable nature of a MANET, in terms of number of harmless nodes,
mischievous nodes, network topology, and mobility pattern. Most studies tend to
restrict one or more of the above parameters to compute the result. Motivated by these
factors, our paper tries to vary as many parameters as possible, while maintaining a
maximum level of accuracy. Our paper uses fuzzy-based approach to detect the level
of severity of flooding attack in a network independent of the nodes in the network,
topology of the network, or the percentage of flooding nodes.

3 Proposed Fuzzy-Based Flooding Attack Detection System

(FFADS)

The proposed Fuzzy-based Flooding Attack Detection System (FFADS) uses fuzzy
logic to detect the number of flooding nodes in a MANET. Fuzzy logic is beneficial
in this situation as it can handle uncertainties and make decisions in a given range.
This paper uses three-input, single-output-based first-order fuzzy Mamdani infer-
ence system for composing the decision. The fuzzy parameters are elicited from the
network traffic and then passed on to the fuzzy interface. In the fuzzy interface, the
fuzzy rules are applied and the number of flooding nodes is estimated.

3.1 Fuzzy Controller in the Proposed System

A MANET network is simulated with an arbitrary number of nodes and various

parameters of the network—routing overhead, throughput, and packet loss ratio are
extracted from the traffic. These parameters are then fed to fuzzy inference system as
shown in Fig. 1, which uses a set of rules to define the output. This can be reported by
the system in terms of the extent of flooding attack in the system, which is classified
as low, medium, and high.
From the simulated MANET environment, the parameters—throughput, packet
loss ratio, and routing overhead—to be used in the intrusion detection system are
extracted. The inputs are then fed to the fuzzy controller. The fuzzification module
converts input data to values of the membership functions and matches data with
conditional rules in the rule base. The Mamdani-type fuzzy inference engine applies
the rules and returns a fuzzy set for the defuzzification block. The output values
are converted to crisp values through the defuzzification module. The output values
are used to decide the extent of flooding attack in the system. The fuzzy output is

[Link]@[Link]
Detection of RREQ Flooding Attacks in MANETs 113

Fig. 1 Fuzzy inference system

generated within the range of [0, 1] which indicates the flooding level where 0–0.3
is classified as low, 0.3–0.7 is classified as medium, and 0.7–1 is classified as high.

3.2 Fuzzy Inputs

Routing Overhead: Routing overhead is determined as the ratio of route request

packets (RREQ packets) among the total number of packets sent. It is observed
that, as the number of flooding nodes increases, the number of route request packets
relative to the total number of packets sent increases. Since routing overhead depends
on the number of RREQ packets, the value will dynamically change depending on
whether the bandwidth is utilized by RREQ packets. The membership function of
this fuzzy input variable is depicted in Fig. 2.
Number of Route request packets sent
Routing overhead 
Total number of packets sent

Packet Loss Ratio: Packet loss ratio is the ratio of the number of route request
packets dropped relative to the total number of RREQ packets sent. It is observed
that, as the number of flooding nodes increases, the packet loss ratio increases as
the packets have to share bandwidth with the increasing number of route request
packets. In a flooding attack, fake RREQs are sent to a destination that does not
exist. Therefore, the fake RREQs are dropped since the node is unable to forward the

[Link]@[Link]
114 B. Nithya et al.

Fig. 2 Routing overhead membership function

Fig. 3 Packet loss ratio membership function

RREQ packet. This increases packet loss ratio. The membership function is depicted
in Fig. 3.
Number of RREQ packets dropped
Packet loss ratio 
Number of RREQ packets sent

Throughput: In this simulation, throughput is determined as the fraction of

TCP/UDP packets received at the destination to the sent packets. It is observed
that, as the number of flooding nodes increases, the throughput decreases as the
TCP/UDP packets are forced to share bandwidth with the increasing route request
packets. As the number of fake RREQ increases, the channel experiences conges-
tion and throughput gradually reduces. The membership function of this fuzzy input
variable is depicted in Fig. 4.
Number of data packets
Throughput 
Time

[Link]@[Link]
Detection of RREQ Flooding Attacks in MANETs 115

Fig. 4 Throughput membership function

Table 1 Fuzzy inference rules

Rule number Routing overhead Packet loss ratio Throughput Flooding level
1 Low Low High Low
2 High High Low High
3 Medium Medium Medium Medium
4 High High Medium High
5 Low Low Medium Low
6 Low Medium High Low
7 High Medium Medium High
8 Medium Medium High Medium
9 High Medium Low High

3.3 Fuzzy Inference Rules

This detection system uses Mamdani-type inference system and takes in three param-
eters as input values—routing overhead, packet loss ratio, and throughput. The rule
base considered in the proposed FFADS is given in Table 1.

3.4 Fuzzy Surface View

The surface view between two input parameters and the output parameter (flood-
ing level) is shown in Figs. 5, 6, and 7. Different colors represent the severity of
flooding. Figure 5 shows the output surface (flooding level) versus the two param-
eters—routing overhead and packet loss ratio. Figure 6 depicts the output surface
(flooding level) versus the two parameters—routing overhead and throughput. The
output surface (flooding level) versus the two parameters—throughput and packet
loss ratio—is plotted in Fig. 7. The different colors of the surface graph—blue, green

[Link]@[Link]
116 B. Nithya et al.

Fig. 5 Routing overhead, packet loss ratio

Fig. 6 Packet loss ratio, routing overhead

and yellow—represent the different levels of the output for the two given inputs, that
is, the different intensities of flooding attack.

3.5 Fuzzy Output

The controller takes in the above-mentioned fuzzy inputs and maps them to their
membership functions. This is then fed to the fuzzy rules. The obtained truth values
are then defuzzified. The output values represent the extent of flooding attack in the
system as shown in Fig. 8.

[Link]@[Link]
Detection of RREQ Flooding Attacks in MANETs 117

Fig. 7 Throughput, packet loss ratio

Fig. 8 Fuzzy output

4 Performance Analysis

This section briefs about the simulation setup and the performance analysis depend-
ing on the routing overhead, packet loss ratio, and throughput.

4.1 Simulation Setup

For this system, Ns2 simulator is used to simulate MANET. The simulation param-
eters are shown in Table 2. The parameters are extracted from the trace files and are
fed to the Fuzzy Toolbox in MATLAB.

[Link]@[Link]
118 B. Nithya et al.

Table 2 Simulation Parameter Value

parameters
Mac type IEEE 802.11
Routing protocol AODV
Antenna Omni-directional
Simulation time 20 s
Traffic type FTP
Topology Random
Simulation area 500 m × 500 m
Number of nodes Random

Fig. 9 Routing overhead versus number of flooding nodes

4.2 Metrics

In the proposed system, the following three parameters are considered to affect the
level of flooding attack in the system.
Routing Overhead: Routing overhead is observed to increase with the extent of
flooding attack in the system. The graph plotted in Fig. 9 shows a setup consisting of
25 nodes and different number of flooding nodes ranging from 0 to 20. It is observed
that routing overhead increases linearly till a certain point beyond which it is observed
to be constant. After detection of flooding attack, if the flooding node is correctly
identified and the RREQ packets sent by the flooding node are ignored by all the
other nodes, the effect of flooding attack is greatly reduced.
Packet Loss Ratio: It is observed from Fig. 10 that there is gradual increase in
packet loss ratio with the increase in number of flooding nodes. Detection of flooding
attacks is simulated by dropping any RREQ packets with unknown destination. The
number of RREQ packets dropped is the same in both cases. Since packet loss ratio

[Link]@[Link]
Detection of RREQ Flooding Attacks in MANETs 119

Fig. 10 Packet loss ratio versus number of flooding nodes

is the number of RREQ packets dropped, the packet loss ratio remains almost the
same. Figure 10 shows that the RREQ packet loss ratio before and after detection.
Throughput: It is inferred from Fig. 11 that the throughput decreases as the num-
ber of flooding nodes increases. Since the fake RREQ packets utilize the bandwidth,
the overall throughput decreases. It is observed that throughput increases signifi-
cantly after detection of flooding attack as the entire bandwidth can now be used by
the data communication taking place. The RREQ packets sent by the flooding node
are dropped by all other nodes as long as it continues flooding the medium. Figure 11
depicts the variation of throughput with number of flooding nodes before and after
detection of the flooding attack.
Predicted Flooding Level: The graph plotted in Fig. 12 depicts the flooding
level of the network which consists of nodes ranging from 0 to 30. We have taken
different setups with different number of nodes and measured the input parameters.
It is observed that the predicted flooding level increases with the increase in number
of flooding nodes. It is also observed that the predicted flooding level is independent
of number of nodes, simulation time, and number of connections.

5 Conclusion

In this paper, a detection system—Fuzzy-based Flooding Attack Detection System

(FFADS)—is proposed for detecting flooding attacks using fuzzy logic. The proposed
system predicts the flooding level using three input parameters—routing overhead,
throughput, and packet loss ratio. The flooding level is independent of the number
of nodes, simulation time, and number of connections and depends only on the
externally measured parameters. The system has an edge over existing systems that

[Link]@[Link]
120 B. Nithya et al.

Fig. 11 Throughput versus number of flooding nodes

Fig. 12 Predicted flooding level versus number of flooding nodes

detect intrusions as it is dynamic in nature and the predicted flooding level changes
with the performance of the network. Also since the system does not restrict the
size or topology of the network, it can be applied on real-world networks to detect
flooding attacks.

[Link]@[Link]
Detection of RREQ Flooding Attacks in MANETs 121

References

1. Ullah, Z., Khan, M.S., Ahmed, I., Javaid, N., Khan, M.I.: Fuzzy-based trust model for detection
of selfish nodes in MANETs. In: 2016 IEEE 30th International Conference on Advanced
Information Networking and Applications (AINA), Crans-Montana, pp. 965–972 (2016)
2. Chaudhary, A., Tiwari, V., Kumar, A.: A novel intrusion detection system for ad hoc flooding
attack using fuzzy logic in mobile ad hoc networks. In: International Conference on Recent
Advances and Innovations in Engineering (ICRAIE-2014), Jaipur, pp. 1–4 (2014)
3. Patel, M., Sharma, S.: Detection of malicious attack in MANET a behavioral approach. In:
2013 3rd IEEE International Advance Computing Conference (IACC), Ghaziabad, pp. 388–393
(2013)
4. Rmayti, M., Begriche, Y., Khatoun, R., Khoukhi, L., Gaiti, D.: Flooding attacks detection in
MANETs. In: 2015 International Conference on Cyber Security of Smart Cities, Industrial
Control System and Communications (SSIC), Shanghai, pp. 1–6 (2015)
5. Choudhury, P., Nandi, S., Pal, A., Debnath, N.C.: Mitigating route request flooding attack in
MANET using node reputation. In: IEEE 10th International Conference on Industrial Infor-
matics, Beijing, pp. 1010–1015 (2012)
6. Geetha, K., Sreenath, N.: Detection of SYN flooding attack in mobile ad hoc networks with
AODV protocol. Arab. J. Sci. Eng. 41, 1161 (2016)
7. Chouhan, N.S., Yadav, S.: Flooding attacks prevention in MANET. Int. J. Comput. Technol.
Electron. Eng. (IJCTEE) 1(3) (2013)
8. Bhalodiya, S., Balakrishnan, V., Chauhan, R., Chouhan, N.S., Fitri, M.H., Ghonge, M.M.,
Hong, F., Jawandhiya, P.M., Joshi, K., Lomte, V., Murthy, S.R., Pushpalatha, M., Rai, M.K.,
Song, J., Tupakula, U., Vaghela, K., Varadharajan, V., Venkataraman, R.: Enhanced detection
and recovery from flooding attack in MANETs using AODV routing protocol (2015)
9. Sathish, T., Sasikala, E.: Dynamic profile based technique to detect flooding attack in MANET.
Int. J. Innov. Res. Comput. Commun. Eng. 2(1) (2014)
10. Guo, Y., Gordon, S., Perreau, S.: A flow based detection mechanism against flooding attacks in
mobile ad hoc networks. In: 2007 IEEE Wireless Communications and Networking Conference,
Kowloon, pp. 3105–3110 (2007)

[Link]@[Link]
Selection of Software Development
Model Using TOPSIS Methodology

Dayanand Gaur and Sakshi Aggarwal

Abstract In software industry, the software project failure is a serious concern for
stakeholders. The company suffers a huge financial loss in a year due to the team’s
negligence and irresponsibility. It can happen because of mismanagement in decision
making at any stage of project development. But the study conducted so far cites
several causes such as unarticulated project goals, mishandling of requirements,
poor estimation of resources, sloppy software development life-cycle model. The
paper tries to reduce the effort of decision-makers and project team by outlining the
significance of TOPSIS model. The statistical and quantitative analysis is the main
feature of TOPSIS. It accomplishes the experts’ job by validating their opinions. It
prioritizes the defined options after evaluating them against confliction and multiple
attributes. The research proposes a decision-making framework for the selection of
software development model using one of the widely accepted multicriteria decision-
making tools, i.e., TOPSIS.

Keywords Software engineering · MCDM · Decision-making methods · TOPSIS

1 Introduction

1.1 Software Engineering

The traditional concept of software engineering, one of the prominent disciplines of

computer science, merely defines the detailed and systematic study to the develop-
ment of software [1, 2]. But in the past 15 years, the emerging trends and technologies,
diversities in customer requirements, market demand, and much more complied the

D. Gaur
Computer Science, Galgotias University, Greater Noida 201310, India
e-mail: er.daya2014@[Link]
S. Aggarwal (B)
Software Engineering, Galgotias University, Greater Noida 201310, India
e-mail: agg.sakshi77@[Link]
© Springer Nature Singapore Pte Ltd. 2019 123
L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
124 D. Gaur and S. Aggarwal

engineers to change their perception for software engineering. The term does not
remain only for the software development practices like elicitation, designing, mod-
eling, codification, debugging, maintenance, and documentation but data collection,
decision support system, knowledge discovery are some of the aspects also incorpo-
rated with the software engineering. Now, data accumulation, gaining experiences,
knowledge extraction, building statistical systems go hand-in-hand with software
development activities.
Domain experts, researchers, engineers, client, trade union involvement, and those
who directly or indirectly influence the development process have become a part of
the software engineering. All of them constitute as stakeholders [1]. They take every
promising step in order to make a successful project as the success or failure of
software is the responsibility of whole team. But apart from these entire endeavors,
the studies conducted by the researchers over the past decade state that the failure rate
of software has crossed the threshold. They identified certain causes of failure [3]
which are enumerated as (1) unarticulated project goals, (2) unclear requirements,
(3) sloppy software development models, (4) inaccurate estimation of resources, and
(5) poor communication among customers and developers.
One of the reasons they cited is the wrong selection of software development
model (SDM). The paper underlines this problem as a subject of study and proposes
a solution using TOPSIS methodology—an operative approach of multicriteria deci-
sion making (MCDM) [4].

1.2 Basics of Multicriteria Decision Making (MCDM)

MCDM is a qualitative and quantitative technique for making effective decisions

when multiple and conflicting parameters are taken into account [5]. It offers experts
an opportunity to utilize its significance in the areas where decision making is con-
sidered as tedious job. The wide variety of MCDM methods provides framework
which selects the best suited alternative after evaluating different criteria, defined
under a particular scenario [6, 7]. The paper focuses on the application of MCDM
approach in software engineering field.
The whole research is divided into several sections. The first section broadly
describes the MCDM and its methods followed by the literature review in software
engineering scenario. The flow of TOPSIS is described in the third section. The
following section is the main concern of this paper. It focuses on the proposed model
of TOPSIS for SDM selection. Results and discussion are briefly presented in the
next section. Finally, the above content is supported by the conclusion of this study.

[Link]@[Link]
Selection of Software Development Model Using TOPSIS Methodology 125

2 MCDM

MCDM has gained the popularity because of its intuitive nature. It begins with the
decomposition of the complex problem into hierarchical structure which is easy for
humans to understand. It is classified into various methods which provide optimal
solution for selecting the best among alternatives [5]. The following section highlights
the MCDM supporting methods and how they have been turned into a reliable and
robust decision-making tool in software engineering applications.

2.1 MCDM Methods

Figure 1 represents the classification tree of MCDM methods. The successive heading
gives a glance over the working of some methods extensively used in corporate sector
for making judgments.

2.1.1 Outranking-Based Methods

The outranking-based methods, such as ELECTRE [7, 8] and PROMETHEE, are

based on concordance analysis. It establishes the “outranking relationship” between
concordance and discordance matrices and uses their indices to choose the best
alternative.

Fig. 1 Hierarchical structure of MCDM methods [6]

[Link]@[Link]
126 D. Gaur and S. Aggarwal

2.1.2 Ranking-Based Methods

The ranking model prioritizes the alternatives after rigorous analysis based on either
geometrical mean, Euclidean distance or ratio. It performs pairwise comparison
among criteria as well as for each alternative. Two approaches, AHP and TOPSIS
[9], are most popular among decision-makers and experts. The reason behind is ease
in the steps of algorithm. In AHP, the Saaty Table 1 [10] plays a crucial role in
constructing pairwise comparison matrices. The weights are computed with the help
of geometric mean. The different alternatives are ranked after analyzing eigenvalues
and eigenvectors.
On the contrary, TOPSIS model uses Euclidean distance for weight computation.
Instead of using the Saaty scale, the preference is given to the view of experience
holders, decision-makers, and experts. Its plus factor is found in the consistency of
result even when the small degree of uncertainty cannot be neglected.

2.1.3 Fuzzy Methods

What happens when decision-makers stuck in certain circumstances where the uncer-
tainties, complexities, vagueness, incompleteness are high? In that case, the classic
approach does not have suitable answers. The introduction of fuzzy set theory offers
stakeholders a novel concept to not only take the discrete values, rather consider the
membership of defined variables. It aims to find a solution in between True and False,
Yes or No, 0 and 1, High or Low. The researchers embedded the fuzzy application
into traditional methods of decision making. Some of the fuzzy approaches are fuzzy
AHP and fuzzy TOPSIS.
Fuzzy AHP, more or less, is similar to AHP. However, the Saaty scale [10] had
been modified by researcher Chang [11]. The simple scale is converted into triangular
fuzzy number (TFN) for pairwise comparison matrices. TFN translates linguistic
term into a tuple containing optimistic, moderate, and pessimistic values [12]. The
entire algorithm runs alike AHP over TFNs.
As TOPSIS is another version of AHP, similarly fuzzy TOPSIS is close to fuzzy
AHP.

Table 1 Defined scale and their corresponding definition

Saaty scale Description
1 Equally important
3 Weakly important
5 Fairly important
7 Strongly important
9 Absolutely important
2, 4, 6, 8 The intermediate values between two adjacent
scales

[Link]@[Link]
Selection of Software Development Model Using TOPSIS Methodology 127

2.2 MCDM in Software Engineering Scenario

Recently in March 2018, TOPSIS approach is used as a tool in selection of location for
property development. The author [13] prescribed 32 criteria for analysis including
soil texture, drainage system, distance from city. Alternatives for locations are sites
A, B, and C which are part of this study through decision support system (DSS).
The steps of TOPSIS process are done as follows: (1) Collect relevant data about
locations, (2) location selection, (3) survey, (4) data processing using TOPSIS, (5)
discussion of results, and (6) recommend the location.
In August 2017, the commendable job is done by the team of research scholar for
the selection of best software engineering practices [4] using PAPRIKA, a decision-
making approach. The study covers 11 criteria as software development practices for
designing a medical application. The proposed PAPRIKA framework is used for the
project development which does not conform to a definite software methodology.
In March 2017, another popular technique fuzzy AHP (or FAHP) has been applied
in the software effort estimation in scarcity of data [14]. The results were validated
with IVR dataset of a software industry. The relative ranking of 20 projects from the
dataset has been achieved after evaluating their weights. The equation for estimating
effort is based on the variables including estimated effort, known effort, and their
respective weights.
In 2016, the application of integrated techniques comes into existence. AHP and
TOPSIS methodologies are integrated for ETL software selection [15]. ETL, stands
for Extract, Transfer and Load, is a powerful action for data preprocessing before
it loaded into data warehouse. The framework is composed of three steps: (1) AHP
technique, responsible for making pairwise comparison among criteria and deter-
mining their weights; (2) TOPSIS technique, responsible for constructing decision
matrices, determining ideal and negative ideal solution and prioritizing alternatives;
(3) final selection of the best alternative, i.e., decision making.

3 TOPSIS Methodology

The paper revolves around the implementation of TOPSIS in software project plan-
ning to answer “Which SDM is best fitted into the scenario?” The successive heading
describes the TOPSIS work flow.
TOPSIS, acronym for Technique for Order Preference by Similarity to Ideal Solu-
tion, follows two rules [13, 15]. One, selected alternative is closest to ideal solution,
and second, it should have the distance from the negative ideal solution. Ideal solu-
tion involves best criteria (maximum profit, minimum time, high value, etc.), whereas
negative ideal solution contains worst criteria (minimum profit, maximum time, low
value, etc.).

[Link]@[Link]
128 D. Gaur and S. Aggarwal

3.1 Major Steps

Suppose there are ‘k’ numbers of decision-makers, ‘m’ number of alternatives, and
‘n’ number of criteria, then the steps of algorithm are as follows:
Step 1 Generation of decision matrices
For every alternative m, the decision matrix D is given by Eq. (1),
 
D p  di j 1 ≤ p ≤ m, 1 ≤ i ≤ n, 1 ≤ j ≤ k (1)

where di j is the rating scale from 1 to 10 or 1 to 100, where 1 stands for low and 10
for very good.
Similarly, the decision matrix is constructed for rating the criteria and represented
by the form given below.
 
W  wi j 1 ≤ i ≤ n, 1 ≤ j ≤ k (2)

where wi j is the rating scale from 1 to 10 or 1 to 100, where 1 stands for low and 10
for very good.
The mean along each row of W is calculated and assign them to criteria as their
weights.

W  [w̄x ] 1 ≤ x ≤ n (3)

where wx is the computed mean along each row.

Step 2 Reduction of m decision matrices into single matrix
Taking mean along each row in m matrices, these are reduced into one decision
matrix of n × m dimension.
Step 3 Standardize the decision matrix
To standardize the matrices, the columns are divided with root of sum of the
squares computed along each row.
 
D  di j 1 ≤ i ≤ n, 1 ≤ j ≤ m (4)

[Link]@[Link]
Selection of Software Development Model Using TOPSIS Methodology 129

Step 4 Weighted decision matrix

To construct the weighted matrix, ratings achieved in step 3 are multiplied with
the respective criteria weights from Eq. (3). The modified decision matrix D from
step 3 is multiplied with the criteria weights wx as,
 
D  di j × wx 1 ≤ i ≤ n, 1 ≤ j ≤ m, 1 ≤ x ≤ n (5)

Step 5 Determining ideal and negative ideal solution

For ideal solution, compute the maximum for each criteria as,
  
I  max di j 1 ≤ i ≤ n, 1 ≤ j ≤ m (6)

For negative ideal solution, compute the minimum for each criteria as,
  
N I  min di j 1 ≤ i ≤ n, 1 ≤ j ≤ m (7)

Step 6 Determining distance from ideal and negative ideal solutions

Firstly, each weight is subtracted from concerned value in I and squared. Then,
root mean distance is evaluated along columns for determining closeness to the ideal
solution.
 2
D  di j − y 1 ≤ i ≤ n, 1 ≤ j ≤ m, y ∈ I for respective i (8)

For particular j, root mean distance for ideal solution Si is


Si  d1 j + d2 j + d3 j + . . . . . . + dn j (9)

Similarly, the above equation is used for determining separation from negative
ideal solution denoted by Sni
Step 7 Relative closeness to ideal solution
 
Sni
RC  (10)
Si + Sni

The RC denotes the relative closeness of each alternative to ideal solution.

[Link]@[Link]
130 D. Gaur and S. Aggarwal

Table 2 TOPSIS for SDM selection [13]

Step 1 Determine number of alternatives, criteria, and
decision-makers’ involvement
Step 2 Construct decision matrices
Step 3 Standardize the decision matrix
Step 4 Compute weighted standardize decision matrix
Step 5 Determine ideal solution and negative ideal
solution
Step 6 Find relative closeness to ideal solution
Step 7 Order the alternatives in descending

Table 3 Decision matrix for criteria

E1 E2 E3 Weights
Flexibility of 10 9 9 10+9+9
3  9.33
requirements
Risk analysis 9 9 8 9+9+8
3  8.67
Customer 9 8 7 9+8+7
3  8.00
feedback

Table 4 Decision matrix for alternative Waterfall Model

E1 E2 E3 Weights
Flexibility of 3 4 3 3+4+3
3  3.33
requirements
Risk analysis 3 3 2 3+3+2
3  2.67
Customer 4 3 3 4+3+3
3  3.33
feedback

4 Proposed Work—TOPSIS for SDM Selection

The study traces the steps of TOPSIS to propose a model for choosing appropriate
SDM. The flow of our TOPSIS model is depicted in Table 2.
Consider
Alternatives: (1) Waterfall Model, (2) Spiral Model, (3) Agile Model
Criteria: (1) Flexibility of requirements, (2) Risk analysis, (3) Customer feedback
Experts: (1) E1, (2) E2, (3) E3 (Tables 3, 4, 5, 6, 7, and 8).

[Link]@[Link]
Selection of Software Development Model Using TOPSIS Methodology 131

Table 5 Decision matrix for alternative Spiral Model

E1 E2 E3 Weights
Flexibility of 6 7 8 6+7+8
3  7.00
requirements
Risk analysis 9 8 9 9+8+9
3  8.67
Customer 7 7 8 7+7+8
3  7.33
feedback

Table 6 Decision matrix for alternative Agile Model

E1 E2 E3 Weights
Flexibility of 10 10 9 10+10+9
3  9.67
requirements
Risk analysis 9 9 9 9+9+9
3  9.00
Customer 8 8 9 8+8+9
3  8.33
feedback

Table 7 Standard decision matrix

Waterfall Spiral Model Agile Model
Model
√
Flexibility of 3.33 7 9.67 3.332 + 72 + 9.672 
requirements 12.39
√
Risk analysis 2.67 8.67 9 2.672 + 8.672 + 92 
12.77
√
Customer 3.33 7.33 8.33 3.332 + 7.332 + 8.332 
feedback 11.58

Table 8 Standard decision matrix (continued)

Waterfall Model Spiral Model Agile Model
Flexibility of 3.33/12.39  0.26 7/12.39  0.56 9.67/12.39  0.78
requirements
Risk analysis 2.67/12.77  0.20 8.67/12.77  0.67 9/12.77  0.70
Customer feedback 3.33/11.58  0.28 7.33/11.58  0.63 8.33/11.58  0.71

5 Results and Discussion

Weight matrix is achieved by multiplying the weights of respective criteria. It is

represented as,

[Link]@[Link]
132 D. Gaur and S. Aggarwal

Table 9 Relative closeness of alternatives

Alternative Waterfall Model Spiral Model Agile Model
Relative closeness 0 0.72 1

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0.26 × 9.33  2.42 5.22 7.27
⎢ ⎥ ⎢ ⎥ ⎣
⎣ 0.20 × 8.67  1.73 ⎦, ⎣ 5.80 ⎦, 6.06 ⎦
0.28 × 8.0  2.24 5.04 5.68

Ideal solution, {7.27, 6.06, 5.68}, and negative ideal solution, {2.42, 1.73, 2.24}
Distance from ideal solution Si  {7.35, 2.15, 0} and negative ideal solution
Sni  {0, 5.67, 7.35}
Finally, relative closeness is calculated using Eq. (10), and we get (Table 9).
Therefore, the order of ranking is defined as below:
Agile Model > Spiral Model > Waterfall Model
So, the TOPSIS model recommends Agile Model because its closeness is maxi-
mum to best criteria.

6 Conclusion

Multicriteria decision making (MCDM) has been becoming indispensable in the

software engineering discipline for more than a decade [16]. Its presence is vital
when making decisions is uncompromising and stern for company. It does not limit
to software industry but also in ecological assessment [17]. The methods of MCDM
involve quantitative and qualitative analyses for linguistic variables in contrast to
discrete variables.
As we prove, the TOPSIS provides robust and simple framework for prioritizing
the alternatives and present the best suitable option before stakeholders. It does
not much require the domain knowledge but the statistical analysis. The results are
acceptable and clear to experts. It “best discriminates” the superior and inferior
alternative on the basis of closeness to attributes.
TOPSIS has wide potential to mark its presence in another area. Its scope is high
in the applications of software engineering and other subfields of computer science
which are yet to be explored.

References

1. Pressman, R.S.: Software Engineering: A Practitioner’s Approach, 7th edn. McGraw Hill, New
York (2010)
2. IEEE Standard Glossary of Software Engineering Terminology: Computer Society of IEEE,
New York (1990)

[Link]@[Link]
Selection of Software Development Model Using TOPSIS Methodology 133

3. Verner, J., Sampson, J., Cerpa, N.: What factors lead to software project failure? In: 2nd IEEE
International Conference on Research Challenges in Information Science, Morocco, pp. 85–93
(2008)
4. Hernández-Ledesma, G., Ramos, E.G., Fernández-y-Fernández, C.A., Aguilar-Cisneros, J.R.,
Rosas-Sumano, J.J., Morales-Ignacio, L.A.: Selection of best software engineering practices:
a multi-criteria decision making approach. Res. Comput. Sci. 47–60 (2017)
5. Velasquez, M., Hester, P.T.: An analysis of multi-criteria decision making methods. Int. J. Oper.
Res. 10, 56–66 (2013)
6. Aruldoss, M., Lakshmi, T.M., Venkatesan, V.P.: A survey on multi criteria decision making
methods and its applications. Am. J. Inform. Syst. 1, 31–43 (2013)
7. Triantaphyllou, E., Shu, B., Sanchez, S.N., Ray, T.: Multi-criteria decision making: an opera-
tions research approach. Encycl. Electr. Electron. Eng. 15, 175–186 (1998)
8. Ermatita, Hartati, S., Wardoyo, R., Harjoko, A.: Electre methods in solving group decision
support system bioinformatics on gene mutation detection simulation. IJCSIT 3, 40–52 (2011)
9. Anitha, M., Sathiya, M.: Improved technique by integrating AHP-TOPSIS for prioritization.
IJETST 2, 1911–1915 (2015)
10. Saaty, T.L.: How to make a decision: analytic hierarchy process. Eur. J. Oper. Res. 48, 9–26
(1990)
11. Chang, D.-Y.: Applications of the extent analysis method on fuzzy AHP. Eur. J. Oper. Res. 95,
649–655 (1996)
12. Ayhan, M.B.: A fuzzy AHP approach for supplier selection problem: a case study in a gear
motor company. IJMVSC 4, 11–23 (2013)
13. Jollyta, D.: TOPSIS technique for selecting of property development location. Softw. Eng. 6,
20–26 (2018)
14. Sehra, S.K., Brar, Y.S., Kaur, N.: Applying fuzzy-AHP for software effort estimation in data
scarcity. IJETT 45, 4–9 (2017)
15. Hanine, M., Boutkhoum, O., Tikniouine, A., Agouti, T.: Application of an integrated multi-
criteria decision making AHP-TOPSIS methodology for ETL software selection. In: Springer-
Plus, pp. 1–17 (2016)
16. Dipti, S.: Maintainability estimation of component based software development using fuzzy
AHP. IJETST 1, 280–285 (2014)
17. Ersayin, K., Tagil, S.: Ecological sensitivity and risk assessment in the Kizilirmak Delta. Fre-
senius Environ. Bull. 26, 6508–6516 (2017)

[Link]@[Link]
Situation-Aware Conditional Sensing
in Disaster-Prone Areas Using Unmanned
Aerial Vehicles in IoT Environment

J. Sathish Kumar, Mukesh A. Zaveri, Saurabh Kumar and Meghavi Choksi

Abstract Environmental sensing is the most crucial task that needs to be performed
in order to analyze the situation of a region during a disaster. The devices deployed
in such regions are responsible for sensing and communication effectively. During a
disaster, the operation of these devices may be affected by the environmental condi-
tions and their respective power constraints. Moreover, the mobility of these devices
in the network leads to a challenging task to perform sensing and communication in
such an environment. The disaster recovery may need different sensor data at various
points of time. In such cases, the selectivity of data from different sensors and its
dissemination in real time are the most important tasks. In this paper, the proposed
algorithm is based on the situation-aware conditional sensing for disaster-prone areas
using unmanned aerial vehicles. The technique presented in this paper focuses on
the control of way points of the aerial vehicles based on the events detected in the
Internet of Things environment.

Keywords Disaster management · Unmanned aerial vehicles · Internet of things

1 Introduction

The disaster management is a crucial research problem that needs attention to address
various critical problems such as environmental monitoring, real-time data collection,

J. Sathish Kumar (B) · M. A. Zaveri · S. Kumar · M. Choksi

Computer Engineering Department, Sardar Vallabhbhai National Institute of Technology, Surat
395007, India
e-mail: ds14co001@[Link]
M. A. Zaveri
e-mail: mazaveri@[Link]
S. Kumar
e-mail: ds14co006@[Link]
M. Choksi
e-mail: meghavichoksi@[Link]
© Springer Nature Singapore Pte Ltd. 2019 135
L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
136 J. Sathish Kumar et al.

rescue recovery operations. It is reported that in the last two decades, approximately
400,000 lives were lost all over the globe in total [1]. However, due to the techno-
logical advancements, it is possible to handle the effects of the disaster to a certain
extent. The most important challenge in such situations is the real-time surveillance
[2] and communication [3]. In addition, the adversity faced in the case of a disas-
ter demands that all the above-mentioned challenges must be solved intelligently,
preferably using automation. The real-time support must be provided irrespective of
the disaster type, i.e., natural or man made [4]. A glimpse of various disaster types
such as fire and smoke in a building due to terrorist attack, extreme fire in the forest,
road damages due to landslide, heavy flood, etc., is shown in Fig. 1 [5]. These situa-
tions need to be sensed both in pre-disaster and post-disaster cases. The sensed data
in the disaster environment must be communicated to the rescue team for preparing
the solution strategy, by bridging the gap between distance and time. The Internet of
things (IoT) environment provides a mechanism, with which the real-time sensing
can be communicated to any place at anytime from anywhere [6]. However, the IoT
environment must be configured and organized in such a way that an optimized com-
munication should result into automation of tasks, reducing the faults and failures in
the system. During the disaster, the real-time data collection needs sensor devices,
which are responsible to sense the situations under surveillance. However, these sen-
sors are very low-powered and energy-constrained devices. Due to this limitation
of the sensing devices, there is a need of IoT device to be computationally efficient
ultimately making it energy efficient using Internet. The IoT environment provides
collaborative processing among sensing devices and IoT devices that best suits these
requirements [7]. The rescue operations and the response in the case of a disaster
must be automated using these IoT devices, so that human lives are not lost in this
process. Due to the adversity and danger in the case of a disaster, these IoT devices
must be protected from the physical damage which is a precautionary requirement.
The unmanned aerial vehicles (UAVs) [8, 9] may serve a great purpose in such
scenarios as discussed above. These vehicles may be operated from anywhere, at
anytime, and to the distances that depend on the requirements of the situation under
observation. UAV is a mobile device and can be controlled by an IoT device for
fulfilling the requirements of real-time communication over the Internet from any-
where to the desired place [10]. Figure 1a depicts the fire and smoke situation in a
building. In such a case, the sensor devices are deployed in and around this building.
These devices sense different parameters and report the abnormality conditions to
the ground control station. In this regard, the dynamic network is to be set up for the
ad hoc data acquisition and on the spot monitoring from all possible directions. The
UAV is responsible for monitoring and data collection using these deployed devices
because of the hostile and restricted entry situations. It is mandatory to have knowl-
edge of the safe distance of operation for the UAV as incidences of fire and smoke
may affect the operations of the sensing and communication. Further, in the case of
any abnormality observed by the UAV, it must be able to respond and rescue itself.
Therefore, the two basic requirements that may be outlined in a disaster scenario are
(1) for real-time response, the effective sensing, and communication operation and
(2) the protection of UAV devices involved in the operation.

[Link]@[Link]
Situation-Aware Conditional Sensing in Disaster-Prone Areas … 137

Fig. 1 A glimpse of various disaster types [5]

This paper focuses on the use of UAV with the help of an IoT device to collect
the sensed data from the sensor devices in case of a disaster. The situation-aware
conditional sensing mechanism is considered to address the problems as specified
above. There is a necessity to communicate the information in real time, even in
adverse situations instantly and protect themselves so as not to burden the operation
in terms of cost. The validation of various situations under observation is performed
by providing different conditions to the UAV device for enhancing its automation
in total. The simulation is performed in IoT environment using DroneKit-Python
simulation tool [11], and the experiments were performed on an actual drone.
The rest of the paper is organized as follows: Section 2 discusses the proposed
algorithm that is used to validate the effectiveness of the UAV operation. Section 3
presents the implementation of the proposed algorithm and explains the situational
awareness of the UAV in different scenarios. Finally, Sect. 4 concludes the work.

2 Proposed Algorithm

In this section, the proposed algorithm performs the sensing and communication
using aerial vehicles by including various conditions. In this context, the problem is
formulated using different notations. Let us assume a geographical region, R. This

[Link]@[Link]
138 J. Sathish Kumar et al.

region may be a city, a forest, a rural area, a coastal area. Let us assume that the region
R is subdivided into n smaller subregions such that R  (R1 , R2 , … Ri …, Rn ). Such
divisions are needed to perform the area monitoring effectively and with ease. The
division of the region depends on different factors related to the type of region or the
type of disaster. In this context, the disaster may be a fire in a building [12], a flood
[13], or a damage to structures such as flyovers, heavy machineries, highways [14]
or rescue of a group of people lost in the forest [15]. The glimpse of various disasters
is shown in Fig. 1. It is further assumed that the sensors deployed in the region R
are for sensing different environmental parameters, such as temperature, pressure,
humidity, stress, strain, atmospheric pressure, luminosity.
Under such circumstances, there is a need to identify the subregion Ri , where
a critical situation has occurred. This is done based on the abnormality observed
in the sensing parameters of the sensors operating in Ri . Next, for monitoring the
situation specifically in the region Ri , the mission planning for UAV navigation at
different locations within Ri is performed at the ground control station. A ground
control station is a place where the trajectory planning and tracking of the UAVs
are performed. For the real-time analysis, communication between UAV and ground
station is performed using the Internet. In this context, the proposed algorithm works
in two steps. In the first step, the UAV senses the situation of environment on the
fly using various sensor devices. In the second step, the algorithm helps the UAV
to make the decision dynamically based on the environmental conditions. After the
surveillance and real-time communication of various parameters that are required,
the relief response team is dispatched to the most required place in the disaster site.
In the second step, for the dynamic navigation of the UAV, various conditions
are considered and incorporated in the UAV. To perform this dynamic navigation
operation of the UAV, let us assume that there are m different parameters sensed in
the region such that P  (P1 , P2 , … Pj …, Pm ). These values are collected by the
aerial vehicle, and the collected values are sent to the ground control station. For
instance, it may be possible that the temperature sensor shows more abnormality in
its reading as compared to the other parameters. Let this be the jth parameter in the
region where the disaster has occurred. Based on this gathering of the sensed data,
two types of threshold can be determined viz maximum and minimum values of the
parameter Pj . There are four different conditions that may be used to navigate the
UAV successfully. If the value of Pj is greater than the threshold maximum, observed
at the region, then an alert is generated. Similarly, if the sensed value of Pj is lesser
than the threshold minimum, then there may exist two different scenarios. Scenario
1 may suggest that the region is not under disaster, whereas on the contrary scenario
2 suggests that there is a false alarm. In both the cases, the alert is generated. Once
the alert is triggered, then the UAV may skip the remaining locations given in the
mission plan and make an immediate decision to return back to its launch location.
If the above-mentioned cases do not occur, then the UAV needs to continue the
navigation through the planned path as per the mission to reach the next location
for surveillance. The collected sensed information is communicated to the ground
station in real time. Following any sensed anomaly, the relief operation at the site
is executed. The relief operation can be initiated in two ways: either after complete

[Link]@[Link]
Situation-Aware Conditional Sensing in Disaster-Prone Areas … 139

monitoring of all the locations in the disaster region or at each and every location
in the planned mission. In the second case, the UAV sends the recorded parameters
continuously to the ground station in real time to prepare for the further solution
strategy relevant to the problems faced at the disaster site. The proposed algorithm
is explained in brief in Algorithm 1.
During a disaster, there are three most crucial operations that need to be per-
formed. First, the real-time data must be sensed and communicated to the ground
control station. This helps in formulating the strategy for the disaster relief. Second,
the disaster site Ri must be isolated from the other regions of R, such that the opera-
tion surveillance is focused on only Ri . Third, the relief response must be provided
in real time, with no loss of life. The proposed algorithm uses UAV as a part of
the disaster response team to collect real-time data from the sensors deployed at the
disaster site and communicate this data to the ground station using IP-based commu-
nication. Finally, the relied response may be provided in real time using UAV without
causing a threat to human life. This process of situation-aware conditional sensing
of the environment using IoT network may prove to be an effective mechanism in
disaster situations. The implementation of the proposed methodology is discussed
with different scenarios in the following section.
Algorithm 1: Condition-based Sensing Algorithm
Require: Sensing device, UAV, and m reading parameters
Ensure: Relief response at disaster site, protection of UAV during relief operation

1: Define a region R. Subdivide R such that R  (R1 , R2 , …, Rn ).

2: for disaster observed in Ri region do
3: /*Define Thresholds */
4: Estimate max and min values of parameter Pm .
5: if val(Pm ) ≥ max(Pm ) then
6: /*Alert Is generated */
7: Navigate UAV to launch location.
8: else if val(Pm ) ≤ min(Pm ) then
9: /*Alert Is generated */
10: Navigate UAV to launch location.
11: else
12: Continue the navigation through the mission path.
13: end if
14: Communicate the data collected to ground station.
15: end for
16: Perform the relief operation.

[Link]@[Link]
140 J. Sathish Kumar et al.

Fig. 2 UAV with its

specifications

3 Scenarios and Results

In this section, for demonstrating the proposed algorithm using UAV, DroneKit-
Python [11] is used and simulation is performed. DroneKit-Python is a simulation
tool that allows the developer to control the navigation of the UAVs. Additionally,
the mission planner tool [16] is used to track the navigation of the UAV in the
region. After successful simulations, the scenarios are emulated using the actual
UAV. As shown in Fig. 2, the Hexa Drone with high stability incorporated with
GPS module, Raspberry Pi 3 onboard microcontroller with pi camera, high accurate
flight controller, high-resolution video transmitter, analog video capture device, and
lithium-powered battery of 20 min fly time is designed for the purpose of post-disaster
management and surveillance. The glimpse of the drone on the fly for surveillance
is shown in Fig. 3.
For the purpose of demonstrating the different scenarios, initially, a region of
interest is defined as shown in Fig. 4. In the case presented in this paper, the region of
interest, R is outlined using Google maps on mission planner, which consists of three
department buildings, residential hostel, playground, health center, and few forest
areas. Further, the whole region is partitioned into nine different subregions given

[Link]@[Link]
Situation-Aware Conditional Sensing in Disaster-Prone Areas … 141

Fig. 3 Snapshot of
on-the-fly UAV

by R  (R1 , R2 , …, R9 ). These subregions are as shown in Fig. 5 with boundaries.

Let us assume that a disaster occurs in region R1 and the unnatural fire condition is
considered as the disaster for experimental purpose.
In such a scenario, the unmanned aerial vehicle is dispatched to perform the
surveillance operation, and at the same time, real-time communication of the sensed
information is sent to the ground control station. The UAV is instructed to navigate
through the region of disaster, which is shown in red color and performs the sensing
and real-time communication operations at eight different locations, L  (L1 , L2 , …,
L8 ).
The planned mission with the waypoints of the UAV is shown in Fig. 6, and their
respective locations are summarized in Table 1. These eight locations are the critical
points from where the required parameters are needed to be sensed. In our case,
the abnormality in temperature value needs to be sensed by the onboard system in
the UAV. The basic mission plan instructs the UAV to start its flight from the home

[Link]@[Link]
142 J. Sathish Kumar et al.

Fig. 4 A demonstration of region of interest in our campus

Fig. 5 Occurrence of disaster in region R1

location and sense the parameters at eight locations before returning back to its home
location again.
The ground station provides the UAV with an approximate maximum and mini-
mum values of temperature in the disaster region. The threshold of the minimum and
the maximum values is assigned during the experiment with 25 and 35 °C, respec-
tively. During its flight from one location to another, the UAV senses the temperature
value and communicates this information to the ground station. For the verification

[Link]@[Link]
Situation-Aware Conditional Sensing in Disaster-Prone Areas … 143

Fig. 6 Basic mission plan of UAV to sense the disaster region

Table 1 A summary of Location ID Latitude Longitude

locations for navigation of the
aerial vehicle Home location 21.161580 72.785524
L1 21.161925 72.785813
L2 21.161915 72.786162
L3 21.161720 72.785974
L4 21.161680 72.786280
L5 21.161370 72.786210
L6 21.161180 72.786776
L7 21.161275 72.785417
L8 21.161625 72.785132

of our proposed algorithm, three scenarios are demonstrated. In the first scenario,
a false alarm is created by altering the temperature value in the code, i.e., by giv-
ing 23 °C, which is lesser than the minimum temperature threshold provided to the
UAV. In such a case, the UAV generates the alert and returns to the home location
at L3 , as shown in Fig. 7, where the UAV path tracking is depicted. Similarly, the
second scenario is demonstrated as shown in Fig. 8, where the sensed temperature
by the UAV is more than the maximum threshold. In this case, the UAV senses 39 °C
at a location L6 that is more than the maximum temperature range provided to the
UAV. In such a scenario, again the UAV generates the alert and returns to the home
location, as the UAV is vulnerable to get affected by the ambient condition and there
is a possibility of crash. The UAV tracking in mission planner is shown in Fig. 8.
In the third scenario, if the UAV senses the temperature within the specified range,
then the UAV should navigate through each location of the disaster region. While
navigation, real-time communication of the sensed information to the ground control

[Link]@[Link]
144 J. Sathish Kumar et al.

Fig. 7 UAV returns to home location in the case false alarm at L3

station is performed. In this way, the control station knows the exact condition of
the site through parametric sensing report received from the UAV. The full mission
traversed by the UAV is as depicted in Fig. 9. Once the navigation area of the disaster
site is decided successfully, there is a requirement to carry out disaster relief opera-
tion. However, the UAV performs the situation analysis by considering the various
parameters and continuously evaluating the given conditions during the disaster. The
disaster control and corresponding recovery are very crucial operations and need to
be carried out with minimum risk of human lives. In this paper, the real communi-
cation using Internet enables the UAV to send the sensed data to the ground control
station or to the cloud. Further, the real-time streaming of video using Raspberry Pi
camera and analog camera is sent to the ground control station for accurate analy-
sis. Therefore, the application of the IoT in such a scenario is efficient by using the
automated devices as an integral part of the rescue team. This paper addresses these
concerns by the use of the UAVs for relief operation and providing a mechanism for
dynamic navigation, which is an essential requirement for such applications in order
to save human lives.

4 Conclusion

The disaster management and the response are very critical that need to be performed
in an effective way so that many lives can be saved. There is a need of real-time
response in the disaster scenarios. Moreover, the sensing and communication tasks
must be performed in such a way that the rescue can be carried out speedily. In this
paper, the use of the UAV is demonstrated for strategy formulation of rescue team
during disaster to sense the critical parameters at disaster site and accomplish the
real-time communication to the ground station. In this context, the use of the UAV
for real-time sensing and communication and an algorithm to protect these vehicles
during extreme conditions is addressed. The implementation of the proposed work is
performed using DroneKit-Python simulation tool and emulated the scenarios using

[Link]@[Link]
Situation-Aware Conditional Sensing in Disaster-Prone Areas … 145

Fig. 8 UAV returns to home location while exceeding the threshold range at L6

Fig. 9 UAV sensing values at eight different locations

actual drone developed for post-disaster management and aerial surveillance with
the aid of mission planner tool. In the future, the proposed work can be tested for its
efficiency with several parameters in extreme conditions.

Acknowledgements This work is supported by the Ministry of Electronics and Information Tech-
nology (MeitY), funded by Government of India (Grant no. 13(4)/2016-CC&BT).

[Link]@[Link]
146 J. Sathish Kumar et al.

References

1. Cavallo, E., Noy, I., et al.: Natural disasters and the economy—a survey. Int. Rev. Environ.
Resour. Econ. 5(1), 63–102 (2011)
2. Benfold, B., Reid, I.: Stable multi-target tracking in real-time surveillance video. In: Pro-
ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
pp. 3457–3464 (2011)
3. Houston, J.B., Hawthorne, J., Perreault, M.F., Park, E.H., Goldstein Hode, M., Halliwell, M.R.,
Turner McGowen, S.E., Davis, R., Vaid, S., McElderry, J.A., et al.: Social media and disasters:
a functional framework for social media use in disaster planning, response, and research.
Disasters 39(1), 1–22 (2015)
4. Rose, A.: Economic resilience to natural and man-made disasters: multidisciplinary origins
and contextual dimensions. Environ. Hazards 7(4), 383–398 (2007)
5. Zaveri, M.A., Kumar, J.S., Pandey, S.K., Choksi, M.: Collaborative Data Processing and
Resource Optimization for Post Disaster Management and Surveillance using IoT. Techni-
cal Report from (MeitY), India (2016)
6. Bandyopadhyay, D., Sen, J.: Internet of Things: applications and challenges in technology and
standardization. Wireless Pers. Commun. 58(1), 49–69 (2011)
7. Kumar, J.S., Zaveri, M.A.: Graph based clustering for two-tier architecture in Internet of
Things. In: Proceedings of 9th IEEE International Conference on Internet of Things (iThings),
pp. 229–233 (2016)
8. Pajares, G.: Overview and current status of remote sensing applications based on unmanned
aerial vehicles (UAVs). Photogram. Eng. Remote Sens. 81(4), 281–329 (2015)
9. Tuna, G., Nefzi, B., Conte, G.: Unmanned aerial vehicle-aided communications system for
disaster recovery. J. Netw. Comput. Appl. 41, 27–36 (2014)
10. Reiter, G.: Wireless connectivity for the Internet of Things. Europe 433, 868 MHz (2014)
11. Mason, I.A., Nigam, V., Talcott, C., Brito, A.: A framework for analyzing adaptive autonomous
aerial vehicles. In: Proceedings of the International Conference on Software Engineering and
Formal Methods, pp. 406–422. Springer (2017)
12. Hasofer, A., Beck, V.R., Bennetts, I.: Risk Analysis in Building Fire Safety Engineering.
Routledge (2006)
13. Plate, E.J.: Flood risk and flood management. J. Hydrol. 267(1–2), 2–11 (2002)
14. Yan, Y., Cheng, L., Wu, Z., Yam, L.: Development in vibration-based structural damage detec-
tion technique. Mech. Syst. Signal Process. 21(5), 2198–2211 (2007)
15. Davids, A.: Urban search and rescue robots: from tragedy to technology. IEEE Intell. Syst.
17(2), 81–83 (2002)
16. Hadi, G.S., Varianto, R., Trilaksono, B., Budiyono, A.: Autonomous UAV system development
for payload dropping mission. J. Instrum. Autom. Syst. 1(2), 72–77 (2014)

[Link]@[Link]
An Empirical Analysis of Collaborative
Filtering Algorithms for Building a Food
Recommender System

Ashique Mohaimin Ornab, Sakia Chowdhury and Seevieta Biswas Toa

Abstract Recommender system has been playing a great role in almost every sectors
starting from online shopping Web sites to online movie sites and social networking
sites. However, the use of recommendation engine has been very little in the food
sector. Sometimes people become tired of having the same kind of meals everyday
because of several reasons. Some people need to consume fixed food due to their
illness; others consume same meals everyday to stay healthy despite having any
diseases. In this paper, we have first discussed two collaborative filtering algorithms
that can be used to build a food recommender system for the people who have been
leading a monotonous food consumption lifestyle and are bored of having the same
kind of meals every day. After that, we have analyzed the two approaches of building
a food recommender system and finally concluded that the model-based approach is
more reliable than the memory-based approach.

Keywords Recommender system · Collaborative filtering · Cosine similarities

Alternating least squares (ALS)

1 Introduction

Recommender system is basically a system that enables items to be suggested to sim-

ilar users based on their personal common preferences and choices. Recommender
system searches for the similarities between items or users and suggests the most
similar items to the users. In a world of full of information and variety of choices,
it becomes difficult for people to choose the best out of the bests which match their

A. M. Ornab · S. Chowdhury (B) · S. B. Toa

Department of Computer Science and Engineering, BRAC University, Dhaka, Bangladesh
e-mail: sakiacm@[Link]
A. M. Ornab
e-mail: amornab@[Link]
S. B. Toa
e-mail: seevieta@[Link]
© Springer Nature Singapore Pte Ltd. 2019 147
L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
148 A. M. Ornab et al.

interests. This is where recommender system comes into play; it helps people find
out the best for them.
Generally, recommender system is an information filtering method finding
out related information. Recommender system can be divided mainly into two
types—content-based and collaborative filtering. There is another type called hybrid
which is usually done by the combination of content-based and collaborative filter-
ing. In our work, we have used collaborative filtering to build the food recommender
system.

1.1 Collaborative Filtering

People have been sharing ideas and opinions for centuries and that is the origin of
what we call collaborative filtering. It is a process that evaluates and filters people’s
likes and dislikes for a particular product or item. Collaborative filtering is a common
and popular recommendation algorithm that predicts and recommends based on past
behavior of the system’s users or ratings given by the users [1, 2].
Collaborative filtering can be divided into two types: memory-based and model-
based. In memory-based approach, the system makes up a relationship between
users and items in order to suggest users an item. The whole dataset is being used to
find out the similar values between users and/or items. Methods such as Euclidean
distance, Pearson correlation [3], Jaccard similarity, cosine similarity, can be used to
implement memory-based system. For model-based system, linear regression, matrix
factorization, Bayesian clustering, associative classification can be used.

2 Related Works

The online food menu recommender system proposed by Bundasak et al. [4] has
been built using collaborative filtering and slope one predictor. It is a memory-based
recommender system that works with users’ ratings and their preferences to suggest
menus. There are mainly three steps in this system—menu items rating, collaborative
filtering, and slope one predictor. And at the end, the system recommends top N menu
items to the user.
Li et al. [5] proposed a recommender system which uses web-based data min-
ing and the system recommends healthy meals to the users. This is a recommender
system especially for health-related issues that analyze health conditions and food
intake of a user. They proposed that there is a Web site where users had been order-
ing different food items of different recipes and from there they collected all the
users’ information using data acquisition. Data mining algorithms like clustering,
classification, association rules have been used to extract the necessary user data.
A personalized suggestion system proposed by Agapito et al. [6] monitors a user’s
health profile and recommends him or her healthy food meal. The profile of the user

[Link]@[Link]
An Empirical Analysis of Collaborative Filtering Algorithms … 149

has been built based on real-time questionnaires by medical doctors, and the answers
are provided by the users. The system has been built up by their self made algorithm
which focuses on users’ profile and their health status and suggest meals based on
that. The result of this system is typical Calabrian foods are suggested by DIETOS to
the users in three ways: (i) according to the user’s health profile, foods are suggested
spontaneously; (ii) showing the place on a map where the Calabrian foods are made;
and (iii) displaying the nutritional properties of each product stored in the database,
benefits, and side effects of pathologies and specific health conditions.
Gaur et al. [7] majorly focused on two initial dimensions of food recommenda-
tions: calories intake by the person and the calories which are left unused for that
person. The author proposed a model that used information taken from the univer-
sity’s students belonging to the age group 18–24. This information includes the basic
details of the users like age, weight, height, lifestyle, disease (if any), and food taken
by that user till evening. This information is given as input to calorie and BMI calcu-
lators for calculating the BMI and calorie consumed by that person, and it is based
on the information of a person defined above which will result in the body mass
index calculator (BMI calculator) is used to calculate the total fat of the body on
the basis of weight and height. After calculating these and taking into accounts, the
user’s BMI, DNA, and genetic disease—a suitable recommendation of food is given
to the user that will be beneficial for his or her health.

3 Methodology

Our food recommender system has been built in two ways—memory-based and
model-based approaches of collaborative filtering [8]. The memory-based approach
is used cosine similarity [9, 10], and model-based approach is used matrix factoriza-
tion [10, 11] which is a machine learning technique.

3.1 Memory-Based Recommender System

Memory-based collaborative filtering can be divided into two types: user–user col-
laborative filtering and item–item collaborative filtering which are basically neigh-
borhood methods.
For our memory-based recommender system, we have used item–item collabora-
tive filtering. In item–item collaborative filtering, the similarity values between items
are found and the most similar item to an item is suggested to the user.
The most significant part of memory-based recommender system is creating the
user–item matrix and then finding the similarity values between the items.
In order to build a food recommender system, we have accumulated an approxi-
mate of 100 users’ information that contains the users’ personal details such as age,
height, weight. We gave the users some selected 100 meal sets with respective calo-

[Link]@[Link]
150 A. M. Ornab et al.

ries for each meal set, in context of Bangladesh, for which they had to rate the meal
sets on a rating scale of 1–5. Any meal set rated a value of 5 means that meal set is
most healthy for the user and any meal set rated a value of 1 means that meal set is
least healthy for the user.
Our food recommender system will contain user profiles who have already rated
our given meal sets in the system by creating a profile of their own. They can also
add new meal sets and rate them.
The following example, with ten users and five meal sets, shown below describes
how our memory-based food recommender system works
I. A meal set (X) is given as user input. The meal set must be present in our database
in order to get a recommendation as replacement for the input meal set.
II. Creation of user–item matrix from the dataset. This small portion of dataset
(Fig. 1) has been used for better illustration and explanation of how our food
recommender system works.
In Fig. 1, ROWS are the users and COLUMNS are the meal sets (items). The
user–item matrix illustrates all the ratings given by each user to the meal sets they
consume. The higher the rating, the healthier is the food for them. For example, User
1 gave Item 1101 (bread, omelet) a 4 out of 5 and gave Item 1102 (omelet, boneless
chicken) a 5 out of 5. All the ratings are in the range of 1–5.
The meal sets (items) which are not given any rating are assumed to be rated a
zero by default. For example, User 2 has not given any rating to Item 1101 so the
rating for Item 1101 by User 2 is a zero.

Meal set (Item)

bread, omelette, bread, bread, low cereal, low
User omelette boneless fruit jam, fat milk, fat milk
chicken green tea poached egg
1101 1102 1103 1104 1105
1 4 5
2 5
3 5 2
4 1 4
5
6 5
7 1
8 2
9
10 2

Fig. 1 user–item matrix of ten users and five meal sets

[Link]@[Link]
An Empirical Analysis of Collaborative Filtering Algorithms … 151

In case of cosine similarity, a similarity value of 0 means least similar and a value
of 1 means most similar. In our user–item matrix, all our ratings are positive. So all
the item–item cosine similarity values will be in the range between 0 and 1.
III. Finding similarity between meal sets (items)
The similarity values between meal sets have been found using cosine similarity.
Cosine similarity finds similarity by calculating the angle between two vector lines.
Two items are considered as two vector lines in an n-dimensional space, and the
angle between them is the cosine angle. The smaller the angle, the more similar the
two items are. If the number of users is ‘n’, then the vector lines of the two items
will be in n-dimensional space. Here we have considered ten users so the similarity
values between any two items will be found in ten-dimensional space.
If we want recommendation for the meal set Item 1101, then we need to find out
the similarity values between Item 1101 and all other remaining items in the dataset.
The cosine similarity value between two items im and ib is calculated by

i m .i b xa,m xa,b
sucos (i m , i b )    (1)
||i m ||||i b || 2 x2
xa,m a,b

where
im  Item m
ib  Item b
x a,m  rating of ‘x’ of user ‘a’ on item ‘m’
x a,b  rating of ‘x’ of user ‘a’ on item ‘b’
Now let us apply this formula on our user–item matrix (Fig. 1).
Item 1101  (4,0,5,1,0,0,0,0,0,0)
Item 1102  (5,0,0,0,0,0,0,0,0,0)
Item 1103  (0,0,0,0,0,5,1,2,0,0)
Item 1104  (0,5,0,0,0,0,0,0,0,2)
Item 1105  (0,0,2,4,0,0,0,0,0,0)

sucos (Item 1101, Item 1102)

(4 × 5) + (0 × 0) + (5 × 0) + (1 × 0) + (0 × 0) + (0 × 0) + (0 × 0) + (0 × 0) + (0 × 0) + (0 × 0)
    
42 + 02 + 52 + 12 + 02 + 02 + 02 + 02 + 02 + 02 × 52 + 02 + 02 + 02 + 02 + 02 + 02 + 02 + 02 + 02
 0.6172133998

sucos (Item 1101, Item 1103)  0

sucos (Item 1101, Item 1104)  0
sucos (Item 1101, Item 1105)  0.4830458915

From the above calculation, we can conclude that Item 1101 and Item 1102 are
the most similar meal sets in this small portion of dataset. So when the user will give

[Link]@[Link]
152 A. M. Ornab et al.

Item 1101 as an input, our food recommender system will suggest him or her to have
Item 1102 among the other four meal sets.

3.2 Model-Based Recommender System

For building the model-based food recommender system, we have used the latent
factor model of collaborative filtering. Latent factor model characterizes both the
user and the item to explain ratings. The rudimentary idea is that vectors of latent
factors are inferred from ratings’ pattern and matrix factorization characterizes users
and items by this. The higher the correspondence between the users and items, the
greater the chance of recommendation.
The most significant part of model-based recommender system is filling up the
sparse matrix by alternating least square algorithm minimizing the error.
We have used the same dataset that has been used before to implement the memory-
based food recommender system.
Our model-based food recommender system will contain user profiles who have
already rated our given meal sets in the system by creating a profile of their own.
They can also add new meal sets and rate them.
The following example, with ten users and five meal sets, shown below describes
how our model-based food recommender system works
I. Give the user id as input. The user must be present in our database in order to
get a recommendation for meal sets that he or she has not eaten yet.
II. Creation of user–item matrix from the dataset (same user–item matrix as shown
in Fig. 1).
III. Factorize the user–item matrix into two different matrices.
In this step, the user–item matrix (Fig. 1) is factorized into individual user matrix
and item matrix by alternating least square algorithm. And if the user matrix and item
matrix are multiplied back together, it produces an approximation of the original
user–item matrix.
For illustration, we have used Fig. 2.
After formation of individual matrices U and P (Fig. 2), alternating least square
(ALS) puts random values in the cells of the matrices, calculating an error term and
then continues to alternate the values back and forth between matrix U and matrix
P until the error is minimized. Once this process is completed, the matrices U and P
are multiplied back together and the blank cells in the user–item matrix (Fig. 2) are
filled up with predicted ratings as shown in Fig. 3.
In Fig. 3, the predicted ratings N are in the range of 1–5 which is calculated
randomly by the algorithm. For instance, if the latent factors of Item 1101 and Item
1104 are similar and the predicted rating of Item 1104 is one of the highest values
among all other items for User 3, then User 3 will get Item 1104 as one of the top
meal sets. Here latent factors can be the type of meal set; that is, both the items are
for breakfast; the calories of the meal sets are similar.

[Link]@[Link]
An Empirical Analysis of Collaborative Filtering Algorithms … 153

4 5
5
5 2
1 4

5
1
2

R (m×n)
User-Item Matrix U (m×k) P (k×n)

Fig. 2 Matrix factorization of user–item matrix

Meal set (Item)

bread, omlette, bread, bread, low cereal, low
User omlette boneless fruit jam, fat milk, fat milk
chicken green tea poached egg
1101 1102 1103 1104 1105
1 3.99 4.95 N N N
2 N N N 4.99 N
3 4.78 N N N 1.93
4 0.98 N N N 3.98
5 N N N N N
6 N N 4.90 N N
7 N N 0.88 N N
8 N N 1.85 N N
9 N N N N N
10 N N N 1.99 N

Fig. 3 Filled up user–item matrix

3.3 Apache Spark

Apache Spark is a big data processing framework which enables us to use some of
the most popular tools to implement big data-related tasks. We used Apache Spark
to utilize Spark MLlib which is a machine learning library that consists of different
algorithms and utilities. We have implemented alternating least square (ALS) algo-
rithm for our recommendation system using MLlib. Moreover, we used Spark so that
in the future this system can take mass volume of data as input and for making it
efficient, scalability, and fast processing are needed.

[Link]@[Link]
154 A. M. Ornab et al.

4 Result and Analysis

For the memory-based food recommender system, if a user gives the unique meal
set id of any existing meal set of the system as input, then he or she will get a list of
top ten meal sets as recommendation that match the user’s preferences. You can see
the user has given 2201 as input for the meal set rice, mixed vegetables as shown in
Fig. 4.
For the model-based food recommender system, the user has his or her personal
unique user id using which he or she can rate any meal set in the system and can get
recommendation for new meal sets which the user has not tasted yet. For that only
the user requires to give his or her user id as input and then he or she will get top ten
meal sets’ recommendation that has been found by comparing the user’s previous
preferences. Figure 5 shows that for the user who has id 8 has got a list of top ten
meal sets which he or she might like to consume according to his or her preferences.
The memory-based food recommender system mainly focuses on how much an
item (meal set) is liked by the users. It figures out the similar ratings of an item (meal
set) pair and checks whether they are similar meals or not. This means if there is

Fig. 4 Top ten meal sets recommendation of memory-based food recommender system

Fig. 5 Top ten meal sets recommendation of model-based food recommender system

[Link]@[Link]
An Empirical Analysis of Collaborative Filtering Algorithms … 155

no similarity of ratings between any pairs of meal sets, then the system will give an
average recommendation for the meal set alternative to what the user is looking for.
So there must be such dataset that has at least one similar meal set for a given meal
set that is both the meal sets must be given similar values by at least one user. Here
no matter how poor or good ratings are given by the users to the meal sets it is not
taken into account. This means if two meal sets are given a rating of 1 by the same
user, then these two meal sets are very similar. But if one meal set is rated a 1 and
another meal set is rated a 5, then these are very dissimilar meal sets thus in such case
we will get average recommendation for a meal set. Even if a meal set is rated very
high by the users if that meal set does not have any similar rating with other meal
sets, then the recommendation of alternative meal sets will be very poor. Moreover,
if a meal set is rated by only one user and the user has not rated any other meal sets,
then the recommendation system will have no similar meal sets for that meal set and
therefore no recommendation for that meal set. The memory-based recommender
system using cosine similarity also assumes that the rating of a meal set is zero if it
has not been rated by any user. So it finds very dissimilarity among items which are
rated and which are not rated. Thus, again makes poor recommendation for a given
meal set.
On the other hand, model-based food recommender system not only focuses on
the values of the ratings given by the users. The machine learning technique matrix
factorization uses the existing ratings of meal sets to predict the ratings of other meal
sets that have not been rated by the users taking into account of other parameters of
the meal sets such as the type of meal set, the calories of the meal set. These predicted
ratings are fully done by the system, and there is no scope of any error if the ratings
on the meal sets given by the users are absolutely meaningful and correct. But in a
case where an item X is rated a 5.0 by user A, a 4.0 by user B, a 1.0 by user C, and D
has not rated item X. Now when the values of the ratings are computed for training
and getting the predicted ratings; sometimes the lowest rating value compared with
other ratings of an item is not considered. Here the rating value 1.0 might not be
computed for training the existing values and computing the predicted ratings. This
can make the recommendation very inefficient sometimes, and thus, scaling is really
necessary in such cases.
Another big factor that works for both the approach is how the users rate the meal
sets. If there is any wrong or misleading rating on the meal sets despite the meal sets
being very good for health, then also there is a possibility of getting wrong or poor
recommendation for meal sets. So the dataset must be arranged in such a way that
there is no scope of incorrect or false ratings on the meal sets.
Here we cannot exactly compare between these two types of recommender sys-
tem. But we can conclude that the model-based recommender system works best
since there is no scope of assuming an unrated meal set a zero that minimizes the
dissimilarity among different meal sets which happen in case of memory-based rec-
ommender system. Moreover, in the model-based food recommender system the
latent factors like calories, type of meal sets are also taken into account during rec-
ommendation which is not done by the memory-based food recommender system.
These kinds of meal information play an important role in recommending people

[Link]@[Link]
156 A. M. Ornab et al.

with healthy food. It narrows down the food recommendation by suggesting users
only with the meal sets that are within the calorie range of their usual food intake.

5 Conclusion

To conclude, we tried to build a food recommender system by using two different

collaborative filtering algorithms. We have seen that the model-based recommender
system is more effective than that of the memory-based system which suggests items
to users based on only the similar ratings no matter how poor the similar ratings are.
In other words, even if a meal set is not healthy for person in terms of calories, it
might be recommended to him or her. But the model-based system takes into account
the other parameters such as calories and suggests users with meal sets that are within
the users’ preferred calorie intake. This model-based recommendation engine will
hopefully help those people who have a boring, fixed eating pattern and make their
food consumption an interesting one by helping them to choose variety of meal sets.

References

1. Leskovec, J., Rajaraman, A., Ullman, J.D.: Mining of Massive Datasets, 2nd edn. Cambridge
University Press, New York, NY, USA (2014)
2. Schafer, J.B., Frankowski, D., Herlocker, J., Sen, S.: Collaborative filtering recommender sys-
tems. In: Brusilovsky, P., Kobsa, A., Nejdl, W. (eds.) The Adaptive Web. Lecture Notes in
Computer Science, vol. 4321. Springer, Berlin, Heidelberg (2007)
3. Arsan, T., Köksal, E., Bozkuş, Z.: Comparison of collaborative filtering algorithms with various
similarity measures for movie recommendation. Int. J. Comput. Sci. Eng. Appl. (IJCSEA) 6(3),
1–20 (2016). [Link]
4. Bundasak, S., Chinnasarn, K.: eMenu recommender system using collaborative filtering and
slope one predictor. In: The 2013 10th International Joint Conference on Computer Science and
Software Engineering (JCSSE), pp. 37–42. IEEE, Maha Sarakham, Thailand (2013). https://
[Link]/10.1109/jcsse.2013.6567316
5. Li, X., Liu, X., Zhang, Z., Xia, Y., Qian, S.: Design of healthy eating system based on web data
mining. In: 2010 WASE International Conference on Information Engineering, pp. 346–349.
IEEE, Beidaihe, Hebei, China (2010). [Link]
6. Agapito, G., Calabrese, B., Guzzi, P.H., Cannataro, M., Simeoni, M., Car´e, I., Lamprinoudi,
T., Fuiano, G., Pujia, A.: DIETOS: a recommender system for adaptive diet monitoring and
personalized food suggestion. In: 2016 IEEE 12th International Conference on Wireless and
Mobile Computing, Networking and Communications (WiMob), pp. 1–8. IEEE, New York,
NY, USA (2016). [Link]
7. Gaur, N., Singh, A.: Recommender system—making lifestyle healthy using information
retrieval. In: 2016 2nd International Conference on Next Generation Computing Technolo-
gies (NGCT-2016), pp. 479–484. IEEE, Dehradun, India (2016). [Link]
2016.7877463
8. Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Item-based collaborative filtering recommen-
dation algorithms. In: WWW ‘01 Proceedings of the 10th international conference on World
Wide Web, pp. 285–295. ACM, New York, NY, USA (2001). [Link]
372071

[Link]@[Link]
An Empirical Analysis of Collaborative Filtering Algorithms … 157

9. Nandi, M.: Recommender systems through collaborative filtering. (2017). [Link]

[Link]/recommender-systems-collaborative-filtering/
10. Levinas, C.A.: An analysis of memory based collaborative filtering recommender systems with
improvement proposals. UPCommons (2014)
11. Koren, Y., Bell, R., Volinsky, C.: Matrix factorization techniques for recommender systems.
Computer 42(8), 30–37 (2009). IEEE. [Link]

[Link]@[Link]
Performance Analysis of Iris Recognition
System

Ruqaiya Khanam, Zohreen Haseen, Nighat Rahman and Jugendra Singh

Abstract A biometric system offers automatic identification of an individual based

on a unique feature or characteristic possessed by the individual. Iris recognition is
regarded as the most reliable and accurate biometric identification system available.
Although iris identification system is based on pattern recognition technique but due
to poor iris boundary detection and high computational time in previous work, we
used neural network and discriminant machine learning technique to obtained high
accuracy. In this work, we implement neural network and discriminant analysis of
machine learning method for iris recognition in iris images to implement in day-
to-day life, using MATLAB 2016a. The emphasis will be only on the software for
performing recognition and not hardware for capturing an eye image. The proposed
method gives better recognition rate than SVM technique with less computational
complexity. Neural network and discriminant methods are used for matching and
finding recognition accuracy. Thus, the accuracy obtained from neural network is
94.44%, whereas from discriminant analysis the accuracy obtained is 99.99%.

Keywords Iris recognition · Neural network · Machine learning

1 Introduction

The advancement of technology and expand importance of security have gained more
awareness toward biometric identification system. Biometric systems are used based

R. Khanam (B) · Z. Haseen · N. Rahman · J. Singh

Galgotias University, Greater Noida 201310, India
e-mail: [Link]@[Link]
Z. Haseen
e-mail: zovahaseen93@[Link]
N. Rahman
e-mail: nighatrehman08@[Link]
J. Singh
e-mail: jugendra12031993@[Link]
© Springer Nature Singapore Pte Ltd. 2019 159
L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
160 R. Khanam et al.

on face, iris, fingerprint, gait parameter, etc. Application of biometric systems is used
for unique identity, border security, airport security, etc. Biometric identification is
the process of an automatic identification based on characteristic or unique feature
of an individual or a person. In 1980s two ophthalmologists, Leonard Flom and Aran
Safir proposed that even in twins no two irises are alike thus making them a good
biometric. By 1994, the algorithms have been developed and patented and further
now used as the basis for iris recognition systems [1, 2]. The iris starts to develop
in third month of gestation and complete its structure by the eighth month, however,
pigment accretion continues to the first postnatal years. Each individual has a unique
iris even the left and right eye of a particular individual differs in iris. Through the
strategies of image processing, uniqueness of an iris pattern can be extracted from
digitized image of the eye and thus encode it into a biometric template, which saved
in database. This biometric template contains a mathematical representation of the
unique data stored in the iris and enables comparisons to be made between templates.

2 Related Work

The first scientist who developed the algorithm for iris recognition system was Daug-
man integrodifferential operator was used for iris localization and for iris normaliza-
tion rubber sheet model of Daugman was used. The matching of two iris codes was
done by Hamming distance. Lim proposed an efficient method of personal identifi-
cation having high level of stability and distinctiveness. In this paper, Haar wavelet
is used to extract the features from iris image. Navjot provides the review of existing
methods as proposed by various researchers for iris recognition. Mohd. Tariq pro-
posed an algorithm by using 1D Gabor filter for the extraction of feature, normalizing,
and segmenting the iris and pupil boundaries of eye from database images. Further,
Proenca proposed a method which encloses three main parts that are eye detection,
iris segmentation, and discrimination of noise-free iris texture. An algorithm imple-
mented by Mayank was used to enhance both accuracy and speed of iris recognition.
Samir relates a GACs (Geodesic Active Contours) with an iris segmentation scheme
which extract the iris from nearby structures. The scheme invokes iris texture further
assisted by local and global properties of the image. Zhaofeng presented an algo-
rithm for both fast and accurate iris segmentation. FAWAZ is known for proposing
multi-algorithmic approach to enhancing the security of iris recognition system and
can be achieved by fusing the data acquired at the feature level and applying the
K-nearest neighbor classifier (K-NN).

[Link]@[Link]
Performance Analysis of Iris Recognition System 161

3 Background Theory of Proposed Work

3.1 Biometric Technology

Biometric generally describes a characteristic or a process. For automatic recognition,

it is used. Its framework provides a programmed recognition of a person taking same
kind of special elements or trademark governed by the person. A biometric system
generally involves three modules as: recognition, verification, identification (Fig. 1).
Sensor Module: It defines the connection of the human with the system and thus
vital to the execution of the biometric system.
Feature Extraction and Quality Assessment Module: The quality of the biomet-
ric data attained by the sensor is assessed for feature extraction. During enrollment,
the extracted feature is stored in the database known as “template”, representing the
identity of an individual.
Matching Module: The identity of an individual is verified by comparing the
template to the input (query) biometric data of an individual.
Decision-Making Module: This is to verify an exact identity or to provide ranks to
the enrolled identities. Usually, the match score is compared against the “threshold”
to determine the authenticity of an individual.

3.2 The Human Iris

To control the amount of light entering through the pupil is the function of iris. The
average diameter of the iris is 12 mm, and the pupil size can vary from 10 to 80% of
the iris diameter [3]. The pigmentation of iris depends on genetics which determines
its color.

Fig. 1 Working of biometric

system

[Link]@[Link]
162 R. Khanam et al.

3.3 Different Recognition System

Biometric recognition, or biometrics, refers to the automatic identification based on

anatomical or behavioral characteristic or unique feature of an individual. Some of the
recognition systems are fingerprint recognition, speech recognition, iris recognition.

3.4 Work Methodology

Steps in Iris Recognition System

Image Acquisition: It is the first step of the image processing and can be defined
as recovering an image from some source. High-quality image acquisition techniques
are used for iris recognition to make accurate models of different surfaces.
Image Preprocessing: It aimed to enhance ability to interpret quantitatively image
components. It removes low-frequency noise, normalizes the intensity of individual
images, and removes reflections.
Segmentation: The accomplishment relies on upon the imaging nature of images
of eye. Because of the exertion of close infrared light for enlightenment, images in
the CASIA iris database [4] do not contain specular reflections.
Normalization: The next stage after iris region segmentation is to change the iris
region, so it has fixed measurements to allow comparisons. There are dimensional
irregularities between eye images due to extending of iris created by pupil expansion
from shifting levels of brightening. The normalization methodology will create iris
regions, which have the same measurements, so two photos of the identical iris under
exclusive conditions will have characteristic features at the same spatial area.
Feature Encoding and Matching: The template that is created in the feature
encoding procedure will require a relating matching metric, which allow comparison
between two iris templates (Fig. 2).

Fig. 2 Technological perspectives in feature encoding module

[Link]@[Link]
Performance Analysis of Iris Recognition System 163

• Feature Encoding Algorithms: Wavelets can be deployed to decompose the infor-

mation in iris region into constituent that show up at different resolutions.
• Gabor Filters: Daugman encodes iris pattern data by using 2D variant of Gabor
filters. A 2D Gabor filter over the image space (x, y) is represented as:

G(x, y)  e−π [(x−xo ) /α 2 +(y−yo )2 /β 2 ] −2πi[u o (x−xo )+vo (y−yo )]

2
e

where (xo , yo ) is position in the image, (α, β) is the effective width

and length, and

(u o ,vo ) determine modulation, which has spatial frequency. ωo  u 2o + vo2 .
• Haar Wavelet: Oppenheim and Lim [5] use the wavelet change to extract features
from the iris region. From multi-dimensionally filtering, a feature vector with 87
estimations is featured. Since every estimation has a real value expanding from
−1.0 to +1.0, the feature vector is sign quantized so that any positive value is
represented by 1 and negative by 0.

3.5 Matching Algorithms

Matching Algorithms are used for verification as well as identification functions.

Three algorithms which are repeatedly used in iris recognition technology discussed
below are:
• Hamming Distance: The measure of same number of bits between two bit patterns
is given by hamming distance. The hamming distance for two bit patterns X and Y
is defined as the sum of discarding bits (sum of the exclusive-OR between X and
Y ) over N, the total number of bits in the bit pattern.

1 
N
HD  X i (XOR)Yi
N i1

• Weighted Euclidean Distance: In order to compare two templates composed of

integer values WED is used. The metric is employed by Zhu et al. [6] and is
specified as

1 
N
WED  ( f i − f i(k) )2 /(δi(k) )2
N i1

where f i is the ith feature of the unknown iris, and f i(k) is the ith feature of iris
template, k and δi(k) is the standard deviation of the feature in iris template k.
• Normalized Correlation: It is represented as

[Link]@[Link]
164 R. Khanam et al.


n 
m
( p1 [i, j] − μ1 )( p2 [i, j] − μ2 )/nmσ1 σ2
i1 j1

where p1 and p2 are two images of size n × m, μ1 and σ 1 are the mean and standard
deviation of p1 , and μ2 and σ 2 are the mean and standard deviation of p2 .

4 Proposed Work and Techniques Used

The implementation of proposed work is shown in Fig. 3.

4.1 Feature Extraction Technique

The feature map of iris image is extracted with the help of Daugman rubber sheet
model which is used for the normalization of image and then Haar wavelet is used
for the feature extraction.

START

PUPIL DETECTION
MODULE

SEGMENTATI IRIS DETECTION

ON MODULE MODULE

EYELIDS
DETECTION
USING
MORPHOLOGICAL
PRE PROCESSING
NORMALISATION USING DAUGMAN
MODULE RUBBER SHEET

FEATURE USING HAAR

ENCODING WAVELET

USING NEURAL
NETWORK AND MATCHING
DISCRIMINANT MODULE STOP

Fig. 3 Flow chart of proposed algorithm

[Link]@[Link]
Performance Analysis of Iris Recognition System 165

Fig. 4 Daugman’s rubber

sheet model

Daugman’s Rubber Sheet Model: The homogenous rubber sheet model elab-
orate by Daugman relocates every point inside the iris area to a couple of polar
directions (r, θ ) where r is on the interval [0, 1] and θ is angel [0, 2π ] (Fig. 4).
The relocating of iris region from Cartesian coordinates (x, y) to the normalized
non-concentric polar representation is modeled as

I (x(r, θ ), y(r, θ )) → I (r, θ )

With, x(r, θ )  (1 − r )x p (θ ) + r x1 (θ )

y(r, θ )  (1 − r )y p (θ ) + r y1 (θ )

where, I(x, y) is iris region picture, (x, y) are the Cartesian coordinates, (r, θ ) are
the relating standardized polar directions, and are the directions of the pupil and iris
limits along the θ direction.

4.2 Matching and Accuracy Determination

In this proposed work, the accuracy of the feature extraction is determined with the
help of discriminant analysis and neural network.
Machine Learning Algorithm: A type of artificial intelligence learning algorithm
which helps the computers the ability to learn without being clear-cut programmed.
There are various types of machine learning algorithm: supervised, unsupervised,
semi-supervised, reinforce.

5 Experimental Result

Every first image in result modules represent our best performance as when we
compare characteristically features of our first image with others we find considerable
difference in image quality but the catch is that for most of images taken we have
got such technical results which are relevant to our above assumptions. Outputs or
results are shown in sequential manner (Fig. 5).

[Link]@[Link]
166 R. Khanam et al.

The image is shown in Fig. 6 is binarized behavior of original iris grayscale

images. The image of this type is used for further segmentation of iris and then its
localization.
In Fig. 7, the image quality indicates a good level of clarity in the image repre-
sentation. The background is noise free and representations of pixels are 0. This is a
standard image which will be used for masking and making the image noise free.
The current image has been used to show the red circles that are marked in the
image. Here when the image is seen highlighted with circles, the radii so evaluated
are marked only for a certain region or at an angular area shown in Figs. 8 and 9.

Fig. 5 Image after

binarization of iris image

Fig. 6 Image after artefacts

removal

Fig. 7 Original iris image

holding the red circles with a
certain angle

Fig. 8 Image after

Daugman filtering

Fig. 9 Detailed, approximated, horizontal, and diagonal components of wavelets

[Link]@[Link]
Performance Analysis of Iris Recognition System 167

Fig. 10 Image shows neural

network training

To enable the transform of circular region into a horizontal region the image
so calculated is divided in two regions. Now if Fig. 10 is compared to Fig. 9, the
circular iris region will be considered as polar coordinates and then the red marked
semicircles are reduced to Daugman Cartesian coordinates. The image so seen here
is a combination of left and right region of iris.
This image is imperative reason being that image here is a segmented image or
partitioned image which will reduce the image size to either half dimension or to a
lower dimension. If this wavelet analysis is done then ultimate aim of the image is
to calculate the results as features to identify the prime features. In this research the
approximate image seen at the upper left corner will be reduced and used, the total
pixels will be approximately 512.
The feature of four images of subject 1 selected shown by number of columns,
here the number of images will be four and for each image 100 prime features are
shown out of 512. If calculations are traced, the features are evaluated using Haar
wavelet.

[Link]@[Link]
168 R. Khanam et al.

5.1 Accuracy Measurement and Comparison with Other

Technique

Accuracy Measurement: In this supervised learning is used which is a part of

machine learning. The techniques neural network and discriminant analysis are com-
ing under supervised learning and they are used to measure the accuracy in the present
work.
Discriminant analysis is a categorization method. It concludes that different
classes produce data, based on different Gaussian distributions:
• To create a classifier, the fitting function evaluates the parameters of a Gaussian
distribution for each class.
• To estimate the classes of new data, the trained classifier finds the class with the
smallest misclassification cost.
Figure 11 shows the comparisons between two techniques used in the work for
the calculation of accuracy of the features abstracted from the images of iris of the
given subjects. The techniques used are (1) discriminant analysis of machine learning
algorithm and (2) neural network.
The accuracy obtained from the neural network is 94.44%, whereas from discrim-
inant analysis the accuracy obtained is 99.99%.
The comparison of results obtained from the research work is well described in
Table 1. Accuracy of the features of human iris images has been derived in this
Accuracy

Different methods

Fig. 11 Matching accuracy of two mentioned methods

[Link]@[Link]
Performance Analysis of Iris Recognition System 169

Table 1 Recognition accuracy with different feature vector on CASIA image database
Methodology Accuracy (%) False acceptance rate (%)
SVM with Haar 91.33 8.66
Hamming distance 99.60 0.33
SVM with 1D log Gabor 99.65 0.33
Proposed work 1 94.76 5.21
(Haar + neural network)
Proposed work 2 94.76 5.22
(Db1 + neural network)
Proposed work 3 99.99 0.01
(Haar + discriminant analysis)
Proposed work 4 99.99 0.01
(Db1 + discriminant analysis)

proposed work after enhancing the acquired images of different subjects obtained
from the CASIA data source [7].

6 Conclusion

In this proposed work, an efficient technique for recognition and feature extraction
was explained. The crisscross collarette region of iris was picked because it is the
most significant region of iris complex layout due to which high acknowledgment
rate has been taken out. Haar wavelet and Daubechies wavelet were used for evicting
out the features; these extracted features were utilized in the iris recognition which
was using the feedforward neural network technique for recognition. The proposed
system also used the discriminant analysis approach for the matching stage with
the use of same extracted feature. This also gives better recognition rate than SVM
technique with less computational complexity. The performance accuracy of present
work is best in the favor of CASIA as well as check image database. So for both
identification and verification, the proposed work is efficient.

7 Experimental Result

Every first image in result modules represent our best performance as when we
compare characteristically features of our first image with others we find considerable
difference in image quality but the catch is that for most of images taken we have
got such technical results which are relevant to our above assumptions.

[Link]@[Link]
170 R. Khanam et al.

References

1. Daugman, J.: High confidence visual recognition of persons by a test of statistical independence.
IEEE Trans. Pattern Anal. Mach. Intell. 15(11), 1148–1161 (1993)
2. Daugman, J.: Biometric personal identification system based on iris analysis. U.S. Patent No.
5,291,560, 1 Mar 1994
3. Wolff, E.: Anatomy of the Eye and Orbit, 7th edn. H. K. Lewis & Co. LTD
4. Chinese Academy of Sciences—Institute of Automation: Database of 756 grayscale eye images.
[Link] Version 1.0 (2003)
5. Oppenheim, A., Lim, J.: The importance of phase in signals. Proc. IEEE 69, 529–541 (1999)
6. Zhu, Y., Tan, T., Wang, Y.: Biometric personal identification based on iris pattern. In: ICPR2000:
The 15th International Conference on Pattern Recognition, pp. 805–808, Barcelona, Spain
(2002)
7. CASIA iris image database. [Link]
8. Lim, S., Lee, K., Byeon, O., Kim, T.: Efficient iris recognition through improvement of feature
vector and classifier. ETRI J. 2, 61–70 (2001)
9. Daugman, J.: How iris recognition works. In: Proceedings of 2002 International Conference
on Image Processing, vol. 1 (2002)
10. Kaur, N., Juneja, M.: A review on iris recognition. In: Proceedings of 2014 RAECS, UIET
Panjab University, Chandigarh, 6–8 Mar 2014
11. Khan, M.T., Arora, D.: Feature extraction through iris images using 1-D Gabor filter on different
iris datasets. IEEE (2013). ISBN 978-1-4799-0192-0/13
12. Proenca, H.: Iris recognition: on the segmentation of degraded images acquired in the visible
wavelength. IEEE Trans. Pattern Anal. Mach. Intell. 32(8) (2010)
13. Vatsa, M., Singh, R., Noore, A.: Improving iris recognition performance using segmentation,
quality enhancement, match score fusion, and indexing. IEEE Trans. Syst. Man Cybern. (2008).
IEEE
14. Shah, S., Ross, A.: Iris segmentation using geodesic active contours. IEEE Trans. Inf. Forensic
Secur. 4(4) (2009)
15. He, Z., Tan, T., Sun, Z., Qiu, X.: Towards accurate and fast iris segmentation for iris biometrics.
IEEE Trans. Pattern Anal. Mach. Intell. (2008)
16. Ashwini, M.B., Imran, M., Alsaade, F.: Evaluation of iris recognition system on multiple feature
extraction algorithms and its combinations. Int. J. Comput. Appl. Technol. Res. 4(8), 592–598
(2015). ISSN 2319–8656
17. [Link]
18. Wildes, R., Asmuth, J., Green, G., Hsu, S., Kolczynski, R., Matey, J., McBride, S.: A system
for automated iris recognition. In: Proceedings IEEE Workshop on Applications of Computer
Vision, pp. 121–128, Sarasota, FL (1994)
19. Kong, W., Zhang, D.: Accurate iris segmentation based on novel reflection and eyelash detection
model. In: Proceedings of 2001 International Symposium on Intelligent Multimedia, Video and
Speech Processing, Hong Kong (2001)
20. Tisse, C., Martin, L., Torres, L., Robert, M.: Person identification technique using human iris
recognition. In: International Conference on Vision Interface, Canada (2002)
21. Ma, L., Wang, Y., Tan, T.: Iris Recognition Using Circular Symmetric Filters. National Labo-
ratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing
(2002)
22. Boles, W., Boashash, B.: A human identification technique using images of the iris and wavelet
transform. IEEE Trans. Signal Process. 46(4) (1998)
23. Lee, T.: Image representation using 2D Gabor wavelets. IEEE Trans. Pattern Anal. Mach. Intell.
18(10) (1996)
24. Jain, A.K., Ross, A.: Introduction to biometrics. In: Jain, A.K., Flynn, P., Ross, A. (eds.)
Handbook of Biometrics, pp. 1–22. Springer (2008). ISBN 978-0-387-71040-2

[Link]@[Link]
Performance Analysis of Iris Recognition System 171

25. Sreekala, P., Jose, V., Joseph, J., Joseph, S.: The human iris structure and its application in
security system of car. In: IEEE International Conference on Engineering Education: Innovative
Practices and Future Trends (AICERA) (2012)
26. Shah, N., Shrinath, P.: Iris recognition system—a review. Int. J. Comput. Inform. Technol. 3(2)
(2014)
27. Wibowo, E.P., Maulana, W.S.: Real-time iris recognition system using a proposed method. In:
International Conference on Signal Processing Systems, pp. 98–102, 15–17 May 2009
28. Zhiping, Z., Maomao, H., Ziwen, S.: An iris recognition method based on 2DWPCA and neural
network. In: Chinese Control and Decision Conference (CCDC ’09), pp. 2357–2360, June 2009
29. [Link]: A new texture analysis approach for iris recognition. Online Available:
[Link] (2015)

[Link]@[Link]
Challenges in Mining Big Data Streams

Veena Tayal and Ritesh Srivastava

Abstract Big data deals with data of very large data size, heterogeneous data types
and from different sources. The data is very complex in nature and having grow-
ing data. Dealing with big data is one of the emerging areas of research which is
expanding at a rapid rate in all domains of engineering and medical sciences. A major
challenge imposes on the analysis of big data is originated from big data generation
source, which generate data with very fast speed with varying data distribution due
to which the classical methods are unable to process big data. This paper discusses
the characteristics, challenges, and issues with big data mining. It also illustrates the
examples taken from various fields like medical, finance, social networking sites,
stock exchange, etc. to realize the application and importance of big data mining.
This paper explains about the use of parallel computing in data mining security
issues and how to deal with them. Furthermore, this paper also discusses challenges
associated with big streaming data with concept drifts.

Keywords Big data · Data mining · Machine learning · Online learning

1 Introduction

Data processing is a challenging task in big data mining because data is sourced from
multiple sources; moreover, they have very complex and evolutionary relationship
that is increasing at a very rapid rate. Today is an era of big data [1–3]. A very large
amount of data is produced every day. It is about thousands of billion bytes and that
data is produced, recently, within 5 years ago [4]. The capability of data generation
has enhanced so much due to advent of information technology. The example of
big data is presidential debate between President Barack Obama and Governor Mitt

V. Tayal (B) · R. Srivastava

CSE Department, FET, MRIIRS, Faridabad, India
e-mail: veena.mittal06@[Link]
R. Srivastava
e-mail: ritesh21july@[Link]

© Springer Nature Singapore Pte Ltd. 2019 173

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
174 V. Tayal and R. Srivastava

Romedy triggered more than 10 billion tweets in just two hours [5]. It is how fast
the rate on which data is expanding. There are many more examples available in
these tweets, the moments that generated discussions of public interests related to
Medicare and vouchers. It provides new means to judge the public opinion and
generates feedback in real time which is more valuable. It gives better response as
compared to radio or TV broadcasting.
It is one of the old saying “A picture depicts thousand words.” The pictures play
a very important role in exploring human society, public affairs, etc. The Flickr is an
example of public sharing sites [6]; it requires storage in various terabytes every day.
Due to large amount of data, the commonly used tools are not sufficient to capture
and manage the data, within allowed time limit. It is very challenging to extract
useful information out of such a large size data [7]. As the storage is required in
large capacity, which is almost infeasible. For example—Square Kilometer Array
[8] is used for large storage. The data generated from it is very large. It requires a
data analysis and prediction that is effective in order to achieve timely response. The
data can be sourced from multiple sources and the improvement in the efficiency
of single-source knowledge discovery methods are discussed in [9]. The dynamic
data mining methods along with its analyzes are discussed in [10]. The multisource
perspective of data mining methods as the large amount of data is sourced from
multiple locations, therefore multisource data mining as discussed in [11, 12]. The
theory proposed in these papers provides the solution for the problem of full search
but also helps in finding of global model of which traditional data mining approach
are not capable. The data classification and clustering gain a lot of importance in
data mining. Furthermore, clustering is one of the analytical methods in data mining
and k-means are one of the popular techniques of clustering. There are various
improvements suggested in the base algorithm of k-means in the past few years and
one of such improvement is discussed in the [13]. There are many machine learning
algorithms that are used for classification and prediction in various applications
like text classification, spam detection, etc. The Naïve Bayesian is an example of
machine learning approach, and the survey of this approach for the classification of
text document is discussed in [14].

2 Challenges with Big Data Mining

The challenges of big data are manifold. It focuses on how to access data and how to
apply computational operations on it related with arithmetic computing. The big data
storage location is not single but sparse at different places and the characteristics of
big data is that it is continuously growing in nature; therefore, there is a requirement
of distributed data storage at large scale. It is also expensive to move data across
different locations. It is required in case of data mining algorithms, the data, to be
loaded in main memory for computing purposes. The challenges are also related with
semantics and domain knowledge of big data. It will provide benefits to data mining
algorithm but disadvantage is that it increased complexity. It is very important to

[Link]@[Link]
Challenges in Mining Big Data Streams 175

understand the semantics at both low level and high level for example—in social
networking— users are linked and shared dependency structures but in some other
applications it can be represented using highly correlated items [15]. The challenges
in the analysis of big data mining can be discussed by considering factors [16] like
complexity, tools, data integration, cost of solution, availability of skills, etc. as shown
in Fig. 1.
These challenges also focus on designing of algorithms in handling the difficul-
ties of complex and dynamic data. The sharing of information is required for better
development and processing at each stage. The challenges of big data can also be
studied by considering the four V’s of big data, i.e., volume, velocity, variety, veracity
as shown in Fig. 2. The volume is increasing continuously in the form of clickstream,
log, event, speech, social media, etc. The velocity is speed of generation of data; it
can also be called as rate of analysis. The variety deals with different types of data
like structured, unstructured, and semi-structured. The veracity is data in doubt, dis-
trusted, and unclean. The uncertainty is due to data inconsistency and incompleteness
of data.

Fig. 1 Biggest challenges for success in big data and analytics

Fig. 2 Challenges of big

data Volume

Velocity Big Data Variety

Veracity

[Link]@[Link]
176 V. Tayal and R. Srivastava

There are many application areas of big data mining like smart health care,
finance, log analysis, traffic control, telecom, manufacturing, trading analytics,
retails, improving search quality, sentiment analysis [17–20], etc. There are various
crowdsourcing models like Google Maps, Wikipedia, etc. The information is shown
in Table 1, considering volume, velocity, variety, veracity of various crowdsourcing
models.
The framework of processing big data can be studied in various tiers as shown in
Fig. 3. Tier 1 discusses low-level detail that deals with accessing of data and how
to perform computations on them. Tier 2 concentrates on high level in the form of
domain knowledge, semantics, sharing of data, and privacy. Tier 3 deals with the
algorithms of data mining having complex and dynamic data.

2.1 Tier 1: Platform of Big Data

The mining requires intensive computing units and there for data and computing
processors are required. There is a requirement of large-scale, medium-scale, and
small-scale resources for data mining tasks. In case of small scale, a personal com-
puter is sufficient but in medium scale, data is large and cannot fit into memory.

Table 1 Crowdsourcing models

Crowdsourcing Volume Velocity Variety Veracity
models
Google maps Terabytes High Low Medium
Duollingo Terabytes Medium High High
reCAPTCHA Petabytes Very high Very high Very high
Amazon Turk Petabytes Medium Very high High
Wikipedia Petabytes Medium High Very high

Fig. 3 Framework of big

data

Tier 3

Tier 2
(Semantics, Sharing and
domain knowledge)

Tier 1
(Low level details)

[Link]@[Link]
Challenges in Mining Big Data Streams 177

It emerges the need of parallel computing [21, 22] and collective mining [23]. For
large scale, programming tools such as MapReduce is required. There is a need of
cluster computers which are having high speed. There is a need of parallel comput-
ing infrastructure for the computing of large size data. For example—data size in
Exabyte, petabyte and furthermore they have complex computing process. It changes
the quantity to quality. The parallel processing models which used to process such a
large data size includes MapReduce, cloud computing, etc. The efficiency of MapRe-
duce is further enhanced to deal with large-scale data along with its real-time nature.
There are many algorithms available for data mining. For example—k-means, Naive
Bayes, linear support vector machines, regression, etc. The data is divided into small
subsets which are independent of each other and then combine their results to produce
the desired result and intermediate results can be taken using summation on mapper
nodes. There are various tools available for data mining like R, Weka, Hadoop, etc.
The integration of R and Hadoop [24] improves the scalability of traditional analysis
and also the poor analysis capabilities. The integration of Weka and MapReduce
[25] also improves the capability of standard Weka tool. The capability of Weka
tool is to run only on single machine and that too with limited memory of 1 Gb. By
its integration with MapReduce, its memory can be extended to 100 Gb, and it is
achieved by exploring the advantage of parallel computing.

2.2 Tier 2: Semantics of Big Data

Semantics of big data is related with sharing of data, privacy, domain knowledge,
and knowledge of application. The parallel computers require sharing of information,
but it can be difficult in case of sensitive information, for example, medical records,
banking transactions credit cards, etc. The privacy issues involved in such type of
information and there can be some measures to resolve such issues, for example, by
adding information such as certificate authority that limits the access of data. The role
of domain knowledge [26] plays a very important role in big data mining algorithms.
For example, hormones test gives better and clear indication about the detection of
deceases such as NCAH, adult acne rather than just external view. Also in stock
exchange, there is great contribution to the stakeholders if the domain knowledge
related with upward/downward is there, it would be much easier to predict the market
trend rather than random guess [27]. The privacy of data is ensured by introducing
randomness in the data so that actual information can be kept hidden. There are many
methods for protecting the data and also presented in [28] such as:
• Anonymity: The suppression and generalization are two techniques used for
imposing anonymity in the data.
• Deleting sensitive values: The data values having sensitive information can be
deleted so that it can be protected. Adding random noise: The addition of some
random information that is not required is also effective method in protection of
information.

[Link]@[Link]
178 V. Tayal and R. Srivastava

• Swapping of values: The matching of values can be avoided if they are swapped.
It is very useful in the protection of sensitive information.
• Replacing of values: The original values can be replaced with some synthetic
values that are generated from simulations.
The distribution of data is at different locations; therefore, users do not have
physical access to such data. The third-party data miners require efficient data access
mechanism. There should no downloading allowed of any local data copies. There
should not be having direct access to original data.

2.3 Tier 3: Complexity of Data: Relationships

and Associations

The activities related to distributed sites are also an issue to produce unbiased view
from them. It is done at various levels; data level, model level, and the knowledge
level. The data level is related with local sites, remote site and to aggregate the
information at different sites. The data is multisource, dynamic and massive in nature.
There is emerging requirement for the expansion of data mining methods. The gradual
improvement in the hardware allows improving the knowledge discovery methods,
so that they can work well for massive data. There are a lot of differences between
single source and multisource data mining due to its real time, heterogeneous and
complex in nature like characteristics. The application of data mining such as financial
analysis, medical field, online trading diagnostic, etc. are inherently very dynamic in
nature where static knowledge is not capable of adapting the change of dynamism.
Evolving knowledge is a continuous process in real time and dynamic applications,
for example, in a class of average students any student among them can perform
extraordinary well, may be because of various external factors like affording of
various expensive coaching, guidelines receive from expert, etc. and various other
factors with the passage of time. It gives birth to knowledge discovery in case of
concept drift. There can be various types of drifts possible in it. It gives rise to new
area of research which is called as concept drifting data stream mining [29–31],
the detail discussion on concept drifting data stream is present in Sect. 3. There are
various types of complex data in big data [15].
• Complex and heterogeneous data: It is having structured data and unstructured
data. It is having data in the form of table, image, audio, video, hypertext, etc. The
data models are incapable of handling such a complex nature of data.
• Complex relationship networks in data: There is existence of social relationship
between individual and that comes under the category of complex social network.
The examples include Facebook, twitter, LinkedIn, and other social networks.
• Complex intrinsic semantics associations in data: The “text–image–video” is
having strong semantics associations. It helps to improve the performance of appli-
cations. It is very challenging to describe the semantic features.

[Link]@[Link]
Challenges in Mining Big Data Streams 179

3 Big Data Streams and Concept Drifts

3.1 Data Stream and Types of Concept Drifts

Big data stream is time-ordered sequence of instances generated by a data generating

source at very high speed. It is often in the case of data stream that its underlying
statistical distribution also changes very frequently with time. Such underlying dis-
tribution data stream is also known as concept drifting data stream. The concept
drifts are formally defined as the change in underlying statistics of data stream. The
classical methods of classification are not suitable for concept drifting data streams.
The concept drifts can be broadly categorized under the following five categories as
shown in Fig. 4.

3.2 Methods for Dealing with Concept Drifting Data Stream

The approaches for concept drifting data stream mining can be broadly classified into
three categories as shown in Fig. 5 namely window-based methods, drift detection-
based methods and ensemble-based methods.
• Window-Based Methods: The window-based method is a kind of incremental
learning in which the classifier is trained by incoming instances for a fixed length
of window. Mostly, sliding window is used to track the time-varying data instances.
The forgetting strategy is the key practice of such method. By forgetting the old
learned concept is dropped and the new concept is learned.
• Drift detection-based methods: In drift detection-based methods, the rate of
change in the accuracy of the classifier is monitored continuously and any signifi-
cant drop in the accuracy above a threshold value is taken as a drift in the concept.
If drift is detected, the classifier is rebuilt to accommodate the change in concept.
• Ensemble-based methods: In ensemble-based methods, a set of diverse compo-
nent classifiers are pooled together. While performing the classification, all com-
ponent classifiers participate in decision making according to their weights and
the decision about the any income instances is taken by using decision integration
rule like weighted majority.

4 Conclusion

The big data mining is very complicated task and challenging too. It is because of
real time, complex, multisource, and hetero generous nature of data. The mining of
such a big data and with such type of characteristics requires big mind to fully exploit
the benefit out of it. We have discussed the challenges faced in big mining at data

[Link]@[Link]
180 V. Tayal and R. Srivastava

1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2
0.2
0
1 3 5 7 9 11 13 0
1 3 5 7 9 11 13
Sudden Recurring
1
1
0.8
0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
1 3 5 7 9 11 13 1 3 5 7 9 11 13
Abrupt Incremental

0.8

0.6

0.4

0.2

0
1 3 5 7 9 11 13
Gradual

Fig. 4 Types of concept drifts in data stream

level, models level, and system level. It requires devices with high computing power
which is both effective and efficient. The challenging situation in big data mining is
issue of privacy, random noise and errors, missing data, etc. which are the problems
faced at data level. The challenge faced at model level is how to integrate the subsets
locally to form a large unified view. The problem of communication at distributed
sites which are multisource arises. The problem at system level is to consider the
relationship between complex data, semantics, along with their evolving changes.

[Link]@[Link]
Challenges in Mining Big Data Streams 181

Fig. 5 Methods for dealing with concept drifting data stream

The system should be designed in such a manner that it can link unstructured data to
extract useful information from it considering their complex relationships. It should
from the basis of prediction of future trend. Lastly, big data is an emerging field
and its need is arising in all the fields of engineering and science domains. With the
advent of big data technique, we can better understand our society by analyzing the
big data available on various social networking sites. It increases the participation of
public audiences to receive response in the various social-economic events. It proves
today is an era of big data. This paper also explains the challenges associated with
concept drifting big data streams, types of concept drifts, and the methods used to
deal with concept drifting data streams.

[Link]@[Link]
182 V. Tayal and R. Srivastava

References

1. Nature Editorial: Community cleverness required. Nature 455(7209), 1 (2008)

2. Howe, D., et al.: Big data: the future of biocuration. Nature 455, 47–50 (2008)
3. Labrinidis, A., Jagadish, H.: Challenges and opportunities with big data. Proc. VLDB Endow-
ment 5(12), 2032–2033 (2012)
4. IBM: What is big data: bring big data to the enterprise. [Link]
bigdata/ (2012)
5. Blog, T.: Dispatch from the Denver debate. [Link]
[Link] (2012)
6. Michel, F.: How many photos are uploaded to flickr every day and month? [Link]
com/photos/franckmichel/6855169886/ (2012)
7. Rajaraman, A., Ullman, J.: Mining of Massive Data Sets. Cambridge University Press (2011)
8. Dewdney, P., Hall, P., Schilizzi, R., Lazio, J.: The square kilometre array. Proc. IEEE 97(8),
1482–1496 (2009)
9. Chang, E.Y., Bai, H., Zhu, K.: Parallel algorithms for mining large-scale rich-media data. In:
Proceedings of the 17th ACM International Conference on Multimedia (MM’09), pp. 917–918
(2009)
10. Domingos, P., Hulten, G.: Mining high-speed data streams. In: Proceedings of the Sixth ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD’00),
pp. 71–80 (2000)
11. Wu, X., Zhang, S.: Synthesizing high-frequency rules from different data sources. IEEE Trans.
Knowl. Data Eng. 15(2), 353–367 (2003)
12. Wu, X., Zhang, C., Zhang, S.: Database classification for multi-database mining. Inf. Syst.
30(1), 71–88 (2005)
13. Na, S., Xumin, L., Yong, G.: Research on k-means clustering algorithm: an improved k-means
clustering. In: Third International Symposium on Intelligent Information Technology and Secu-
rity Informatics, pp. 63, IEEE (2010)
14. Vidhya, K.A., Aghila, G.: A survey of Naïve Bayes machine learning approach in text document
classification (IJCSIS). Int. J. Comput. Sci. Inf. Secur. 7(2) (2010)
15. Wu, X., Zhu, X., Wu, G., Ding, W.: Data mining with big data. IEEE Trans. Knowl. Data Eng.
26(1), 97–106 (2014)
16. TM Forum: Challenges of big data (2012)
17. Srivastava, R., Bhatia, M.: Offline vs. online sentiment analysis: issues with sentiment analysis
of online micro-texts. Int. J. Inf. Retrieval Res. (IJIRR) 7(4), 1–18 (2017)
18. Srivastava, R., Bhatia, M.: Real-time unspecified major sub-events detection in the twitter data
stream that cause the change in the sentiment score of the targeted event. Int. J. Inf. Technol.
Web Eng. (IJITWE) 12(4), 1–21 (2017)
19. Srivastava, R., Bhatia, M.: Challenges with sentiment analysis of on-line micro-texts. Int. J.
Intell. Syst. Appl. 9(7), 31 (2017)
20. Srivastava, R., Bhatia, M.: Ensemble methods for sentiment analysis of on-line micro-texts.
Presented at the International Conference on Recent Advances and Innovations in Engineering
(ICRAIE) (2016)
21. Shafer, J., Agrawal, R., Mehta, M.: SPRINT: a scalable parallel classifier for data mining. In:
Proceedings of the 22nd Conference on VLDB (1996)
22. Luo, D., Ding, C., Huang, H.: Parallelization with multiplicative algorithms for big data mining.
In: Proceedings of the IEEE 12th International Conference on Data Mining, pp. 489–498 (2012)
23. Chen, R., Sivakumar, K., Kargupta, H.: Collective mining of Bayesian networks from dis-
tributed heterogeneous data. Knowl. Inf. Syst. 6(2), 164–187 (2004)
24. Das, S., Sismanis, Y., Beyer, K.S., Gemulla, R., Haas, P.J., McPherson, J.: Ricardo: integrating
R and Hadoop. In: Proceedings of the ACMSIGMOD International Conference on Management
Data (SIGMOD’10), pp. 987–998 (2010)

[Link]@[Link]
Challenges in Mining Big Data Streams 183

25. Wegener, D., Mock, M., Adranale, D., Wrobel, S.: Toolkit-based high-performance data mining
of large data on MapReduce clusters. In: Proceedings of the International Conference on Data
Mining Workshops (ICDMW’09), pp. 296–301 (2009)
26. Kopanas, I., Avouris, N., Daskalaki, S., The role of domain knowledge in a large scale data
mining project. In: Vlahavas, I.P., Spyropoulos, C.D. (eds.) Proceedings of the Second Hellenic
Conference on AI: Methods and Applications of Artificial Intelligence, pp. 288–299 (2002)
27. Bollen, J., Mao, H., Zeng, X.: Twitter mood predicts the stock market. J. Comput. Sci. 2(1),
1–8 (2011)
28. Machanavajjhala, A., Reiter, J.P.: Big privacy: protecting confidentiality in big data. ACM
Crossroads 19(1), 20–23 (2012)
29. Mittal, V., kashyap, I.: Online methods of learning in occurrence of concept drift. Int. J. Comput.
Appl. 117(13), 18–22 (2015)
30. Mittal, V., Kashyap, I.: Empirical study of impact of various concept drifts in data stream
mining methods. Int. J. Intell. Syst. Appl. 8(12), 65 (2016)
31. Mittal, V., Kashyap, I.: An overview of real world applications with concept drifting data
streams (2018)

[Link]@[Link]
Controlling of Non-minimum Phase
System Using Harmony Search
Algorithm

Vivek Kumar Jaiswal, Anurag Singh, Shekhar Yadav

and Shyam Krishna Nagar

Abstract In this paper, a non-minimum phase system with dead time is controlled
by intending the optimized proportional, integral, and derivative controller (PIDC).
A system problem is formulated to the non-minimum phase system in which zeroes
in the right half-plane (RHP) make system insignificant as delay raises. To enhance
the system performance, the factors of the conventional PIDC are optimized by a
heuristic algorithm (HA), chosen harmony search (HS). HA, copying the invention of
music players, chosen harmony search algorithm (HSA). The HSA looks to acquire
a global optimal magnitude of the conventional PIDC and genetic algorithm (GA)
in the area portrayed by the regular PID controller and GA. The simulation results
demonstrate the transient responses such as settling, rise, peak time, undershoot, and
overshoot. Thus to optimized the parameters by minimizing integral square error
(ISE) of the given system is improved by the proposed method.

Keywords Non-minimum phase (NMP) system · Conventional PID controller

Harmony search algorithm (HSA) · Harmony memory considering rate (HMCR)
Pitch adjusting rate (PAR)

V. K. Jaiswal · A. Singh (B) · S. Yadav

Department of Electrical Engineering, MMMUT, Gorakhpur, India
e-mail: anuragsinghucer123@[Link]
V. K. Jaiswal
e-mail: vivekjaiswal190@[Link]
S. Yadav
e-mail: syee@[Link]
S. K. Nagar
Department of Electrical Engineering, IIT (BHU), Varanasi, India
e-mail: [Link]@[Link]

© Springer Nature Singapore Pte Ltd. 2019 185

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
186 V. K. Jaiswal et al.

1 Introduction

In spite of the fact that the advanced control hypothesis has been connected to numer-
ous regions, for example, robots and airplane, and the industry universe are as yet
ruled by the crude PIDC. Over 90% of controllers utilized as a part of plant control
are yet PIDC. People may ask why the interval is so huge. By looking at the two
perspectives, one may discover the appropriate response [1]. For a PIDC outline, no
data about the given system is expected, which implies the progression of the given
system is thought to be obscure. This really constructs it simple to execute and is more
useful. For control configuration classical, i.e., established or modern—present-day
control theory, a numerical model of the system is required at the absolute starting
point, which implies a large portion of the elements of the system is thought to be
known. The achievement of the conventional PID controller acquired utilizing har-
mony search (HS) algorithm is verified by controlling the NMP system with the delay
in time. And basically, the stable system is easy to control than the unstable system.
A system having at least one zero in the right half-plane (RHP) of s domain is called
NMP stable system, and when the delay in time raises in the system, the scope of
controller gain moves toward becoming smaller; in this manner, the response of the
system will additionally be decayed. Accordingly, the basic criterion for NMP stable
system may not be obtainable.
For controlling the unstable system without or with the delay in time, researchers
are working progressively. There are many techniques for controlling unstable system
with or without delay, and numerous researchers have watched the traditional PIDC
less powerful to enhance the response of the NMP dead-time system [2–4].
So, to enhance the overall response, dead-time or delay-time compensating (DTC)
model is used in the system. There are many researchers who have used the Smith
predictor (SP) and extensions of SP to control time delay system. In this, SP presence
of delay in time is wiped out from the given characteristic equation of the closed-
loop (CL) system. From these advantages of SP, a time delay system is converted into
without time delay system. Although, disadvantages of SP it cannot be employed to
open-loop (OL) unstable system [5–9].
In this paper, the given system is optimized to get a better transient response
of the given system using the HSA. Music is a standout among the most fulfilling
forms created by human effort. Another HA created by an artificial event establish
in music response, in particular, the way toward hunting down better harmony, can
be presented.
This paper has been organized as follows. Section 2 exhibits the concise depiction
of NMP systems with dead time. In Sect. 3, conventional PIDC is displayed. In Sect. 4,
HSA is clarified. Section 5 shows the result and discussion of this paper. Section 6
introduces the conclusion of this paper with related references to this paper.

[Link]@[Link]
Controlling of Non-minimum Phase System Using Harmony … 187

2 Non-minimum Phase System with Dead Time

In control theory, the nonlinear systems which have at least one system zero in the
RHP. A system recognized in this paper is fine-modeled by suitable linear time-
invariant (LTI) systems by feasible NMP components and delay in time. For any
control system, internal stability is the essential criterion of actual CL system that
can be calibrated by testing the observability and controllability gramians of a given
system. And G(s) is the transfer function of the system [2]:
K N (s) K Nmp(s)N nmp(s) −sT
G(s)   e (1)
D(s) Ds(s)Du(s)

where K stands for gain, Nmp(s) is minimum phase, Nnmp(s) is NMP, Du(s) and
Ds(s) are unstable and stable polynomials, and T stands for time delay. In NMP
systems, beginning response of the system is sluggish because of the presence of
RHP zero in the system, and dead time is called the transport lag, and at higher
frequency, thundering phase lags along no attenuation. The amplitude is always
unity or 0 dB in log amplitude, and phase angle changes linearly along frequency of
transport lag [3].

|G( jω)|  |cos ωT − j sin ωT |  1 (2)

 G( jω)  −ωT (rad.) (3)

3 Typical PID Controller

PIDC is widely used in industrial control. The universality of PIDC can be certified
most of the way to its strong implementation in a broad variety of working conditions
and halfway to its reasonable ease that empowers the designer to work them in
coordinate and an essential manner. As the name proposes, PID controller includes
three basic terms: proportional, integral, and derivative which are given to get a
perfect response [10].
In Fig. 1, e(t) indicates error signal which is the difference between R(t), reference
input signal, and Y (t), obtained output signal. The e(t) works as input to PIDC and
output of the PIDC control signal, u(t), to be applied to given plant that is equivalent
to proportional gain (K P ) times the amplitude of e(t) accumulation to the integral
gain (K I ) times the fundamental of e(t) in accumulation to the derivative gain (K D )
times the derivative of e(t).
Thus, time domain demonstrations of u(t):

de(t)
u(t)  K P e(t) + K I e(t)dt + K D (4)
dt

[Link]@[Link]
188 V. K. Jaiswal et al.

Fig. 1 Typical block

diagram of PID controller

A plant on accepting the input u(t) will create a changed output Y (t) which will
be again contrasted with the reference signal until the point that the desired level is
come to in this way shaping a nearby closed-loop system [11].

4 Description of Harmony Search Algorithm

Presently, the time has come to make an inquiry. Is it conceivable to build up another
heuristic algorithm with better response (better results with less iteration) than exist-
ing heuristic calculations? This algorithm would fill in as an appealing contrasting
option to another effectively settled algorithm [11].
Since the 1970s, numerous heuristic calculations have been created that consoli-
date standards and haphazardness copying natural event. These procedures or tech-
niques incorporate simulated annealing (SA), physical annealing (PA), evolutionary
algorithms (EA), human memory, etc. Consequently, another calculation may like-
wise be based on a natural event, or in synthetic one. A synthetic event, musical or
melodic harmony, can fill in as the mold for concocting another technique. Harmony
or music is a standout among the most fulfilling forms created in human being’s
effort [12]. Another heuristic algorithm is obtained from a synthetic event found in
melodic response (for instance, a jazz triad); in particular, the way toward looking
for better harmony with less iteration can be presented.
Music concordance is a mix of sounds thought about satisfying from an elegant
perspective. There is a peculiar connection between a few sound effects that have
distinctive frequencies in nature of harmony. Melodic responses look for an optimum
state (awesome harmony) controlled by elegant estimation, and the optimization
techniques look for an optimum state (state of global optimum: maximum gain with
minimum worth) dictated with objective function assessment. Elegant evaluation is
controlled with the arrangement of sound played with connected instruments, and
similarly, objective function assessment is dictated by the arrangement of values
created by factors variables; for better sound, elegant estimation could be enhanced
with a great many practices; similarly, superior objective function assessment could

[Link]@[Link]
Controlling of Non-minimum Phase System Using Harmony … 189

be enhanced by many iterations [13–16]. Short introductions of these perceptions

appear inside Table 1.
Harmony search (HS) algorithm procedure can be understood with the help of
flowchart in this way:
Step 1: Initializing a harmony memory (HM).
Step 2: Improvise or create a new harmony from step 1.
Step 3: In this step, if the new harmony is superior to anything least harmony in HM,
incorporate the new and exclude the least harmony in HM.
Step 4: In this step, if halting norms are not fulfilled, then again follow step 2.
The global optimum solution lies at first in HM. At the point when this is not
the situation, keeping in mind the end goal to search global optimum, HS starts a
parameter, HMCR, which ranges between 0 and 1. Let a HMCR of 0.90 implies on
that subsequent stage, the algorithm picks a variable incentive from HM with 90%
probability [17, 18].
Enhancing the results and getting away local optima, the additional choice might
be presented. Let aPAR 0.20, it means this algorithm picks an adjacent with 20%
probability (lower and upper values by ±10%) [19] (Fig. 2).
The proposed parameters of the conventional PIDC, GA, and HSA are shown in
Table 2.

5 Result and Discussion

For controlling the NMP time delay system by optimized PIDC [2], consider a
transfer function of actual NMP system with time delay:

Table 1 Comparison of musical and optimization response

Comparing factors Optimization procedure Response procedure
Optimum state Globally optimum Awesome harmony
Estimated by Objective function Elegant norm
Estimated with Variables values Pitch of musical instruments
Procedure unit Every iteration Every practice

Table 2 Gain of the Parameters HAS GA PID

controllers
KP 4.4295 4.9043 5.1610
KI 0.0073 0.1975 0.2775
KD 20.5130 23.8970 24.0
ISE 9.7605 11.9572 13.2720

[Link]@[Link]
190 V. K. Jaiswal et al.

Fig. 2 Flowchart of HSA

0.1
G(s)  G(s)  e−8s (5)
0.05s 4 + 1.25s 3 + 5.2s 2 + 4s
Step response of actual NMP system is shown in Fig. 3.
Approximation of the system using Pade’s:
First of all, the given NMP system transfer function with dead time is approximated
with the help of Pade’s first-order approximation in MATLAB with command Pade
(G,1) that produces the approximated transfer function as shown in Eq. 6, and step
response of the system is shown in Fig. 4.
−0.1s + 0.025
G(s)  (6)
0.05s 5 + 1.263s 4 + 5.513s 3 + 5.3s 2 + 5
Step response of approximated system using PIDC:

When we used PIDC to approximate a system, we got the result that settling time
of system is reduced but the maximum overshoot and undershoot of the system are

[Link]@[Link]
Controlling of Non-minimum Phase System Using Harmony … 191

increased as shown in Fig. 5. These problems are overcome by using evolutionary

techniques.
Step response of the system using HSA and GA:

As shown in Fig. 6, comparison of step response of different algorithms can be

understood with the help of Table 3.

Fig. 3 Actual response of

NMP system

Fig. 4 Pade’s approximated

response of NMP system

[Link]@[Link]
192 V. K. Jaiswal et al.

Fig. 5 Step response of

NMP system with PIDC

Fig. 6 Comparison of step

response of NMP system

6 Conclusion

In this paper, the performance of HSA is considered for optimizing the parameters
of PIDC. The proposed PIDC is optimized by HSA to control the NMP system. The
given NMP system is approximated with Padé’s first-order approximation. The HSA
consolidates ISE for optimizing the factors of the PIDC. From the above discussion,

[Link]@[Link]
Controlling of Non-minimum Phase System Using Harmony … 193

Table 3 Comparison of response specification of different technique

Specification Approximated PIDC GA HSA
system
Peak time 229.8469 22.8672 26.5317 14.5837
Settling time 123.90 74.8433 58.6122 11.7892
Overshoot 0 37.5177 27.7547 1.6629
Undershoot 2.8013 46.6808 45.5332 37.6136
Rise time 65.4610 3.2765 3.4684 5.0763

it has been noticed that HSA gives the global optimum results and enhances the
characteristics compared to other evolutionary algorithms like GA and conventional
control techniques like PIDC tuned using Ziegler–Nichols method.

References

1. Hagglund, T.: An industrial dead-time compensating PI controller. Control Eng. Pract. 4(6),
749–756 (1996)
2. Verma, S.K., Yadav, S., Nagar, S.K.: Optimized fractional order PID controller for non-
minimum phase system with time delay. In: International Conference on Emerging Trends
in Electrical Electronics and Sustainable Energy Systems (ICETEESES), pp. 169–173, IEEE,
Sultanpur (2016)
3. Yadav, S., Verma, S.K., Nagar, S.K.: Design of an optimized PID controller for non-minimum
phase system with delay. In: 2015 39th National Systems Conference (NSC), IEEE (2015)
4. Palmor, Z.J.: Time-delay compensation Smith predictor and its modifications. In: Levine, W.
(ed.) The Control Handbook, pp. 37–48. CRC Press, Boca Raton, FL (1996)
5. Watanabeand, K., Ito, M.: A process-model control for linear systems with delay. IEEE Trans.
Autom. Control. 26(6), 1261–1269 (1981)
6. Ogata, K.: Modern Control Engineering, 4th edn, Prentice Hall, New York (2002)
7. Gu, K., Niculescu, S.I.: Survey on recent results in the stability and control of time-delay
systems. J. Dyn. Syst. Meas. Control 125(2), 158–165 (2003)
8. Richard, J.P.: Time-delay systems: an overview of some recent advances and open problems.
Automatica 39(10), 1667–1694 (2003)
9. García, P., Albertos, P.: A new dead-time compensator to control stable and integrating pro-
cesses with long dead-time. Automatica 44(4), 1062–1071 (2008)
10. Åströmand, K.J., Hägglund, T.: PID Controllers: Theory, Design, and Tuning, vol. 2, Research
Triangle Park, NC: Instrument Society of America (1995)
11. Chidambaram, M., Sree, R.P.: A simple method of tuning PID controllers for integrator/dead-
time processes. Comput. Chem. Eng. 27(2), 211–215( 2003)
12. Deb, K.: Multi-objective Optimization Using Evolutionary Algorithms, vol. 16, John Wiley &
Sons (2001)
13. Geem, Z.W., Kim, J.H., Loganathan, G.V.: A new heuristic optimization algorithm: harmony
search. Simulation 76(2), 60–68 (2001)
14. Lee, K.S., Geem, Z.W.: A new structural optimization method based on the harmony search
algorithm. Comput. Struct. 82(9–10), 781–798 (2004)
15. Lee, K.S., Geem, Z.W.: A new meta-heuristic algorithm for continuous engineering optimiza-
tion: harmony search theory and practice. Comput. Methods Appl. Mech. Eng. 194(36–38),
3902–3933 (2005)

[Link]@[Link]
194 V. K. Jaiswal et al.

16. Mahdavi, M., Fesanghary, M., Damangir, E.: An improved harmony search algorithm for
solving optimization problems. Appl. Math. Comput. 188(2), 1567–1579 (2007)
17. Geem, Z.W. (ed.): Music-Inspired Harmony Search Algorithm: Theory and Applications, vol.
191, Springer (2009)
18. Yang, X.S.: Harmony search as a metaheuristic algorithm. Music-Inspired Harmony Search
Algorithm, pp. 1–14, Springer, Berlin, Heidelberg (2009)
19. Yang, X.S.: A new metaheuristic bat-inspired algorithm. Nature Inspired Cooperative Strategies
for Optimization (NICSO 2010), pp. 65–74, Springer, Berlin, Heidelberg (2010)

[Link]@[Link]
Framework for Real-World Event
Detection Through Online Social
Networking Sites

Ritesh Srivastava, M. P. S. Bhatia, Veena Tayal and J. K. Verma

Abstract In recent few years, due to the exponential growth of users on online
social networking sites (OSNs), mainly over micro-blogging sites like Twitter, the
OSNs now resemble the real world very cohesively. The excess of continuously
user-generated online textual data by OSNs that encapsulates almost all verticals
of the real world has attracted many researchers who are working in the area of
text mining, natural language processing (NLP), machine learning, and data mining.
This paper discusses the feasibility of OSNs in detecting real-world events from the
horizon of the virtual world formed over OSNs. Moreover, this paper also describes
the framework for real-world event detection through online social networking sites.

Keywords Online social network · Event detection · Social network analysis

Data mining · Text mining

1 Introduction

The evolution of Web 2.0 [1] allows users to interact and collaborate with each other
on OSNs platform [2]. In recent years, an exponential growth in the users of OSNs
has been witnessed. The numbers of users on OSNs are getting double in every five
years with about 2.5 billion users this year. A projected result shows that about 39.9%

R. Srivastava (B) · V. Tayal

CSE Department, FET, MRIIRS, Faridabad, India
e-mail: ritesh21july@[Link]
V. Tayal
e-mail: veena.mittal06@[Link]
M. P. S. Bhatia
NSIT, University of Delhi, New Delhi, India
e-mail: mpsbhatia@[Link]
J. K. Verma
Galgotias University, Greater Noida 201310, India
e-mail: [Link]@[Link]

© Springer Nature Singapore Pte Ltd. 2019 195

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
196 R. Srivastava et al.

of world population will become OSN users with the end of 2020. The users of micro-
blogging sites (like Twitter) actively participate in sharing their opinions on various
hot topics (e.g., personalities, products, and events) by posting their comments about
the topics. Usually, the comments written by users of micro-blogging sites are short
snippets of text that are limited to only a few characters. These short snippets of textual
comments over OSN are also termed as online micro-texts [3]. The consistent posting
of comments by millions of users of micro-blogging sites and the exponential increase
in the number of users on micro-blogging sites have flourished two interesting aspects
of social-networking-enabled micro-blogging sites:
(i) The WWW has now become a massive source of online micro-texts. Micro-
texts are extremely subjective in nature as they generally contain the positive or
negative opinion of users about an entity. Data mining for searching some inter-
esting information from such data has gained the affection of many researchers
in the previous years. Sentiment analysis (SA) is one of the most prominent
ways for the analysis of this valuable collection of subjective data for making
predictions in various events.
(ii) Another interesting facet of OSN is the creation of virtual communities. A
virtual community is a set of social entities, especially human beings that are
connected to each other on the basis of some common interests on any topics
and events over OSN [2].
With the exponential growth of users on OSNs, the OSNs now resemble the cor-
responding real-world community very cohesively. As a conscience, any event that
may be initiated in any of these communities has a significant reaction in both the
communities and vice versa. This cohesiveness among the virtual and its correspond-
ing real community has motivated many researchers and data analyst to sense the
happenings of the real-world events through the contents of OSN.
Knowing about future has always been fascinating. Predictive analytics is a way
by which one can predict the unknown future events. In the predictive analysis,
historic and the current data are analyzed to make a prediction, which utilizes many
methods such as statistics, data mining, and machine learning. The problem of the
predictive analysis can be abstractly classified into two different set of problems.
(i) Making predictions of future by utilizing the current data.
(ii) Making predictions of some attributes of one observation space from another
observation space at the same time.
Nowadays, the OSNs offer great opportunities for making predictions about the
real-world happenings in both cases. From Fig. 1, it can be easily understood that
the degree of cohesion (d) between a real-world community and its corresponding
online virtual community formed over OSN is directly proportional to the number of
active users on the online virtual community and the number of users commenting
about an event occurred in the real world as given in (1) and (2).

d ∝ Number of active users on Online Virtual Communitiy (1)

d ∝ Number of users commenting about an event (2)

[Link]@[Link]
Framework for Real-World Event Detection … 197

Fig. 1 Degree of cohesion between real world and virtual world of OSN

The consistent posting of comments by millions of users on micro-blogging sites

and the exponential increase in the number of users on micro-blogging sites have
made the WWW a massive collection of online micro-texts and increased the degree
of cohesiveness between the real-world communities and their corresponding online
virtual communities, respectively. The availability of massive data on OSN that
represents the social and behavioral aspects of a big section of the population provides
the immense possibilities to us to observe the public behaviors from the virtual
world formed over OSN. Furthermore, it also opens up new opportunities to conduct
predictive analysis for the future events. Consequently, an emerging area of research
is to make the prediction of real-world happenings and behaviors of people in the
real world by analyzing their behaviors in the virtual worlds of OSNs.
The social networking sites have proven their huge power of prediction in predict-
ing the results of the events of real world. Recently, Twitter is utilized in various tasks
such as monitoring, predicting, and analyzing the real-world events and activities like
breaking news tracking [4], election prediction [5], natural disasters like earthquake
[6], and crime, radicalization, and terrorism [7]. Certainly, the text stream mining of
Twitter data can substitute the traditional polling [8].
We believe that the real-world activities could be monitored in real time by per-
forming the real-time analysis of contents of micro-blogging site like Twitter. Such
type of real-time analysis can enhance the real-time decision-making and an alter-
native for both types of predictive analysis as mentioned above.

[Link]@[Link]
198 R. Srivastava et al.

2 Events and Event Detection Through OSN

The Oxford Dictionary defines the word event as a thing that happens or takes place,
especially one of importance. An event is usually associated with time and location.
The process of event detection from temporally ordered textual data can be explained
as an automatic process for identifying novel events evolved meanwhile in that textual
stream [9]. In early years of 2000s, with the evolution of Web 2.0, the enormous use
of computer-mediated communications motivated researchers for automatic event
detection from user-generated text stream. The event detection from textual data
has long been discussed as topic detection and tracking (TDT) [9–11]. Most of the
previous works related to event detection are implemented on the conventional news
based textual contents from various news broadcasting media [12–14].
The task of event detection in the Twitter data stream can be broadly categorized
into two: (i) specified or targeted events and (ii) unspecified (untargeted) events [15].
The targeted event detection is a kind of supervised process in which a sufficient
amount of prior information is known in advanced; based on this information, the
events are detected. For specified event detection, the prior information may include
place, time, domain, description, and features about the event. For example, election
in any country is kind of specified event as the date of the election, names of contesting
parties as well as the names of the candidates are known in advance for performing
the analysis. Conversely, in the case of unspecified event detection process, no clues
about the event are known in advance. Moreover, an event can undergo with some sub-
events. Sub-events can be defined as those events that occur in between the discussion
duration of any event and cause significant impact over the points of discussion. Sub-
events generally change the sentiment of the main event significantly. We state such
event as sentimental events or sub-events. The task of event detection in the Twitter
data stream can be broadly categorized in two: (i) specified or targeted events and (ii)
unspecified (untargeted) events (Fig. 2) [15]. The targeted event detection is a kind
of supervised process in which a sufficient amount of prior information is known
in advance; based on this information, the events are detected. For specified event
detection, the prior information may include place, time, domain, description, and
features about the event. For example, election in any country is a kind of specified
event as the date of the election, names of contesting parties as well as the names of
the candidates are known in advance for performing the analysis. Conversely, in the
case of unspecified event detection process, no clues about the event are known in
advance. Moreover, an event can undergo with some sub-events. Sub-events can be
defined as those events that occur in between the discussion duration of any event and
cause significant impact over the points of discussion. Sub-events generally change
the sentiment of the main event significantly.

[Link]@[Link]
Framework for Real-World Event Detection … 199

Fig. 2 Types of event detection task on OSN

2.1 Types of Events on OSN

Based on the ways, in which the events evolved in real world and discussed on OSN,
we can classify the events on OSN into the following three types:
(i) Periodic events.
(ii) Sudden events.
(iii) Long-duration gradual events.
As depicted in the graphs of Fig. 3, the periodic events are those events that reoc-
curred periodically on certain pre-decided date and time. Over OSNs, such type of
events hugely attracts comments of users on the date or time of event occurrence
periodically. The periodic events are generally designated by certain prior infor-
mation such as a list of frequent keywords and time of occurrence. For example,
#FollowFriday is a periodic event that involves discussions on new movie release on
every Friday of a week. The periodic events can further be classified on the basis
of their recurrence intervals; for example, daily event may be (#GoodMorning), the
weekly event may be follow Friday (#FollowFriday), and the yearly event may be
any festival (#HappyChrismas). The next type of events is sudden events; such type
of events gains the sudden interest of users of OSN in their posts after the occurrence
of the events. For example, after the occurrence of an earthquake, the user’s posts
mentioning the word earthquake increase suddenly. Unlike the periodic events, such
events do not offer any prior information concerning the time and the place of events;
moreover, in many cases, they do not designate any predefined keywords or hashtags
keywords also. Terrorist attacks and riots generally belong to such category of events.
Another essential category of events discussed on OSN is long-duration gradual
events. Such kinds of events are specified and designated by the date, place, and other
related information such as keywords and terminologies in advance. The discussions
about a topic in a long-duration event persist for a long period. The users on OSN have
prior knowledge of such kind of events. An example of such kind of events includes
an election in any country. Usually, the date of an election is declared in advance. The
users of OSN start discussing the election few days prior to the election (e.g., one

[Link]@[Link]
200 R. Srivastava et al.

Fig. 3 Types of events on OSN

month) and persist it throughout the election campaign. Some major keywords for an
election are comprised of contesting party names, contesting candidates’ names, etc.

3 Framework for New Event Detection in OSN

A generic framework for new event detection is text data stream process which is
a two-step process as depicted in Fig. 4. The first step is responsible for feature-
based signal generation from the input text stream. The second step is responsible
for detecting the burst in the signals.
(i) Signal Generation: The signal generation depends on features of the incoming
text streams. Identification of the best features of the incoming text stream is
very crucial for accurate event detection.

Fig. 4 Model for event detection in text data stream

[Link]@[Link]
Framework for Real-World Event Detection … 201

(ii) Burst Detection Strategy: Any unprecedented change in the observed signal
generated by signal generator the incoming text streams can be considered as a
burst in the feature-based signals. Bursts in the generated signal are the potential
indicators of the occurrence of events.
The rest parts of this section describe various strategies widely used for signal
generation and burst detection in textual data stream.

3.1 Features for Signal Generation in Text Data Stream

1. Frequency of words: The most prominent burst detection methods in text data
streams relied on the burst detection based on the examination of the frequencies
of words present in the text stream [16–18]. The term frequency and inverse
document frequency (TF-IDF) is the most popular method for generating term-
frequency-based signal. The TF-IDF can be explained as follows: Let d be a
document in a corpus of N documents and let t be a term in documents, then
TF-IDF can be calculated by using (3).

TF − I D Ft, d  1 + log T Ft, d × I D(t) (3)

where TF(t, d) is the term frequency of t in d, and IDF(t) is the inverse document
frequency, i.e.,N /n(t), where n(t) is the number of documents containing the
term t.
In Twitter data streams, instead of calculating the TF-IDF, the Term Frequency
and Inverse Tweet Frequency (TF-ITF) has been calculated for a given temporal
window. For the real-time event detection, the TF-ITF is monitored periodically
and any unprecedented changes in the TF-ITF are recorded as the bursts in the
signal.
2. Platform-Specific Features: Online social networking sites like Twitter gener-
ally offer some specific notations to emphasize the topic of discussion in order
to grasp the attention of other users. The notion of hashtags (#) is widely used
by almost all OSN these days. Hashtags are generally created by users to indi-
cate event or issue and floated over OSNs for drawing comments of other users
over the events. Hashtags are the most prominent features for detecting events
via OSNs. Instead of monitoring frequency of all words belonging to the text
stream, the monitoring of the frequency of hashtags is more beneficial.
3. Signal Generation based on online analysis of raw features: There are certain
signal generators which take raw features of the text data streams and process
them online for generating the signals, for example, change in the sentiment score
and change in domain of discussion.

a. Sentiment-Analysis-Based Signal Generation: The sentiment analysis is

defined as the automatic process of determining the sentiment of digitally

[Link]@[Link]
202 R. Srivastava et al.

stored textual documents [19–21]. While performing online sentiment anal-

ysis [18, 20, 22, 23] for a specified event, a significant change in the sentiment
score with respect to time is a strong indicator of the occurrence of sub-event.
Such significant changes can be utilized for tracking the occurrence of sub-
events.
b. Domain-Specific Features: In text streams, the features represent the domain
changes very frequently; for instance, the discussion of users on OSNs may
change from the topic politics to sports after any crucial sport result. Observ-
ing the domain-specific features and generating the signal accordingly is
also very prominent way for accurate event detection. However, such kind of
signal generation required online domain classification. During the time of
change in the domain of discussion, the underlying data distribution of the
text stream also changes significantly.

3.2 Burst Detection in Text Data Stream

Burst detection is a process of detecting high-frequency period of in time series data

analysis. The burst detection methods are utilized in variety of ways:
1. Fixed threshold value.
2. Minimum period between two bursts.
3. Duration of bursts.
4. Adaptive threshold parameter.

4 Conclusion

With the exponential growth of users on OSNs, the OSNs now resemble the cor-
responding real-world community very cohesively. The events occurred in the real
world are often discussed on the virtual world of OSNs; hence, the analysis of con-
tents of online OSNs provides immense possibility to track the real-world happening
from the horizon of virtual world formed over social network. In keeping the view
of the feasibility of OSNs in real-time tracking of the new events through OSNs, this
paper discusses the general framework for real-world event detection through online
social networking sites.

References

1. Web 2.0: January 2016, cited 2016. Available from: [Link]

2. Kaplan, A.M., Haenlein, M.: Users of the world, unite! The challenges and opportunities of
social media. Bus. Horiz. 53(1), 59–68 (2010)

[Link]@[Link]
Framework for Real-World Event Detection … 203

3. Rosa, K.D., Ellen, J.: Text classification methodologies applied to micro-text in military chat.
In: ICMLA’09. International Conference on Machine Learning and Applications. IEEE (2009)
4. Jackoway, A., Samet, H., Sankaranarayanan, J.: Identification of live news events using Twitter.
In: Proceedings of the 3rd ACM SIGSPATIAL International Workshop on Location-Based
Social Networks. ACM (2011)
5. Srivastava, R., et al.: Analyzing Delhi assembly election 2015 using textual content of social
network. In: Proceedings of the Sixth International Conference on Computer and Communi-
cation Technology. ACM (2015)
6. Sakaki, T., Okazaki, M., Matsuo, Y.: Earthquake shakes Twitter users: real-time event detection
by social sensors. In: Proceedings of the 19th International Conference on World Wide Web.
ACM (2010)
7. Weimann, G.: New Terrorism and New Media. Wilson Center Common Labs (2014)
8. O’Connor, B., et al.: From tweets to polls: linking text sentiment to public opinion time series.
ICWSM 11(122–129), 1–2 (2010)
9. Yang, Y., Pierce, T., Carbonell, J.: A study of retrospective and on-line event detection. In:
Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Devel-
opment in Information Retrieval. ACM (1998)
10. Allan, J.: Topic Detection and Tracking: Event-Based Information Organization, vol. 12.
Springer Science & Business Media (2012)
11. Allan, J.: Introduction to topic detection and tracking. In: Topic Detection and Tracking,
pp. 1–16. Springer, Berlin (2002)
12. AlSumait, L., Barbará, D., Domeniconi, C.: On-line LDA: adaptive topic models for mining
text streams with applications to topic detection and tracking. In: Eighth IEEE International
Conference on Data Mining. IEEE (2008)
13. Fiscus, J.G., Doddington, G.R.: Topic detection and tracking evaluation overview. In: Topic
Detection and Tracking, pp 17–31. Springer, Berlin
14. Brants, T., Chen, F., Farahat, A.: A system for new event detection. In: Proceedings of the 26th
Annual International ACM SIGIR Conference on Research and Development in Information
Retrieval. ACM (2003)
15. Atefeh, F., Khreich, W.: A survey of techniques for event detection in twitter. Comput. Intell.
31(1), 132–164 (2015)
16. Cordeiro, M.: Twitter event detection: combining wavelet analysis and topic inference sum-
marization. In: Doctoral Symposium on Informatics Engineering (2012)
17. Bahir, E., Peled, A.: Real-time major events monitoring and alert system through social net-
works. J. Conting. Crisis Manag. 23(4), 210–220 (2015)
18. Cui, L., et al.: Topical event detection on Twitter. In: Australasian Database Conference.
Springer, Berlin (2016)
19. Srivastava, R., Bhatia, M.: Ensemble methods for sentiment analysis of on-line micro-texts.
In: International Conference on Recent Advances and Innovations in Engineering (ICRAIE).
IEEE (2016)
20. Srivastava, R., Bhatia, M.: Challenges with sentiment analysis of on-line micro-texts. Int. J.
Intell. Syst. Appl. 9(7), 31 (2017)
21. Srivastava, R., et al.: Exploiting grammatical dependencies for fine-grained opinion mining.
In International Conference on Computer and Communication Technology (ICCCT). IEEE
(2010)
22. Srivastava, R., Bhatia, M.: Offline versus online sentiment analysis: issues with sentiment
analysis of online micro-texts. Int. J. Inf. Retr. Res. (IJIRR) 7(4), 1–18 (2017)
23. Srivastava, R., Bhatia, M.: Real-time unspecified major sub-events detection in the twitter data
stream that cause the change in the sentiment score of the targeted event. Int. J. Inf. Technol.
Web Eng. (IJITWE) 12(4), 1–21 (2017)

[Link]@[Link]
Novel Architecture for Internet of Things
and Blockchain Technologies

Chetna Dabas and Aniket Dabas

Abstract Diverse companies and platforms are falling madly in love with the
blockchain technology in the present scenario. So, it is of due importance that the rela-
tions among the different types of blockchains are established. Further, the extraction
of capabilities of strengths of diverse blockchains catering to different kind of micro-
services is a crucial step toward the realistic blockchain. This issue even became of
utmost significance when specifically the researchers are dealing with Internet of
things (IoT) purposes. The IoT in general is composed of a wide spectrum of tech-
nologies varying from soft real time to hard real time, from stateless through state
full, and from unconstrained to constrained. Therefore, for all definite implementa-
tions, no definite blueprint of the IoT architecture may be referenced. When it comes
to blockchain technology melded with IoT, the existing literature seems to lack a
hybrid architecture which contains the best of both these worlds. This research paper
proposes a hybrid architecture for IoT and blockchain along with the evaluation of
this architecture by implementing an application for the same. The application is
for a common user using these technologies which enable the user to select among
diverse cryptocurrencies efficiently. The evaluation of the proposed architecture is
carried out by implementing blockchain on IoT devices. The results retrieved as a
part of this work are encouraging in terms of execution times.

Keywords Internet of things · Blockchain · Radio frequency identification

IoT architecture

C. Dabas (B)
Department of CSE, Jaypee Institute of Information Technology, Noida, India
e-mail: [Link]@[Link]
A. Dabas
Erricson India Global Services Pvt. Ltd., Noida, India
e-mail: [Link]@[Link]

© Springer Nature Singapore Pte Ltd. 2019 205

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
206 C. Dabas and A. Dabas

1 Introduction

In almost all the information industries across the globe, the concept of Internet of
things is prevalent in its wider most form. It has deeply impacted the life of a common
man and has caught the attention of academia, industries, and the governments world-
wide. An Internet of things (IoT) [1] is an advanced and extending technology having
its basis grounded on the well-known Internet technology. There lie immense possi-
bilities of expansion and scalability with IoT. The IoT includes any kind of objects or
things connected through wired or wireless technologies involving orientation trac-
ing, intelligent recognition, equipments like Radio Frequency Identification (RFID)
[2, 3] tags and readers, numerous kinds of sensors, GPS, laser scanners to name a
few. Further, the IoT can be well connected to cloud computing [4–9] technology to
make the matter more complicated to address [10].
There are various initial developments in the direction of the Internet of things
which have paid special attention to the product life cycle applications and RFID
infrastructures in Business-to-Business (B2B) logistics but still the IoT is farther to
go and make even a common man capable to effectively access as well as contribute
rich, secure, and cost-effective information about locations and things.
Due to the heterogeneous environment of the IoT scenarios, there is a lack of
precise and generic architecture which can govern the majority of applications in
diverse scenarios. Apart from this, there exist lots of constraints in such an environ-
ment which hampers the performance of applications and increases the cost involved
in developing methods to assist fast and reliable transactions across the network.
Further, there come security issues embedded with such heterogeneous and scalable
environments where IoT is connected to the cloud as well.
Blockchain [11–13] can perform a crucial role in such a scenario to suggest reliable
and secure solutions with the Internet of things [1].
Specifically, the blockchain is considered as a distributed and public ledger. The
transactions in blockchain are kept in a linked list consisting of blocks. This chain
expands upon the addition of new blocks which are added continuously on the Inter-
net. Various cryptography and distributed algorithms have been utilized for maintain-
ing the security and consistency of the ledger. Blockchain comes with advantages of
persistence, decentralization, auditability, and anonymity which enable the improve-
ment of the efficiency and saves time.
Specifically, mentioned in [13] paper, the blockchain holds great potential in
constructing the future Internet systems to come in due years.
In this work, a hybrid IoT architecture is proposed, the possible applications
of the proposed architecture are suggested, and the architectural elements of the
proposed IoT architecture are discussed. Further, the integrated view of the proposed
hybrid IoT architecture composed of technologies like RFID, cloud computing, and
blockchain is presented. Apart from this, IoT data analytics is performed along with
result assessment.
The rest of the research paper is ordered in the way mentioned ahead. The related
work is presented in Sect. 2. Section 3 presents the proposed Internet of things

[Link]@[Link]
Novel Architecture for Internet of Things … 207

hybrid architecture. This section also contains subsections which contain the possible
applications for the proposed architecture, the architectural elements of the proposed
IoT architecture, and the integrated view of the proposed architecture. In Sect. 4,
authors give a detailed explanation of our proposed work. Section 4 presents the
IoT data analytics along with result assessment. Section 5 presents open challenges
and future vision. The research paper concludes with Sect. 5 which presents the
conclusions of the work and summary is discussed as well.

2 Related Work

This related work section describes the existing literature on various existing architec-
tures of IoT-based systems along with the possibilities of the existing technologies
like cloud computing, Radio Frequency Identification, and blockchain along with
IoT.
In [10], authors have discussed architectural aspects of integrating cloud and RFID
for the scalability. The authors have also talked about the possibilities of kinds of
applications that can be running on this integrated architecture. The authors of this
paper have revealed the blending possibilities of the two important technologies,
namely cloud and RFID.
The authors of paper [14] have given a cloud-centric vision for IoT scenarios. This
paper also makes a point that with the present existence of a variety of numerous RFID
tags and other wireless sensor technologies and sensor equipments, IoT is believed
to be the forthcoming revolutionary technology which is supposed to transform the
present Internet to future Internet which is fully integrated.
In the paper [15], the authors present a basic architecture of IoT with three layers,
namely perception layer, network layer, and application layer. This paper lacks the
workflow of these layers in detail.
The authors of research work [16] have tabulated the security issues in IoT. Further,
they have presented the role of blockchain in solving a huge number of security
issues associated with IoT and hence revealing the significance of blockchain in IoT
scenarios.
In [17], the research work by the authors discussed the future architecture for
the IoT. This paper discusses the definitions and existing developments in the area,
along with key requirements. This paper also discusses a proposal for the future IoT.
Apart from this, in both user-centric and business-centric environments, this paper
examines the usability by stakeholders.
The authors of the paper [18] discuss the significance of IoT in day-to-day life,
present the architecture of IoT, the protocols for IoT, and further discuss possible
IoT applications.
The research work in [19] presents a framework for IoT architectures. The frame-
work proposed by the authors is designed for the applications which are service-based
and in smart cities. The data input for this framework is retrieved from data sources
which form data cloud.

[Link]@[Link]
208 C. Dabas and A. Dabas

In [20], the authors have proposed the classification along with a comparison of the
blockchains and the blockchain-based systems. This paper claims that the presented
work helps in assisting the assessment and design procedure of blockchains on the
existing architectures.
The authors of [21] highlight the existing trends and generic architecture of IoT
composed up of application layer, middleware layer, perception layer, business layer
along with the future application. The authors of this paper specifically stress on the
fact that IoT is gaining popularity across diverse fields like industry, academia to
name a few. Further to this, the authors convey that IoT bears the capability to bring
benefits to professionals and even to individuals.
In [22], the authors of this paper have proposed specific media-aware traffic secu-
rity architecture. This architecture depends on the given traffic classification for
catering various multimedia services.
In the existing literature, so far no work has been noted on the integration of all
these technologies like blockchain, RFID, cloud computing in a single architecture
melded with IoT.
The authors of this paper took this step further in the proposed research work as
follows.
The research aim of this proposed work is to offer fusion architectures with Internet
of things and blockchain technologies, while utilizing Radio Frequency Identification
and cloud computing perception altogether. The vision of this proposed research work
is to provide data security, management, analytics, and implications to a common
man while keeping low cost and optimum performance.

3 The Proposed Internet of Things Hybrid Architecture

System blockchains are the inventions of the present era that have revolutionized
the way one distributes personal information over the cloud and Internet in secured
manner. Blockchain has been of great importance to the daily personal (both textual
and image) data playing its crucial role on social Web sites, etc., and hence in the
commercial aspect. RFID has on the other hand revolutionized the association of
unique identities and tracking to things, places, animals, and even human beings.
The hybrid use of RFID and blockchain leverages a benefit of distributed ledger
which fortifies both security and transparency.
Aryanvika is a self-improved RFID and blockchain system and method which
has the ability to boost the performance and decline the cost of the data analyzed in
optimum ways in which data can be handled.
The idea was first conceived by the authors when they were busy observing their
surroundings to seek an innovative design architecture on how to use their creativity
for the enhancement of the society. As a consequence, the attention of the authors
fell on the day-to-day problem of storing the magnanimous amount of data generated
by photographs taken by people every day all over the world and not being able to

[Link]@[Link]
Novel Architecture for Internet of Things … 209

store them properly due to issues like lack of security. Then the authors thought why
not to think of some innovative way of solving this problem.
This idea is not sold; neither has it gone through any official process. No prototypes
have been made. Authors do have a hard copy of our calculations and workouts and
researches that made us finalize the features for the system and method.

3.1 Example Scenarios of Applications of the Proposed IOT

Architecture

In this section, the authors explain with the help of a few examples the possible
applications that may be used on the proposed architecture.
• For instance, suppose a person in India was working on a technology ‘X’ and had
his personal photograph album for it stored on his personal cloud at his home.
He also wants to associate a unique identification to his photographs for security
reasons. Suppose he now moves to a different country, let’s say Australia, for a
job and wants to show his secure photographs to his new colleagues. This can be
easily accomplished with the proposed hybrid IoT architecture which he might
have constructed in his home back in India.
• Another interesting example of this scenario is related to mining of user choice-
based cryptocurrencies pertaining to scalable, efficient, and cost-effective way in
a heterogeneous IoT environment. The proposed architecture delivers a significant
performance in a cost-effective manner in such a scenario.

3.2 Elements of the Proposed IoT Architecture

The proposed architecture in the present research work is described in Fig. 1 in

abstract as a system model. It makes use of blockchain, Radio Frequency Identifi-
cation, and cloud technologies. This system model is primarily composed of four
layers as is self-explanatory in Fig. 1. The authors have named it as Aryanvika.
The detailed workflows of the proposed hybrid IoT architecture are presented in
the subsections ahead.

3.2.1 Workflow of Layer 1

Layer 1 corresponds to the virtual private network on the cloud. It is composed of

three heterogeneous devices, namely EndUser Device 1 (laptop), EndUser Device 2
(smartphone), and EndUser Device 3 (raspberry pi 3). These heterogeneous devices
may be scaled up or down according to the requirements. Further, these devices have
their own respective user interfaces (may be same or different). With the help of these

[Link]@[Link]
210 C. Dabas and A. Dabas

Fig. 1 System model

user interfaces, the user can see or update the data on these devices. These end user
devices form a virtual private network. This layer 1 is further connected to layer 2 of
the proposed hybrid IoT architecture which is explained in the next subsection. The
system model of layer 1 is presented in Fig. 2.

3.2.2 Workflow of Layer 2

Layer 2 as depicted in Fig. 3 takes its input from layer 1. The interface of this layer
interacts with the layer 1 as well as the RFID reader. There exists a sound possibility
of using more than one RFID reader in this layer of the architecture. The RFID reader
offers a unique identification of the asset linked to another interface which is linked
to layer 3 of this architecture. The output of layer 2 is fed as an input to layer 3.

[Link]@[Link]
Novel Architecture for Internet of Things … 211

Fig. 2 System model: layer 1

Fig. 3 System model: layer 2

[Link]@[Link]
212 C. Dabas and A. Dabas

Fig. 4 System model: layer 3

3.2.3 Workflow of Layer 3

Layer 3 of the proposed novel architecture is shown in Fig. 4. This layer takes its
input from layer 2 and passes on its output to layer 4. In layer 3, there exists a single
node (to save cost) although there exists a strong possibility to scale it to multiple
nodes depending on the requirements. There exist three computational elements in
the present architecture (which may be extended to n in numbers). By this layer of
the architecture, a transaction is initiated and layer 4 is invoked.

3.2.4 Workflow of Layer 4

Layer 4 corresponding to the proposed hybrid IoT architecture is presented in Fig. 5.

This layer is composed of the blockchain. The data in the block of the chain may be
like the time stamp, date hash, and link as is clearly indicated in the figure. The data
block is added securely and uniquely with tracking details.

[Link]@[Link]
Novel Architecture for Internet of Things … 213

Fig. 5 System model: layer

4

[Link]@[Link]
214 C. Dabas and A. Dabas

Fig. 6 Proposed hybrid Internet of things architecture (with blockchain and RFID technologies)

3.3 Integrated View of Hybrid Internet of Things

Architecture

Here, the authors describe the integrated functioning of the proposed hybrid architec-
ture of the Internet of things. The pictorial view of the proposed novel IoT architecture
is presented in Fig. 6.
The functioning of the proposed IoT architecture is described by the following
four major components or layers:
(1) Virtual Private Network (VPN): Accessing Data
Whenever a request from the user comes, a data file is generated locally on the
virtual private network (local cloud) from one of the heterogeneous devices, for
example, from EndUser Device 1 (laptop) or EndUser Device 2 (smartphone) or
EndUser Device 3 (raspberry pi 3). These devices may be further generalized to
n numbers. The data file may be in a different format depending on the source.
There exists a Python script to convert the requested file in the desired format.
(2) RFID System: Reading Data
The RFID reader is connected to the VPN and reads data from the same. The
Python script is used to connect the RFID reader to the VPN and reading the data
and associating a unique identification to the data. This ensures the uniqueness
of the desired data. There can be a single RFID system or a collection of RFID
devices. In case of more than one RFID reader, there will be a necessity of
incorporation of the RFID anti-collision protocols [23].
(3) Computing system: Processing Data
This data from the RFID reader is further fed to a cluster of small fast processing
computers at a single node (there is a possibility of multiple nodes connected at

[Link]@[Link]
Novel Architecture for Internet of Things … 215

this point consisting of n number of computational elements depending on the

horizontal scalability parameters) which contains execution of Python scripts.
(4) Blockchain: Managing Data
The concepts of proofs are written according to the requirements, the specific
or generic algorithms are up and running on the computational nodes, and the
data is appended as a block in the chain in the most reliable and cost-effective
manner.

4 Proposed Architecture Assessment and Results

This section discusses the experimental setup and assumptions made while execut-
ing the designed architecture, and then, results are displayed corresponding to the
successful execution of the proposed hybrid IoT architecture. The application under
consideration is written in Python and aims to offer a user the best-rated cryptocur-
rency among the four considered in the present scenario. The four cryptocurrencies
considered in this work are BTC, Etherum, Ripple, and Zcash.

4.1 Environment of Execution

The environment under consideration is as discussed in the proposed hybrid archi-

tecture for IoT under Sect. 3 of this research paper.

4.2 The Scenario and Application Under Consideration

A user of the proposed architecture has migrated to some other location, and his IoT
architecture setup exists at his home in India. The user who is temporarily migrated
to some other location now wishes to offer a new business proposal to his new
organization where he wishes to seek the old mined data kept on his cloud at his
home. He wishes to offer an application which makes use of all the technologies
like RFID, cloud, blockchain, and IoT to make secure transaction and offer his new
organization a secure, unique, and reliable solution.

4.3 Execution of the Application

The data in layer of the architecture was successfully mined with the help of designed
Python scripts. Each of the data center storage had one of the four different types
of files, namely low priority type-1, high priority type-1, low priority type-2, and

[Link]@[Link]
216 C. Dabas and A. Dabas

Fig. 7 Timing interval versus price of various cryptocurrencies

high priority type-2. For simulating the configuration, the authors used three end
user devices, a cluster of three graphical processing units, one RC522 RFID reader,
specially designed user interfaces, one raspberry pi 3 module with Wi-Fi module,
one Samsung smartphone, and one HP laptop (core i3). The results were successfully
obtained at the user interface of a different node as depicted in Fig. 7.

5 Open Challenges and Future Visions

The challenges associated with the proposed work are to bring the instances of the
proposed architectures to life for amplification of the performance of the hybrid IoT
system and also to reduce the cost of data handled to a prime level while keeping into
consideration the heterogeneous devices utilized in Internet of things. More chal-
lenges associated with the proposed research work are multi-technologies awareness
and integration capabilities along with the blockchain vulnerabilities. The future
vision for the proposed work imbibes the possibilities of implementation of the pro-
posed architectures with polyglot programming and polyglot approaches.

[Link]@[Link]
Novel Architecture for Internet of Things … 217

6 Summary and Conclusions

The prime idea of this research work is that the proposed IoT architecture would
require no external characteristics to govern the cost or the performance of the hybrid
system. Further, the instances resulting from implications of the proposed architec-
ture would possess the self-capability for analyzing the performance of the private
data which is in the system and hence set the cost, timing to list a few among cru-
cial parameters. Further, there lie intensive market opportunities for the proposed
research work. This is due to the fact that one can make discrete Aryanvika kind
of system and methods for diverse stretches across the globe, catering to their own
needs of preserving data from heterogeneous sources in a secure and cost-effective
manner. Aryanvika is a huge research, business, and expansion opportunity.

References

1. Huckle, S., Bhattacharya, R., White, M., Beloff, N.: Internet of things, blockchain and shared
economy applications. Procedia Comput. Sci. 98, 461–466 (2016)
2. Molnar, D., Wagner, D.: Privacy and security in library RFID: issues, practices, and archi-
tectures. In: Proceedings of the 11th ACM Conference on Computer and Communications
Security, pp. 210–219. ACM, Oct 2004
3. Bolic, M., Simplot-Ryl, D., Stojmenovic, I. (eds.): RFID systems: research trends and chal-
lenges. Wiley, New York (2010)
4. Buyya, R., Yeo, C.S., Venugopal, S., Broberg, J., Brandic, I.: Cloud computing and emerging
IT platforms: vision, hype, and reality for delivering computing as the 5th utility. Future Gener.
Comput. Syst. 25(6), 599–616 (2009)
5. Moreno-Vozmediano, R., Montero, R.S., Llorente, I.M.: IaaS cloud architecture: from virtual-
ized datacenters to federated cloud infrastructures. Computer 45(12), 65–72 (2012)
6. Zhu, J.: Cloud computing technologies and applications. In: Handbook of Cloud Computing,
pp. 21–45. Springer, Boston (2010)
7. Al-Dhuraibi, Y., Paraiso, F., Djarallah, N., Merle, P.: Elasticity in cloud computing: state of the
art and research challenges. IEEE Trans. Serv. Comput. 11(2), 430–447 (2018)
8. Malik, S.U.R., Khan, S.U., Ewen, S.J., Tziritas, N., Kolodziej, J., Zomaya, A.Y., Malluhi,
Q.M.: Performance analysis of data intensive cloud systems based on data management and
replication: a survey. Distrib. Parallel Databases 34(2), 179–215 (2016)
9. Kapoor, S., Dabas, C.: Cluster based load balancing in cloud computing. In: Eighth International
Conference on Contemporary Computing (IC3), pp. 76–81. IEEE, Aug 2015
10. Dabas, C., Gupta, J.P.: A cloud computing architecture framework for scalable RFID. In:
Proceedings of the International MultiConference of Engineers and Computer Scientists, vol.
1, Mar 2010
11. Zheng, Z., Xie, S., Dai, H.N., Wang, H.: Blockchain challenges and opportunities: a survey.
Work Paper (2016)
12. Cachin, C.: Architecture of the hyperledger blockchain fabric. In: Workshop on Distributed
Cryptocurrencies and Consensus Ledgers, vol. 310, July 2016
13. Zheng, Z., Xie, S., Dai, H., Chen, X., Wang, H.: An overview of blockchain technology: archi-
tecture, consensus, and future trends. In: IEEE International Congress on Big Data (BigData
Congress), pp. 557–564. IEEE, June 2017
14. Gubbi, J., Buyya, R., Marusic, S., Palaniswami, M.: Internet of things (IoT): a vision, architec-
tural elements, and future directions. Future Gener. Comput. Syst. 29(7), 1645–1660 (2013)

[Link]@[Link]
218 C. Dabas and A. Dabas

15. Yang, Z., Yue, Y., Yang, Y., Peng, Y., Wang, X., Liu, W.: Study and application on the archi-
tecture and key technologies for IOT. In: International Conference on Multimedia Technology
(ICMT), pp. 747–751. IEEE, July 2011
16. Khan, M.A., Salah, K.: IoT security: review, blockchain solutions, and open challenges. Future
Gener. Comput. Syst. 82, 395–411 (2018)
17. Uckelmann, D., Harrison, M., Michahelles, F.: An architectural approach towards the future
internet of things. In: Architecting the Internet of Things, pp. 1–24. Springer, Berlin (2011)
18. Kraijak, S., Tuwanut, P.: A survey on IoT architectures, protocols, applications, security, pri-
vacy, real-world implementation and future trends (2015)
19. Al-Fagih, A.E., Al-Turjman, F.M., Alsalih, W.M., Hassanein, H.S.: A priced public sens-
ing framework for heterogeneous IoT architectures. IEEE Trans. Emerg. Top. Comput. 1(1),
133–147 (2013)
20. Xu, X., Weber, I., Staples, M., Zhu, L., Bosch, J., Bass, L., Pautasso, C., Rimba, P.: A taxonomy
of blockchain-based systems for architecture design. In: IEEE International Conference on
Software Architecture (ICSA), pp. 243–252. IEEE, Apr 2017
21. Khan, R., Khan, S.U., Zaheer, R., & Khan, S.: Future internet: the internet of things architecture,
possible applications and key challenges. In: 10th International Conference on Frontiers of
Information Technology (FIT), pp. 257–260. IEEE, Dec, 2012
22. Zhou, L., Chao, H.C. Multimedia traffic security architecture for the internet of things. IEEE
Netw. 25(3) (2011)
23. Dabas, A.C., Balhara, B.M., Gupta, C.J.P.: CDMA based anti-collision deterministic algorithm
for RFID tags. Int. J. Recent Trends Eng. 1(1), 603 (2009)

[Link]@[Link]
Homomorphism Between Fuzzy
Set-Valued Information Systems

Waseem Ahmed, M. M. Sufyan Beg and Tanvir Ahmad

Abstract Communication between various information systems (IS) has been an

urgent issue that needs to be discussed in the granular world. Compatible and consis-
tent homomorphism is an important analytical tool to study communication among
various IS. Fuzzy set-valued information systems (FSIS) are those IS that contain
fuzzy set-values for some attribute. This paper aims to discuss important properties
related to communication between FSIS. Fuzzy relation mapping from the perspec-
tive of FSIS is discussed. The proposed approach proved that feature selection and
other properties of the original FSIS and the corresponding image FSIS are assured
under consistent and compatible homomorphism. Finally, a real-life example demon-
strated the utilization of the proposed work.

Keywords Fuzzy set-valued information systems · Homomorphism

Feature selection · Fuzzy relation mapping

1 Introduction

Rough set theory (RST) is an influential soft computing tool in the area of infor-
mation technology and has attracted much attention and interest [1–5]. Continuous
or real-valued IS cannot be handled directly by the traditional rough set because it
requires discretization of real attributes before feature selection, so to handle this,
fuzzy rough set (FRS) is introduced [2]. In FRS, fuzzy similarity relation is used to
handle real-valued attributes. RST begins with a single-valued information system

W. Ahmed (B) · T. Ahmad

Department of Computer Engineering, Jamia Millia Islamia, New Delhi, India
e-mail: waseem.ahmed86@[Link]
T. Ahmad
e-mail: tahmad2@[Link]
M. M. S. Beg
Department of Computer Engineering, Aligarh Muslim University, Aligarh, India
e-mail: mmsbeg@[Link]

© Springer Nature Singapore Pte Ltd. 2019 219

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
220 W. Ahmed et al.

[2–5]; however, in real-life situations, some attribute values may contain set-values,
resulting in the formation of SIS [6, 7]. Dai and Tian [6] investigated FRS model
for SIS. SIS studied so far contains crisp set-values [6, 7] and does not consider the
case of fuzzy set-valued attributes. To handle these types of attributes, a FSIS was
introduced [1].
In today’s Information era, information system (IS) is the realm of information
technology, and the major concern here is communication between IS. Due to differ-
ent nature of IS, it becomes necessary to transfer information within these IS [3, 5].
This motivates us to examine the communication among various IS’s. In this work,
important properties related to communication between FSIS using FRS model are
discussed. Communication means transformation of information between IS, along-
side keeping unblemished its prime properties and functions.
From mathematical viewpoint, mapping is an efficient way to communicate
between IS by analyzing their properties [5]. Homomorphism mapping plays a vital
role to accomplish this task [3, 5]. In this paper, important properties related to
communication between FSIS under FRS model are discussed.
The proposed approach proved that feature selection and other properties of the
original FSIS and the corresponding image FSIS are assured under consistent and
compatible homomorphism.
The structure of the remaining paper is defined below:
A brief description about FSIS is given in Sect. 2. In Sect. 3, fuzzy relation
mapping for FSIS is formalized. Section 4 discusses the notion of homomorphism
between FSIS. In Sect. 5, the conclusion is outlined.

2 Preliminaries

2.1 Introduction to FSIS

Definition 1 A FSIS  (S, C, V , f ) is an IS that contains real or fuzzy set-values for

some of the attribute [1].

Where S and C are set of sample and set of fuzzy multivalued attributes, respec-
tively. f : S × C → V , where for m ∈ S, c ∈ C, f (m, c) ∈ V c assigns fuzzy set-values
to samples. Also, V c is the set of pairs (x, µVc (x)) such that µVc (x) is the membership
value of x in V c between [0, 1].

Example 2 FSIS is given in Table 1. Here, S  {m1 , m2 , m3 , m4 } is a set of objects; C

 {R, W , S} is a set of attributes; V = {French, German, English} is domain values.
For ease, French, German, and English are denoted by f , g, and e, respectively.

If suppose, ‘c’ is an attribute ‘Reading’= {f , g, e}. Then c(m) = {(f , 0.8), (g, 0.6),
(e, 0.9)} illustrates that reading abilities of ‘m’ in French, German, and English are
0.8, 0.6, and 0.9, respectively.

[Link]@[Link]
Homomorphism Between Fuzzy Set-Valued Information Systems 221

Table 1 An example of FSIS

FSIS Reading (R) Writing (W ) Speaking (S)
m1 {(f , 0.8), (g, 0.6), (e, {(f , 0.3), (g, 0.7), (e, {(f , 0.6), (g, 0.9), (e,
0.9)} 0.5)} 0.7)}
m2 {(f , 0.8), (g, 0.6), (e, {(f , 0.3), (g, 0.7), (e, {(f , 0.6), (g, 0.9), (e,
0.9)} 0.5)} 0.7)}
m3 {(f , 1), (g, 0.3), (e, 0)} {(f , 0.8), (g, 0.3), (e, {(f , 0.9), (g, 0.6), (e,
0)} 0.3)}
m4 {(f , 0.5), (g, 0.9), (e, {(f , 0.7), (g, 0.7), (e, {(f , 0.5), (g, 0.8), (e,
0.4)} 0)} 0.7)}

2.2 Fuzzy Similarity Relation for FSIS

Definition 3 For FSIS  (S, C) and a ∈ C and m, n ∈ S, the fuzzy similarity relation
Ra is defined as:

inf(a(m), a(n))
µ Ra (m, n)  
sup(a(m), a(n))

3 Fuzzy Relation Mapping for FSIS

A definition of fuzzy relation mapping to communicate between FSIS using Zadeh’s

extension principle is given below:
For two universal sets S 1 and S 2 , let R(S 1 × S 1 ) and R(S 2 × S 2 ) represent classes
of all fuzzy binary relation on S 1 and S 2, respectively.

Definition 4 Let f : S 1 → S 2 be a mapping. ‘f ’ generate a mapping from R(S 1 × S 1 )

to R(S2 × S 2 ) as:

supm 1 ∈ f −1 (z1 ) supm 2 ∈ f −1 (z2 ) R(m 1 , m 2 ), (z 1 , z 2 ) ∈ f (S1 ) × f (S1 )
f (R)(z 1 , z 2 ) 
0, (z 1 , z 2 ) ∈
/ f (S1 ) × f (S1 )

3.1 Consistent Functions

Definition 5 Let S 1 and S 2 are universe of discourse, f : S 1 → S 2 is their mapping

from S 1 to S 2 , and B, B1 , B2 ∈ R(S 1 × S 1 ). Let [a]f  {b ∈ U: f (a)  f (b)};

For any m1 , m2 ∈ [a]f, and z1 , z2 ∈ [b]f , we say ‘f ’ is compatible with relation B,

if B(m1 , z1 )  B(m2 , z2 ).
For any a, b ∈ S 1 , if any of the condition holds:

[Link]@[Link]
222 W. Ahmed et al.

(1) B1 (m1 , z1 ) ≤ B2 (m1 , z1 ) for any (m1 , z1 ) ∈ [a]f × [b]f

(2) B1 (m1 , z1 ) ≥ B2 (m1 , z1 ) for any (m1 , z1 ) ∈ [a]f × [b]f , then the mapping ‘f ’ is
consistent to relation B1 and B2 .

Theorem 6 Let f: S1 → S2 , B, B1 , B2 ∈ R(S1 × S1 ); then, we have the following:

(1) f(B1 ∪ B2 )  f(B1 ) ∪ f(B2 )
(2) f(B1 ∩ B2 ) ⊆ f(B1 ) ∩ f(B2 ); if ‘f’ is a consistent mapping, then they are equal.

Proof (1) For any z1 , z2 ∈ S 2

f (B1 ∪ B2 ) (z 1 , z 2 )  sup sup (B1 ∪ B2 )(m 1 , m 2 )
m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 sup sup (B1 (m 1 , m 2 ) ∪ B2 (m 1 , m 2 ))

m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 ( f (B1 ) ∪ f (B2 ))(z 1 , z 2 ).

f (B1 ∩ B2 )(z 1 , z 2 )  sup sup (B1 ∩ B2 )(m 1 , m 2 )
m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 sup sup (B1 (m 1 , m 2 ) ∩ B2 (m 1 , m 2 ))

(2) m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

≤ sup sup B1 (m 1 , m 2 ) ∩ sup sup B2 (m 1 , m 2 )

m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 ) m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 ( f (B1 ) ∩ f (B2 ))(z 1 , z 2 ).

Now, it will be proved that the equality holds, if the mapping ‘f ’ is consistent
mapping:
Now, since ‘f ’ is consistent mapping to B1 and B2 , it follows from Definition 5
that it satisfies one of the following conditions:
1. B1 (m1 , m2 ) ≤ B2 (m1 , m2 )
2. B1 (m1 , m2 ) ≥ B2 (m1 , m2 ).
For any (m1 , m2 ) ∈ f −1 (z1 ) × f −1 (z2 )
For case 1,

f (B1 ∩ B2 ) (z 1 , z 2 )  sup sup (B1 ∩ B2 )(m 1 , m 2 )

m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 sup sup (B1 (m 1 , m 2 ) ∩ B2 (m 1 , m 2 ))

m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 sup sup B1 (m 1 , m 2 )
m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 f (B1 ) (z 1 , z 2 )

Now taking RHS,

( f (B1 ) ∩ f (B2 )(z 1 , z 2 )  f (B1 )(z 1 , z 2 ) ∩ f (B2 )(z 1 , z 2 )

( sup sup (B1 (m 1 , m 2 ))
m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

[Link]@[Link]
Homomorphism Between Fuzzy Set-Valued Information Systems 223

∩( sup sup B2 (m 1 , m 2 ))
m 1 ∈ f −1 (z 1 ) m 2 ∈ f −1 (z 2 )

 f (B1 ) (z 1 , z 2 )

Hence f (B1 ∩ B2 )  f (B1 ) ∩ f (B2 ).

Example 7 (continued to example 2) Let S 1  {m1 , m2 , m3 , m4 } and S 2  {z1 , z2 ,

z3 } f 2 : S 1 → S 2 is a mapping from S 1 to S 2 and defined as: f 2 (m1 )  f 2 (m2 )  z1 ;
f 2 (m3 )  z2 ; f 2 (m4 )  z3 ;

Fuzzy similarity relations for attributes reading (R), writing (W ), and speaking (S)
given in Table 1 are calculated using Definition 3 and shown in Table 2. (R R ∩RW ∩R S )
is given in Table 3.
Images of R, W, and S using ‘f 2 ’ are computed and presented in Table 4. f 2 (R R ) ∩
f 2 (RW ) ∩ f 2 (R S ) and f 2 (R R ∩ RW ∩ R S ) are given in Table 5.
We find that f 2 (R R ∩ RW ∩ R S )  f 2 (R R ) ∩ f (RW ) ∩ f (R S ), since mapping f 2
is a compatible mapping.

4 Homomorphism Between FSIS

This section focuses on the concept of homomorphism and discusses important prop-
erties of FSIS using homomorphism.

Definition 9 Let S 1 and S 2 are two universes, f: S 1 → S 2 be its mapping from S 1 to

S 2 . Assume R  {B1 , B2 , …, Bn } is collection of fuzzy relations on S 1 ; then f (R) 
{f (B1 ), f (B2 ), …, f (Bn )}. Then (S 1 , R) is termed as FSRIS, and the corresponding
(V 1 , f (R)) is its induced FSRIS [5].

Definition 9 Let (S 1 , R) be a FSRISs and ‘f ’ is a function mapping. We define

homomorphism using ‘f ’ satisfying certain conditions as follows [5]:

(1) ∀ Bi , Bj ∈ R, if ‘f ’ is consistent with every Bi and Bj, then we say that ‘f ’ is

consistent homomorphism.
(2) ∀ Bi ∈ R, if is compatible with each Bi, then we say that ‘f ’ is compatible
homomorphism.

Definition 10 Let (S 1 , R) be FSRISs and P ⊆ R satisfies the following:

(1) ∩P  ∩R
(2) ∀Bi ∈ P, ∩P ⊂ ∩(P − {Bi }).

Then we say P is a reduct of R

Theorem 11 Let (S1 , R) be FSRISs and ‘f’ be a consistent homomorphism mapping

from S1 to S2 . P ⊆ R is a reduct of R only when f(P) is reduct of f(R) and vice versa.

[Link]@[Link]
 224

Table 2 Similarity relations for attribute R, W, and S

RR m1 m2 m3 m4 RW m1 m2 m3 m4 RS m1 m2 m3 m4
m1 1 1 0.4 0.6 m1 1 1 0.3 0.5 m1 1 1 0.6 0.9
m2 1 1 0.4 0.6 m2 1 1 0.3 0.5 m2 1 1 0.6 0.9
m3 0.4 0.4 1 0.3 m3 0.3 0.3 1 0.7 m3 0.6 0.6 1 0.6
m4 0.6 0.6 0.3 1 m4 0.5 0.5 0.7 1 m4 0.9 0.9 0.6 1

[Link]@[Link]
W. Ahmed et al.
Homomorphism Between Fuzzy Set-Valued Information Systems 225

Table 3 Relation for (RR ∩ RW ∩ RS )

(RR ∩ RW ∩ RS ) m1 m2 m3 m4
m1 1 1 0.3 0.5
m2 1 1 0.3 0.5
m3 0.3 0.3 1 0.3
m4 0.5 0.5 0.3 1

Table 4 Images of R, W, and S using f 2

f 2 (RR ) z1 z2 z3 f 2 (RW ) z1 z2 z3 f 2 (RS ) z1 z2 z3
z1 1 0.4 0.6 z1 1 0.3 0.5 z1 1 0.6 0.9
z2 0.4 1 0.3 z2 0.3 1 0.7 z2 0.6 1 0.6
z3 0.6 0.3 1 z3 0.5 0.7 1 z3 0.9 0.6 1

Table 5 Relation for f 2 (RR ) ∩ f (RW ) ∩ f (RS ) and f 2 (RR ∩ RW ∩ RS )

f 2 (RR z1 z2 z3 f 2 (RR ) ∩ f 2 z1 z2 z3
∩ RW (RW ) ∩ f 2
∩ RS ) (RS )
z1 1 0.3 0.5 z1 1 0.3 0.5
z2 0.3 1 0.3 z2 0.3 1 0.3
z3 0.5 0.3 1 z3 0.5 0.3 1

Proof ⇒ Let us suppose, P be reduct of R. Therefore, ∩P  ∩(P − {Bi }).

Then, there must be m1 , m2 ∈ S 1 , such that ∩P(m1 , m2 ) < ∩ (P − {Bi })(m1 , m2 ),
which implies,

f (∩(P − {Bi }))( f (m 1 ), f (m 2 ))

 sup sup ∩(P − {Bi }) (z 1 , z 2 )
z 1 ∈ f −1 f (m 1 )) z 2 ∈ f −1 f (m 2 ))

> sup sup ∩P (z 1 , z 2 )

z 1 ∈ f −1 f (m 1 )) z 2 ∈ f −1 f (m 2 ))

 f (∩P) ( f (m 1 ), f (m 2 ))  f (∩R)( f (m 1 ), f (m 2 ))

Now, ∩P  ∩R. Hence, f (∩P)  f (∩R)

Using Theorem 6, ∩f (P)  ∩f (R).
Assume, f (P) is not reduct of f (R), ∃ Bi ∈ P such that ∩(f (P) − f (Bi ))  ∩f (P).
Since, f (P) − f (Ri )  f (P − {Bi }); therefore, ∩f (P − {Bi })  ∩f (P)  ∩f (R)
Again, by Theorem 6, f (∩(P − Bi ))  f (∩R)
which is a conflict to the assumption that f (P) is not reduct of f (R).
∴ f (P) is not reduct of f (R).
Example 12 (continued to example 7) Let S 1  {m1 , m2 , m3 , m4 } and S 2  {z1 , z2 ,
z3 }

[Link]@[Link]
226 W. Ahmed et al.

It is evident from Tables 2 and 4 that {RR , RW } is reduct of ‘R’, if and only if
{f 2 (RR ), f 2 (RW )} is a reduct of f 2 (R).

5 Conclusion

This paper aims to discuss important properties related to communication between

FSIS using FRS model. The definition of fuzzy relation mapping from the perspective
of FSIS is discussed. The proposed approach proved that feature selection and other
properties of the original FSIS and the corresponding image FSIS are assured in the
case of consistent and compatible homomorphism.

References

1. Ahmed, W., Beg, M.S., & Ahmad, T.: Entropy based feature selection for fuzzy set-valued
information systems. 3D Res. 9(2), 19 (2018)
2. Dubois, D., Prade, H.: Rough fuzzy sets and fuzzy rough sets. Int. J. Gen. Syst. 17, 191–208
(1990)
3. Grzymala-Busse, J.W., Sedelow, W.A.: On rough sets, and information system homomorphism.
Bull. Pol. Acad. Sci. Tech. Sci. 36(3–4), 233–239 (1988)
4. Pawlak, Z.: Rough Sets. Int. J. Comput. Inf. Sci. 11, 341–356 (1982)
5. Tsang, E.C.C., Wang, C., Chen, D., Wu, C., Hu, Q.: Communication between information
systems using fuzzy rough sets. IEEE Trans Fuzzy Syst. 21(3), 527–540 (2013)
6. Dai, J., Tian, H.: Fuzzy rough set model for set-valued data. Fuzzy Sets Syst. 229, 54–68 (2013)
7. Guan, Y.Y., Wang, H.K.: Set-valued information systems. Inf. Sci. 176(17), 2507–2525 (2006)

[Link]@[Link]
A Novel Approach for Predicting
the Outcome of Request in RAOP Dataset

Amreen Ahmad, Tanvir Ahmad and Abhishek Bhatt

Abstract In today’s era, online social communities such as Q&A sites are widely
used for asking favors, so it would be beneficial to formulate a technique that would
help in predicting the success of the response. The objective of the paper is to enhance
the accuracy of prediction of the success of altruistic request that follows the same
approach as used by ADJ (Proceedings of AAAI International Conference on Web
and Social Media, ICWSM, 2014 [1]. Three more features are proposed, i.e., topic,
role, and centrality in addition to the features proposed by ADJ [1]) to capture user’s
interaction in the past and topic effect on the prediction of response. We also propose
a graph-based success prediction (GSP) model that uses feature weights and uses the
underlying graph structure for the propagation to predict the outcome of a request.
Experiments were conducted on the RAOP dataset which belongs to sub-community
of [Link] using GSP, and it outperformed ADJ and other baseline methods using
limited training data.

Keywords Success prediction · Altruistic request · Social interactions

1 Introduction

In today’s era of the World Wide Web, online social communities are being widely
used for the noble cause. These social communities are often used by needy people for
donations or help. Recently, researchers have focussed on studying the underlying
hidden factors that increase the probability of a request being responded [2] and
help in the promotion of such request. Some key factors were identified by existing

A. Ahmad (B) · T. Ahmad · A. Bhatt

Department of Computer Engineering, Jamia Millia Islamia, New Delhi, India
e-mail: amreen.ahmad10@[Link]
T. Ahmad
e-mail: tahmad2@[Link]
A. Bhatt
e-mail: abhishek1995bhatt@[Link]
© Springer Nature Singapore Pte Ltd. 2019 227
L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
228 A. Ahmad et al.

researchers that included social interaction between the giver and user, the scale of
the request, and whether the giver reciprocates. ADJ [1] work focussed on uncovering
the effect of the linguistic factor on the success prediction of request given a situation
in which the giver is not receiving any reward. It was also proved that the linguistic
factor that included politeness, evidentiality, and narrative structure had a great effect
on decision making and is strongly correlated with the success of requests.
The paper aims to improve the prediction accuracy of the altruistic request, and
it follows the same approach as proposed by ADJ [1]. ADJ work suffered from the
following limitations: (i) It required a large dataset for prediction; (ii) it did not take
into consideration user interaction in the history that could be one of the important
reason for the success of the request. The central assumption of the paper is that
inclusion of three extra features, i.e., topic, centrality, and role, along with temporal,
social, and textual features would prove to be effective in boosting the accuracy of
the success prediction of the request. A graph-based success prediction (GSP) model
is proposed to predict the outcome of the request. Finally, for evaluation of GSP
model, ADJ, LR, and SVM are used.
The remaining part is structured as follows: The dataset and the proposed fea-
tures are explained in detail in Sect. 2. Next, we proceed with the explanation of the
proposed graph-based success prediction model in Sect. 3. We proceed with an anal-
ysis of results obtained from evaluation described in Sect. 4 followed by conclusion
discussed in Sect. 5.

2 Dataset and Features

A brief description of the dataset and features selection is given below:

2.1 Request Dataset

The pizza request dataset consists of 5728 pizza requests and 1.87 million relevant
posts that are used for computing feature values.

2.2 Features

A detailed description of the three proposed features, i.e., topic, role, and centrality,
is discussed.
1. Centrality features: We capture user interaction that takes place in [Link]
through comments so that their effect can be analyzed which would affect the
success of a request.

[Link]@[Link]
A Novel Approach for Predicting the Outcome of Request … 229

2. Role features: The role of a user in communities can be measured in the following
ways:
(a) Core node: Individuals who are highly active in a network community imply
that they have high TPR, i.e., triangle participation ratio.
(b) Bridging effect: Some individuals involved in the network community play
the role of transmitting information. RawComm is used to measure the
bridging effect [4] of a node.
3. Topic features: It is assumed that the success of altruistic request is highly depen-
dent on the topic to which it belongs. For modeling latent topics, the following
ways are used:
(a) Bag-of-Word: The words with AD (adverb), VA (adjective), VV (verb), NR
(proper noun), and NN (noun) are considered as BOW features.
(b) N-gram: Recursive feature elimination technique [5] is used for the identi-
fication of discriminative and important features of bi-gram and tri-gram.

3 Graph-Based Success Prediction (GSP) Model

Graph-based success prediction (GSP) model contains two phases:

3.1 Graph Construction

In the underlying graph, request text is represented as nodes and feature correlation
between nodes is computed using probabilistic approach that is used to predict the
outcome of unseen request. The graph is used to map the interaction between request
nodes denoted by G =(V ;E). Let V  U train ∪ U test , where U train is a set of training
nodes (t r ) and U test is a set of testing nodes (t st ). Let node u represents testing request
node, i.e., u ∈ U test , and v represent training request node, i.e., v ∈ U train . Now there
are two steps: (a) Connect every node u ∈ U test to the top-λ1 similar node v ∈ U train
and (b) connect every u ∈ U test to the top-λ2 similar nodes uw such that u  uw
and uw ∈ U test , where λ1  k × |U train | and λ2  k × |U test |, k is a parameter ∈[0,1].
In the proposed approach, the value of k is varied to determine its efficiency and
effectiveness. The probability of each request node for each label sc(x) (for node
x) can either be 1 (successful) or 0 (unsuccessful) denoted by Psc (x). Firstly, the
probabilistic scores for trained request nodes are set. For v ∈ U train , let P1 (v)  1 and
P0 (v)  0 if sc(v)  1 else P1 (v)  0 and P0 (v)  1. Also for u ∈ U test , let P1 (u)  0
and P0 (u)  0.
The edge of graph G denotes the correlation between request texts based on their
features. The process of edge creation is explained further. Let us assume that there is
a certain feature f t where f t represents one of the features of nodes, then feature-based

[Link]@[Link]
230 A. Ahmad et al.

request correlation factor F c (n1 ; n2 ) is computed for nodes (n1; n2) ∈ E that can be
derived from difference in features Fc (n 1 ; n 2 )  ft(n 1 ) − ft(n 2 ). For a given set of
features denoted by F  f t (where t  1; …; z) and z is the count of features, feature-
aware correlation factor can be computed using feature-based request correlation
factor and is given as follows:
 
   z
Far c n 1, n 2  exp − wt2 × Fc (n 1 , n 2 ) (1)
t1

where wt denotes the weight of respective feature f t . The F a rc(n1 ; n2 ) is edge weight
which can be denoted as wc (n1 ; n2 ) for edge (n1 ; n2 ) ← ∈ E in the graph G.

3.2 Learning-Based Propagational Optimization

It works in two steps:

1. The probabilistic values of P1 (u) and P0 (u) are inferred from P1 (v) and P0 (v).
2. The probabilistic values of P1 (u) and P0 (u) can also be derived from u’s neigh-
boring nodes. Let the degree of node u be denoted as du, then P1 (u) and P0 (u)
can be derived using the given approach that enables us to get optimal set of edge
weights in the graph:

1 
Psc (u)  wc (u, y) × Psc (y) (2)
du (u,y)∈E

where y denotes other nodes(both training and testing dataset) present in the graph
G.
Furthermore, learning and optimizing wt can enable us to compute similar prob-
abilities for nodes in the close vicinity of the testing node. In this model, a heuristic
approach is proposed to learn the feature weights wt based on entropy of the proba-
bilities and the probabilities are fine-tuned toward more accurate results. E p denotes
the probability of success that is given as:
1 
E p (u)  ×− Psc (u)(1 − Psc (u)) log(1 − Psc (u)) (3)
Utest sc∈(0,1)

where U test denotes the count of requests in testing dataset of graph G and sc denotes
success label.

Learning Optimization It is assumed that success probabilities computed would

be of higher confidence if they are achieving lower entropy values. Gradient descent
algorithm is used to minimize the entropy w.r.t wt given by:

[Link]@[Link]
A Novel Approach for Predicting the Outcome of Request … 231

1  (log(1 − P(y))
δ E p (u)  × × δ P(y) (4)
Utest y∈(U ) P(y)
test

It is further differentiated using chain rule, and the output is:

δwt  2 × wc (n 1 × n 2 ) × Fc (n 1 × n 2 ) × wt (5)

Prediction Likewise, the feature weights wt are updated with each iteration where
wt  wt − δwt . These updated feature weights wt are used to derive new edge
weights wc (n1 ; n2 ) using Eq. (1) and can be further used for the derivation of success
probabilities of testing request P1 (y) and P0 (y) from the underlying graph as follows:

max u  2 × wc (n 1 × n 2 ) × Fc (n 1 × n 2 ) × wt . (6)
Psc

The next step proceeds with the updation of E p using Eq. (3), and the procedure
is repeated iteratively till E p converges. The success status of y nodes where y ∈ U test
can be derived using success probability is obtained from Eq. (5).

4 Evaluation

The t r data is varied in the ratio of 30%:70% and 70%:30% for analysis. Initially,
k is set to 0.1 and is varied to assess the efficiency of GSP model using different
percentages of t r data. The performance of GSP is evaluated using AUC score and
accuracy, where accuracy  C tPs R where CPR is the count of requests that are correctly
predicted and t s is the sum total of testing data. The percentage of t s and t r data is
selected, and the experiment is repeated up to 200 times. Finally, average of AUC
score and average of accuracy score is computed and used for analysis purpose.

4.1 Competitive Methods

The proposed GSP model is compared with ADJ work [1] and other competitive
classifiers such as SVM [6] and LR [3]. It can be seen from the obtained results
shown in Figs. 1 and 2 that GSP model outperforms ADJ and other competitive
rivals for every case. As can be seen from Fig. 2a that the ADJ approach achieved
0.67 AUC score with 70% t r data while 0.81 AUC score and accuracy score was
achieved by the proposed GSP, with same amount of training data. This is mainly
because of Fig. 3a, b reports the AUC score and execution time for different values
of k. As can be seen from Fig. 3a, when the value of k increases, AUC score also
increases since it increases the size of the graph G. In Fig. 3b, although good results
are achieved by increasing k value, alongside execution time also increases that adds

[Link]@[Link]
232 A. Ahmad et al.

(a) AUC score (b) Accuracy score

Fig. 1 AUC and accuracy score when training data percentage is varied from 5 to 30%

(a) AUC score (b) Accuracy score

Fig. 2 AUC and accuracy score when training data percentage is varied from 40 to 90%

(a) AUC score for different values of k (b) Execution time for different values of k

Fig. 3 AUC score and execution time values obtained under the GSP approach the incorporation of
three extra features, i.e., centrality, role, and topic, that played a major role in the success prediction

to its disadvantage. But for k  0.10, GSP takes 4 s to execute and it is feasible for
practical use, so we set k  0.10.

[Link]@[Link]
A Novel Approach for Predicting the Outcome of Request … 233

5 Conclusion

The proposed work focusses on collaborative knowledge obtained from sites Red-
[Link] and evaluates the efficiency of proposed GSP model for the prediction of
real-world events. Considering the work done by ADJ as the base, a new method-
ology is proposed to predict the outcome of the request. By considering three more
features, i.e., topic, role, and centrality apart from those proposed by ADJ and using
graph-based success prediction (GSP) model, 0.81 AUC score was obtained. It was
observed that the proposed GSP model has outperformed other classifiers and has
shown an improvement over ADJ in terms of AUC score with limited training data.

References

Althoff, T., Danescu-Niculescu-Mizil, C., Jurafsky, D. (ADJ): How to ask for a favor: a case study
on the success of altruistic request. In: Proceedings of AAAI International Conference on Web
and Social Media, ICWSM (2014)
Backstrom, L., Huttenlocher, D., Kleinberg, J., Lan, X.: Group formation in large social networks:
membership, growth, and evolution. In: Proceedings of the ACM SIGKDD International Con-
ference on Knowledge Discovery and Data Mining, KDD, pp. 44–54 (2006)
Freedman, D.A.: Statistical Models: Theory and Practice, pp. 128 (Cambridge University Press,
2009)
Scripps, J., Tan, P.N., Esfahanian, A.H.: Exploration of link structure and community-based node
roles in network analysis. In: Proceedings of the IEEE International Conference on Data Mining,
ICDM, pp. 649–654 (2007)
Ugander, J., Backstrom, L., Marlow, C., Kleinberg, J.: Structural diversity in social contagion. Proc.
Nat. Acad. Sci. U.S.A. (PNAS) 109(16), 5962–5966 (2012)
Vapnik, V.: Support-vector networks. Mach. Learn. 20(3), 273297 (1995)

[Link]@[Link]
A Soft Computing Methodology
for Estimation and Forecasting of Daily
Global Solar Radiation (DGSR)

J. Christy Martina and T. Amudha

Abstract Energy is one of the most crucial building blocks of economic develop-
ment. The energy sector in India has a rapid growth in recent years. The factors that
took the nation to a very acute energy crisis are increase in population, transporta-
tion, urbanization, industrialization, high standard of living and fast depleting fossil
fuels. Energy demand problems have increased all over the world, and renewable
energy sources are more crucial to solve these problems. Solar energy is a stepping
stone to satisfy the growing energy demands in India and across the globe. In this
research work, Coimbatore location was considered for daily global solar radiation
(DGSR) analysis and the meteorological parameters used for this research are min-
imum air temperature, global solar radiation, maximum air temperature, sunshine
hours, wind speed, mean air temperature, extraterrestrial radiation, average atmo-
spheric pressure, average precipitation and relative humidity. Statistical models and
artificial intelligence computational technique with ANFIS, i.e. adaptive neuro-fuzzy
inference system, were used for forecasting and estimation of DGSR. The results of
ANFIS model were found to be the best fit for forecasting and estimation of DGSR
in any region.

Keywords Solar energy · Estimation and forecasting of solar radiation

Statistical models · ANFIS · ANFIS chaotic time series prediction

1 Introduction

Renewable energy is abundant in India and used as an alternative source for electricity
generation. Solar energy, biomass, hydro and wind are the renewable energy sources
that offer sufficient supply of clean energy, economic prosperity, social and envi-

J. Christy Martina (B) · T. Amudha

Bharathiar University, Coimbatore, India
e-mail: martinachristy16@[Link]
T. Amudha
e-mail: amudhaswamynathan@[Link]

© Springer Nature Singapore Pte Ltd. 2019 235

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
236 J. Christy Martina and T. Amudha

ronmental benefits [1]. Indian energy sector approach towards meeting the energy
demands has focussed on resource exploration and exploitation, capacity additions,
creating awareness on the potential and promoting use of renewable energy. Solar
energy is adequate and free resource of energy that plays a vital role in meeting
the energy demands in India and across the globe. India has rich solar resources.
There are 250–325 bright days in a year in India. Most of the regions in India receive
4.7 kWh/m2 /day of solar radiation [2]. The main objective of this research work is
to forecast and estimate DGSR for solving the energy demand in India and across
the globe.
In this research work, Coimbatore region was considered for daily global solar
radiation (DGSR) analysis. Climatological data including daily solar radiation, max-
imum air temperature, relative humidity, minimum air temperature, extraterrestrial
radiation, mean air temperature, wind speed, average atmospheric pressure, average
precipitation, day length and sunshine hours were used for estimation and forecast-
ing of DGSR. The data sets for 11 years (2005–2015) were collected from Agro
Climatic Research Centre (TNAU), Coimbatore, and NASA near real-time data. Sta-
tistical methods were used to develop models using the temperature, sunshine and
meteorological-based parameters. A soft computing technique, ANFIS model, was
developed for the estimation of DGSR based on available climatological parameters.
The developed ANFIS model results were compared with the best statistical mod-
els. The historical data sets of daily observed solar radiation from January 2005 to
December 2015 were used to forecast solar radiation using ANFIS chaotic time series
prediction method. Using the daily observed solar radiation in Coimbatore region,
the corresponding daily solar radiation from January 2016 to December 2026 was
calculated. The newly developed ANFIS models were found to be the best models
and are suitable for forecasting and estimation of DGSR in any region of the world.

2 Estimation and Forecasting of Solar Radiation

Solar radiation estimation analysis helped to understand the need for solar energy and
identify the best models to estimate DGSR. Coimbatore region in Tamil Nadu was
considered for this research work. Statistical models and a soft computing technique,
ANFIS method, were used for estimation of DGSR using the available climatological
parameters. ANFIS chaotic time series prediction method was used for forecasting
DGSR. In this research work, existent statistical models were categorized into three
types: temperature, sunshine and climatological parameter-based models. Nine sta-
tistical models were used in which they include one temperature model, four sun-
shine models and four different climatological parameter-based models to evaluate
the performance of DGSR estimation. The statistical indices such as mean abso-
lute error (MAE), RMSE root-mean-square error, determination coefficient (R2 ) and
mean absolute relative error (MARE) were used to find the best-fit model for imple-
menting solar radiation estimation techniques. The statistical models mentioned in
Table 1 were found as the best models on the basis of investigations on DGSR esti-

[Link]@[Link]
A Soft Computing Methodology for Estimation … 237

Table 1 Statistical models

S. No Type Model Equation
1 Temperature based Hunt et al. [13] Z  j(T max − T min )0.5
Z0 + k
2 Sunshine based Ogelman et al. [14] Z/Z 0  j +
k(SH/SH0 ) +
l(SH/SH0 )2
3 Bahel et al. [15] Z/Z 0  j +
k(SH/SH0 ) +
l(SH/SH0 )2
+m(SH/SH0 )3
4 Newland [16] Z/Z 0  j +
k(SH/SH0 ) +
llog(SH/SH0 )
5 Bakirci [17] Z/Z 0  j +
k(SH/SH0 ) +
lexp(SH/SH0 )
6 Meteorological based Swartman et al. [18] Z/Z 0  j +
k(SH/SH0 ) + lRH
7 Abdalla [19] Z/Z 0  j +
k(SH/SH0 ) + lT +
mRH
8 Ododo et al. model 1 Z/Z 0 
[20] j(SH/SH0 )k T lmax RHm
9 Ododo et al. model 2 Z/Z 0  j +
[20] k(SH/SH0 ) + lT max +
mRH +
nT max (SH/SH0 )

mation models where ‘Z’ is the global solar radiation, ‘Z 0 ’ is the extraterrestrial solar
radiation, ‘SH’ is the actual sunshine hour, ‘SH0 ’ is the maximum possible sunshine
duration/day length, ‘RH’ is the relative humidity, ‘T ’ is the daily mean air temper-
ature, ‘T max ’ is the daily maximum air temperature, ‘T min ’ is the daily minimum air
temperature and ‘j, k, l, m, n’ are empirical coefficients.

3 Adaptive Neuro-Fuzzy Inference System (ANFIS)

ANFIS technique is an integration of ANN knowledge representation, learning abili-

ties, implication capabilities of fuzzy logic into a single system that drives to achieve
desired performance by self-modifying their membership function. Fuzzy logic is
used in ANFIS data learning technique. It transforms the given input into a target
output. ANN learning ability is capable of creating interrelationship with the input
and output variables. The input and output parameters can be related using If-Then

[Link]@[Link]
238 J. Christy Martina and T. Amudha

sentences which are called as rules of fuzzy system. The salient features of ANFIS are
used to manage complex decision-making systems successfully, easy to implement,
and former human knowledge is not required. Figure 1 shows the ANFIS architecture
of Sugeno fuzzy model which consists of two If-Then fuzzy rules [3].
RULE 1: If r is C1 and s is D1 Then k1 = t1r + u1s + v1
RULE 2: If r is C2 and s is D2 Then k2 = t2r + u2s + v2
Here, ‘r’ and ‘s’ are considered as inputs; ‘C i and Di ’ are the fuzzy sets; ‘k i ’ are the
outputs in the fuzzy region specified by the rules of fuzzy; and the design parameters
are ‘t i , ui , vi ’. ANFIS architecture has five layers, namely input fuzzification, second
layer is the application of fuzzy operators, third layer is the application method,
fourth layer is the aggregation of output, fifth is defuzzification layer [4]. Each layer
is given as follows:
Layer one is called as the input fuzzification layer in which all the inputs pass
through membership functions.

OP1,i  μCi (r ); i  1, 2 (1)

OP1,i  μDi (s); i  3, 4 (2)

where ‘r’ and ‘s’ are the inputs to node ‘i’, and ‘C i ’ and ‘Di ’ are the linguistic labels
such as high, low associated with this node function. μC i (r) and μDi (s) can adopt
any fuzzy membership function. Triangular, Pi-shaped, Gaussian are some of the
input membership functions, and there are two output membership functions such
as constant and linear. The nodes are fixed in layer two and represented as a circle
labelled  which multiplies the incoming signals for fuzzification of inputs using
the AND operator [5]. Thus, rules weight is achieved in this layer. Following is the
output of this layer,

Fig. 1 ANFIS architecture with two rules, two inputs and one output [3]

[Link]@[Link]
A Soft Computing Methodology for Estimation … 239

OP2,i  wi  μCi (r ).μDi (s); i  1, 2 (3)

Every node in layer three consists of a circle node labelled as N specifies that they
play a normalization role to the firing strengths from the preceding layer. Following
is the output of this layer,
wi
OP3,i  w̄i  ; i  1, 2 (4)
w1 + w2

Defuzzification is the fourth layer. In this every node, it is an adaptive node, with
consequent parameters that include ti , ui and vi . Given below is the output of this
layer

OP4,i  w̄i ki  w̄i (ti r + u i s + vi ); i  1, 2 (5)

In layer five, single node is labelled as . Summation of all incoming signals is

performed in this layer in which it represents the overall output of the ANFIS system
which is given below

 wi ki
OP5,i  w̄i ki   ; i  1, 2 (6)
wi ki

3.1 Solar Radiation Estimation Using ANFIS

Besides the statistical models, artificial intelligence ANFIS technique is one of the
powerful tools to estimate DGSR using the climatological parameters. The learning
procedure of ANFIS is similar to neural networks [6]. In the present work, ANFIS
model was built using grid partition method with back-propagation learning algo-
rithm for estimation of GSR. The ANFIS GUI editor was used to load both the
training and checking data sets. The climatological parameters including observed
solar radiation, maximum air temperature, extraterrestrial radiation, wind speed, min-
imum air temperature, precipitation, relative humidity, sunshine hours, atmospheric
pressure and day length were used as input parameters for training and checking
the data. Gaussian curve membership function with three numbers of membership
function and linear type of output was used to generate FIS structure. Then the
back-propagation optimization method was used to train the data set. The number
of training epochs (1000) and the training error tolerance (0) were set to stop the
training process. Checking process was done after training the FIS and plotting the
checking data against the FIS output. Finally, the average testing error for training
and checking data was obtained by applying the checking data to the model. The
best model was obtained based on the RMSE error rate.
ANFIS Chaotic Time Series Prediction for Forecasting Solar Radiation: Time
series prediction is one of the most important prediction methods that refer to the past

[Link]@[Link]
240 J. Christy Martina and T. Amudha

recorded observations of variables. The values of the variables are measured at spec-
ified intervals of time and analysed to obtain the underlying relationships between
observations and finding out a descriptive example [7]. This model is applied to
extrapolate future time series. The present work was performed with ANFIS chaotic
time series prediction modelling on 11 years of meteorological data during the period
2005–2015 for Coimbatore, Tamil Nadu. In time series prediction, Ti represents the
known values of the time series up to the point in time and to predict the future value
at some point, Ti + P are used. The prediction can be done by creating a mapping
from DP sample data points, sampled every  units in time, x (Ti − (DP − 1) ),
…, x (Ti − ), x (Ti), to a predicted future value x =(Ti + P) [8]. To predict the solar
radiation, DP is set to 4 and   P  3. For each Ti, the input training data for ANFIS
is a four-column vector of the form, ω(Ti)=[x(Ti − 9), x(Ti − 6), x(Ti − 3), x(Ti)].
The output training data that corresponds to the trajectory prediction is represented
as δ (Ti)= x (Ti + 3). For each Ti, ranging in values from 12 to 4013, the training
input–output data is a structure whose first component is the four-dimensional input
ω, and whose second component is the output δ. There are 4017 input–output data
values. The first 2812 data values were used for training the ANFIS model, while
the others were used as checking data for validating the identified fuzzy model.

4 Implementation Results and Discussion

In this research work, estimation and forecasting of solar radiation for Coimbatore
region were implemented successfully with real-time data collected for the period
2005–2015. MATLAB and Microsoft Office Excel 2007 were used for implemen-
tation. Microsoft Office Excel 2007 was used to compute statistical equation for
the estimation of DGSR, and statistical indices including RMSE, MARE, MAE,
correlation coefficient using the formulae. Soft computing technique, ANFIS, was
implemented using MATLAB R 2010 a. ANFIS GUI editor was used to train, test
and plot the data. RMSE error rate was obtained in order to find the best model for
the estimation of DGSR. Best empirical equations based on sunshine, temperature
and climatological parameters were identified and implemented for estimation of
DGSR. ANFIS models were developed using different combinations of available
meteorological parameters for estimation of daily GSR. Statistical evaluation was
performed using performance measures RMSE, MAE, MARE, R2 , and the obtained
results showed that there was a good agreement between the observed and estimated
value by Bahel et al. model and ANFIS_M11 model. ANFIS chaotic time series
prediction method was implemented for forecasting of solar radiation for the period
2015–2026. The developed model is suitable for the regions with similar climatic
condition and unavailability of recorded DGSR data.

[Link]@[Link]
A Soft Computing Methodology for Estimation … 241

4.1 Performance Measures for Evaluation of Statistical

Models

In this research work, statistical indices such as mean absolute error (MAE), root-
mean-square error (RMSE), determination coefficient (R2 ) and mean absolute relative
error (MARE) were used to find the best-fit model for implementing solar radiation
estimation techniques. The following are the formulas used to find the statistical
indices [9].

N
i1 (Hi,c − Hi,m )
2
RMSE  (7)
N
1   
N
MAE  Hi,c − Hi,m  (8)
N i1
N  
1   Hi,c−Hi,m 
MARE  ∗ 100% (9)
N i1  Hi,m 
⎛ ⎞2
N
⎜ Hi,c − Hc,avg Hi,m − Hm,avg ⎟
R2  ⎜
⎝
i1 ⎟
⎠ (10)
N 2 N 2
i1 Hi,c − Hc,avg i1 Hi,m − Hm,avg

where ‘H i,c ’ represents the ith estimated daily solar radiation, ‘H i,m ’ is the ith
observed daily solar radiation, ‘H c,avg ’ is the mean of estimated values, and ‘N’
is the number of daily solar radiation values. The lower values of MAE, MARE and
RMSE indicate good precision of the model, and higher value of R2 indicates the best
model. R2 takes the value between 0 and 1. The performance measures of empiri-
cal equation were obtained by Eqs. (7–10). Performance analysis of the statistical
models implemented in this research work is shown in Table 2.
The results showed that there was a good agreement between the estimated and
observed value by Bahel et al. model with lower value of RMSE 1.096, MAE 0.821,
MARE 20.56% and higher value of R2 0.969. Thus, Bahel et al. sunshine-based model
proved to be best fit among other statistical models for estimating DGSR.
Estimation and Forecasting of GSR Using ANFIS: Meteorological parameters
data sets with 4017 samples were used to build ANFIS model in which 2812 were
used for training and 1205 for testing the model. Grid partition with back-propagation
learning algorithm and Sugeno fuzzy inference system was used. Three membership
functions for each input were assigned, such that rules were framed based on the
number of input parameters. ANFIS model was trained effectively, and testing was
performed using the data samples that were not involved in the training process. Sta-
tistical performance measures such as RMSE, MAE, MARE and R2 were calculated
to compare the estimated and predicted values in which the best fit of the model
was obtained. Fuzzy inference system (FIS) was generated using grid partition with

[Link]@[Link]
242 J. Christy Martina and T. Amudha

Table 2 Performance evaluation of statistical models

Model RMSE MAE MARE (%) R2
Hunt et al. model 1.626 1.334 34.08 0.955
Ogelman et al. 1.101 0.821 20.55 0.969
model
Bahel et al. 1.096 0.821 20.56 0.969
model
Newland model 1.123 0.821 21.62 0.968
Bakirci model 2.088 0.859 21.28 0.882
Swartman et al. 1.276 0.971 25.28 0.969
Model
Abdalla model 1.264 0.959 24.95 0.969
Ododo et al. 2.957 2.753 60.77 0.962
model 1
Ododo et al. 1.318 1.017 26.54 0.969
model 2

back-propagation optimization method. Each input parameters were assigned with

three membership functions and a linear type of output. The proposed ANFIS model
structure is shown in Fig. 2.
Gaussian curve membership function was identified as the best membership func-
tion based on evaluation of temperature-based model during the year 2009. Average
testing error of training and checking was obtained to find the best fit of the member-
ship function. The meteorological parameters for Coimbatore region were used as
inputs to train the ANFIS model. Various combinations of meteorological parameters
were used to train and test the ANFIS models. Statistical indices of ANFIS model

Fig. 2 Structure of ANFIS model

[Link]@[Link]
A Soft Computing Methodology for Estimation … 243

were used to find the best model for estimation of DGSR. The results of statistical
analysis of ANFIS models proved that ANFIS_M11 with available climatological
parameters such as relative humidity RH, bright sunshine hours S, day length S 0 ,
mean air temperature T , precipitation P and extraterrestrial radiation H 0 gave better
result than other models. ANFIS_M11 model obtained lower value of RMSE 0.722,
MAE 0.545, MARE 13.62% and higher value of R2 0.981. Therefore, ANFIS_M11
model showed good performance and can be used for estimation of DGSR across
the globe with similar climatic condition with available meteorological parameters.
The performance comparison of statistical and ANFIS model using statistical per-
formance measures given in Table 3 showed that ANFIS_M11 model outperformed
the Bahel et al. statistical model. Thus, ANFIS_M11 model proved to be the best
model for estimating daily GSR among other models. Figure 3 shows comparison
graph of daily observed GSR, statistical and ANFIS_M11 model.
ANFIS chaotic time series prediction method was used for forecasting daily global
solar radiation. Eleven years of observed solar radiation during the period 2005–2015
for Coimbatore region were given as input parameter in ANFIS chaotic time series
modelling. Data sets of 2812 × 1 for training and 1205 × 1 for checking FIS were
used. This model was able to predict DGSR for the period 2016–2026. The developed
model was validated against actual DGSR for the period January 2016 to October
2017. The results of this model showed better performance for forecasting DGSR
in Coimbatore region, and thus, the proposed model can be implemented easily for

Table 3 Performance comparison of statistical and ANFIS model using statistical performance
measures
Model RMSE MAE MARE (%) R2
Bahel et al. 1.096 0.821 20.56 0.969
model
ANFIS_M11 0.722 0.545 13.62 0.981
model

10.00
9.00
Solar Radiation

8.00
7.00 Observed GSR
6.00 Bahel et al model

5.00 ANFIS_M11 model

4.00

Year

Fig. 3 Comparison graph of daily observed GSR, statistical and ANFIS_M11 model

[Link]@[Link]
244 J. Christy Martina and T. Amudha

8.00

Solar RadiaƟon kWh/m2/day

7.00
6.00
5.00
4.00
3.00
2.00
1.00
0.00

Fig. 4 Monthly average solar radiation from 2005 to 2026

any other region. Tamil Nadu has rich source of solar energy. There are around
300 clear bright days that receive high solar radiation. This shows that the state
has higher potential for solar power generation. Coimbatore region in Tamil Nadu
was considered for this research work has annual average of 5.06 kWh/m2 /day solar
radiation [10]. The days are bright and sunny in the month of February to April and
receive rainfall in the months June to August. The climate is warm and humid during
September to November and winter from December to January. Figure 4 shows the
predicted and observed monthly average solar radiation for the period 2005–2026.

5 Conclusion and Future Enhancements

India is in the world’s third place to produce largest electricity and fourth in largest
consumer of electricity. The conventional electricity generation using coal, oil and
gas has a high impact on environment. The pollutants added to the environment
by these methods affect the ecological balance of plants, animals and humans [11].
Thus, there is an earnest requirement for other ecological friendly resources. These
resources are green source of energy including solar, wind, hydro, biomass and
geothermal, and among these, solar is readily available during the day, abundant,
clean and free source of energy. Solar radiation forecasting acts as a primary key to
ensure the efficient use of energy that can help in better planning and distribution of
energy in photovoltaic system. Solar radiation value is not available for all locations,
and accurate forecasting is very important in solving energy problems. The major
goal is to motivate the local governments to implement the green source of energy
with efficiency measures. Coimbatore City is also one among 48 cities identified for
developing as solar cities, up to 50 lakhs per city is provided for the development of

[Link]@[Link]
A Soft Computing Methodology for Estimation … 245

the project [12]. Every year the generation of electricity by solar increases with higher
demand. This research work was done to depict the importance of solar energy and
utilize them in a better way to meet energy demands with pollution free environment
in future. The developed ANFIS model is well suitable and can be adopted to esti-
mate DGSR with available meteorological parameter for any location. This research
work can be extended in the field of analysing the performance of power production
in photovoltaic systems based on DGSR. The present work results are best fit for
estimating DGSR with available meteorological parameters over Coimbatore region.
In future, satellite images can be used as an alternative to ground-based measuring
networks to minimize the cost for installing expensive measuring equipment’s and
for environments with inadequate meteorological data. Further, investigation of the
relationship between solar radiation and climatological parameters is recommended
to minimize the input parameters with more accurate results.

References

1. [Link]
energy-inexhaustible-sources
2. Dwivedi, A., Bari, A., Dwivedi, G.: Scope and application of solar thermal energy in India—a
review. Int. J. Eng. Res. Technol. 6(3), 315–322 (2013)
3. Al-Hmouz, A., Shen, J., Al-Hmouz, R., Yan, J.: Modeling and simulation of an adaptive neuro-
fuzzy inference system (ANFIS) for mobile learning. IEEE Trans. Learn. Technol. 5(3) (2012)
4. Mehdizadeh, S., Behmanesh, J., Khalili, K.: Comparison of artificial intelligence methods and
empirical equations to estimate daily solar radiation. J. Atmos. Solar Terr. Phys. 146, 215–227
(2016)
5. Sumithira, T.R., Nirmal Kumar, A.: Prediction of monthly global solar radiation using adap-
tive neuro fuzzy inference system (ANFIS) technique over the state of Tamilnadu (India): a
comparative study. Appl. Solar Energy 48(2), 140–145 (2012)
6. [Link]
7. [Link]
8. [Link]
9. Krishnaiah, T., Srinivasa Rao, S., Madhumurthy, K., Reddy, K.S.: Neural network approach
for modelling global solar radiation. J. Appl. Sci. Res. 3(10), 1105–1111 (2007)
10. [Link]
11. [Link]
12. [Link]
13. Hunt, L.A., Kuchar, L., Swanton, C.J.: Estimation of solar radiation for use in crop modelling.
Agric. For. Meteorol. 91, 293–300 (1998)
14. Ogelman, H., Ecevit, A., Tasdemiroglu, : A new method for estimating solar radiation from
bright sunshine data. Sol. Energy 33(6), 619–625 (1984)
15. Bahel, V., Bakhsh, H., Srinivasan, R.: A correlation for estimation of global solar radiation.
Energy 12(2), 131–135 (1987)
16. Newland, F.J.: A study of solar radiation models for the coastal region of South China. Sol.
Energy 43(4), 227–235 (1989)
17. Bakirci, K.: Correlations for estimation of daily global solar radiation with hours of bright
sunshine in Turkey. Energy 34, 485–501 (2009)

[Link]@[Link]
246 J. Christy Martina and T. Amudha

18. Swartman, R.K., Ogunlade, O.: Solar radiation estimates from common parameters. In: Solar
Energy Conference, Tempe, Arizona (1967)
19. Abdalla, Y.A.G.: New correlations of global solar radiation with meteorological parameters
for Bahrain. Int. J. Solar Energy 16, 111–120 (1994)
20. Ododo, J.C., Sulaiman, A.T., Aidan, J., Yuguda, M.M., Ogbu, F.A.: The importance of maxi-
mum air temperature in the parameterisation of solar radiation in Nigeria. Renew. Energy 6(7),
751–763 (1995)

[Link]@[Link]
A Shapley Value Approach
for Community Detection in Social
Network

Amreen Ahmad, Tanvir Ahmad, Abhishek Bhatt and Sadaf Siddiqui

Abstract The increasing popularity of online social networks (OSNs) has attracted
significant attention from the research community. One of the significant tasks of
analyzing structure is to detect communities within these OSN. Nodes within a
community tend to be densely connected in comparison with nodes outside the com-
munity. Community detection is a challenging task in the field of social networks
(SNs). Graph model is used to represent OSN where the nodes are representative of
actors and the edges represent the interactions between these actors. A novel com-
munity detection algorithm named SCD is proposed in this paper. In the underlying
graph, nodes are modeled as rational players where each player wants to maximize
its Shapley value in terms of interactions. Extensive experiments are conducted on
real-world datasets to establish the efficiency of the proposed approach.

Keywords Online social network · Community detection · Shapley value

1 Introduction

Community detection in graphs is gaining serious attention from researchers since

it has many applications in different fields such as recognizing functional groups in
metabolic networks [1], discovering friendship network in social network [2], and
discovering the relationship between topics in World Wide Web [3]. In addition, the

A. Ahmad (B) · T. Ahmad · A. Bhatt · S. Siddiqui

Department of Computer Engineering, Jamia Millia Islamia, New Delhi, India
e-mail: amreen.ahmad10@[Link]
T. Ahmad
e-mail: tahmad2@[Link]
A. Bhatt
e-mail: abhishek1995bhatt@[Link]
S. Siddiqui
e-mail: sadafsiddiqui1992@[Link]

© Springer Nature Singapore Pte Ltd. 2019 247

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
248 A. Ahmad et al.

performance of recommender system can be enhanced with the detection of related

customers in purchase network.
Graph model is adopted to model social networks where actors are represented by
nodes and edges represent the interactions such as financial exchange, common inter-
est, and friendship between these nodes. A lot of interesting properties is exhibited
by social networks, to be mentioned as community structure, power-law degree dis-
tributions, and network transitivity. A community can be defined as a cluster in which
members within a group are likely to be closely knitted as compared to members
outside the group.
Since last decade researchers have proposed a considerable amount of algorithms
to detect communities in a social network. The limitation of classical community
detection algorithms is time complexity that often fails in case of large-scale social
networks. Clauset [2] presented a greedy approach to detect considered every node
as a community, and then, two nodes were merged only when they had maximum
modularity value.
Researchers [4] have analyzed that user’s interactions play a significant role in
detecting the communities within the SN since users interact with a limited number of
their social group. Hence, analysis of community structure within these SNs should
be based on activity instead of the relationship. Weighted links can be used to capture
activity in these SNs. The proposed research focuses on static community detection
approach.

1.1 Contributions of the Paper

In the proposed work, a novel clustering algorithm is proposed to detect communities

in real-world social networks. The work can be summarized as:
1. A Shapley value-based clustering algorithm is proposed for real-world networks.
Graph model is adopted to model nodes as rational players that want to maximize
its Shapley value in terms of profit (where profit represents interactions).
2. The efficiency of the proposed work is established by comparing it with other
competitive methods.
The remaining part is structured in the following way: Sect. 1 proceeds with the
discussion of related work followed by Sect. 2 that gives a detail explanation of the
proposed methodology. We proceed with an analysis of results obtained from the
evaluation discussed in Sect. 3 followed by conclusion discussed in Sect. 4.

2 Proposed Methodology

This section gives a detailed explanation about the procedure for clustering algorithm
SCD and is divided into the following modules:

[Link]@[Link]
A Shapley Value Approach for Community Detection … 249

1. In the underlying graph, salient nodes are determined by using SPIN algorithm
[5] for every node.
2. The salient nodes identified are used to determine the clusters in the underlying
graph discussed in Algorithm 1.

2.1 Clustering Algorithm (SCD)

It is a four-step algorithm that discusses the creation of clusters in the underlying

graph. A clustering algorithm is given in a pseudocode form in algorithm 1, and the
explanation is given as: (i) Initially, the graph model is adopted to represent actors
as nodes and edges representing the interaction between these actors. (ii) Next, we
proceed with the identification of salient nodes that is determined by

Algorithm 1 Clustering Algorithm

Also let [] contain the Shapley values of all nodes in decreasing order
Let k represent the number of salient nodes that is given as input
procedure GETCLUSTERS(G, Φ)
Let n represent the nodes in graph G
S  GETSALIENTNODES(G, , k)
Let Sa represent the set of SalientNodes that are smallest subset of S such that it
gives a coverage of 100% in the underlying graph G
for every node in G different from Sa do
Compute shortest path to all Sa
closestSalientNode  Identify the closest SalientNode
end for
for every node in G do
newGraph  The node belongs to the community of its S a
end for
return newGraph
end procedure

SPIN algorithm. (iii) Finally, the task remains to identify clusters of the graph. In
algorithm 1, first, we have to identify those set of SalientNodes that are the smallest
subset of S such that it gives a coverage of 100% in the underlying graph G and name
it as Sa. Then for every non-salient node, its closest salient node is identified. After
the closest salient node of a node is identified, the non-salient node belongs to the
community of its closest salient node. In the last, n clusters (clusters representing
community) are given as output (where n is the number of SalientNodes).

[Link]@[Link]
250 A. Ahmad et al.

3 Evaluation

3.1 Dataset

1. Zachary karate club [6]—This dataset represents a friendship network of karate

club consisting of 34 nodes and 74 edges where some nodes are influential than
others present in the network. This is used as a benchmark dataset for detecting
communities within a network.
2. Dolphin [7]—This dataset is an unweighted and undirected network that repre-
sents frequent interaction among 62 dolphins and is divided into two communities
due to some reasons.
3. Football network [2]—This dataset is a network of football teams split into inde-
pendent teams and 11 conferences.
4. US political books(UPB)1 —This dataset represents a network for US political
books constituting 105 books, an interaction network of primary school students,
and a collaborative network of 241 physicians.

3.2 Evaluation Metrics

1. Modularity: It is one of the well-known metrics for measuring the quality of

clusters in social networks (SNs). Modularity can be defined in the following
way:

  2 
q
ec dc
Qm  − (1)
c1
e 2e

where q denotes the number of detected communities, eC represents the count of

edges in community C, e denotes the aggregate count of edges in the graph G, and
d C represents the aggregate sum of the degrees of nodes in community C.
2. Clustering coefficient: It is a metric that measures the extent to which nodes tend
to form clusters. It is a local measure, and it can be defined in the following way
for a node:

Ei
Ccoeff(i)  (2)
dgi (dgi − 1)

where Ccoeff(i) denotes the clustering coefficient for node i, dgi computes degree of
node i, and E i is count of edges between neighbors of node i.

1 [Link].

[Link]@[Link]
A Shapley Value Approach for Community Detection … 251

Table 1 Performance comparison based on Qm

Algorithms Dataset
Zachary Dolphin Football UPB
CC-GA 0.420 0.529 0.594 0.527
Newman 0.381 0.465 0.524 –
GN 0.401 0.519 0.599 –
CNM 0.0.381 0.495 0.577 –
SCD 0.51 0.73 0.90 0.70

Table 2 Performance comparison based on C coef f (i)

Algorithms Dataset
Zachary Dolphin Football UPB
Bi-section 0.56 0.21 0.40 –
K means
SCD 0.56 0.26 0.60 –

3.3 Comparative Study and Analysis

Experiments are performed on four real-world datasets to analyze the efficiency of the
SCD. The results are further analyzed by comparing SCD with well-known methods:
(i) CC-GA [8], (ii) Newman [2], (iii) GN algorithm [2], (iv) CNM algorithm [9],
and (v) bi-section k means [10]. The modularity and clustering coefficient values for
different algorithms are presented in Tables 1 and 2. As can be seen from Tables 1 and
2, the modularity and clustering coefficient values of the proposed SCD community
detection algorithm are better than the other competitive methods. The reason can
be inferred as the incorporation of information diffusion concept along with Shapely
value has lead to the creation of good quality clusters. Hence, we arrive at a conclusion
that the proposed SCD is better than other algorithms.

4 Conclusion

A Shapley value-based approach for detecting communities in the underlying net-

work is proposed. For every node in the graph, SPIN algorithm is used to identify
salient nodes. Then, the clustering algorithm is applied to determine communities.
The results obtained prove the efficiency of the proposed approach over other
algorithms. The future work remains to extend it to dynamic networks on large
real-world networks.

[Link]@[Link]
252 A. Ahmad et al.

References

1. Chen, J., Yuan, B.: Detecting functional modules in the yeast protein–protein interaction net-
work. Bioinformatics 22(18), 2283 (2006)
2. Girvan, M., Newman, M.E.J.: Community structure in social and biological networks. Proc.
Natl. Acad. Sci. 99(12), 7821–7826 (2002)
3. Flake, G.W., Lawrence, S.R., Giles, C.L., Coetzee, F.M.: Self-organization and identification
of web communities. IEEE Comput. 35, 66–71 (2002)
4. Viswanath, B., Mislove, A., Cha, M., Gummadi, K.P.: On the evolution of user interaction in
facebook. In: Proceedings of the 2nd ACM workshop on Online Social Networks, pp. 37–42.
ACM, New York (2009)
5. Narayanam, R., Narahari, Y.: A shapley value-based approach to discover influential nodes in
social networks. IEEE Trans. Autom. Sci. Eng., pp. 1–18 (2011)
6. Zachary, W.: An information flow model for conflict and fission in small groups. J. Anthropol.
Res. 33, 452–473 (1977)
7. Lusseau, D.: The emergent properties of a dolphin social network. Proc. Natl. Acad. Sci.
270(Suppl. no. 0962-8452), S186–8 (2003)
8. Saida, A., Abbasi, R.A., Maqbool, O., Daudb, A., Aljohani, N.R.: CC-GA: a clustering coeffi-
cient based genetic algorithm for detecting communities in social networks. Appl. Soft Com-
put., pp. 59–70 (2018)
9. Moore, C., Newman, M.E.J.: Phys. Rev. E Stat. Phys. Plasmas Fluids Relat. Interdiscip. Top.
61, 5678–5682 (2000)
10. Raju, E., Hameed, M.A., Sravanthi, K.: Detecting communities in social networks using unnor-
malized spectral clustering incorporated with bisecting K-means. In: Proceedings of 2015
IEEE International Conference on Electrical, Computer and Communication Technologies
(ICECCT) (2015)

[Link]@[Link]
Multiple Face Detection Using Hybrid
Features with SVM Classifier

Sandeep Kumar, Sukhwinder Singh and Jagdish Kumar

Abstract Nowadays, multiple face detection (MFD) and extraction play an impor-
tant role in face identification for various applications. In the proposed algorithm,
Support Vector Machine (SVM) has been used for multiple face detection, and Dis-
crete Wavelet Transform (DWT), Edge Histogram (EH), and Auto-correlogram (AC)
are used for feature extraction. The proposed methodology worked on two different
database i.e. Carnegie Mellon University (CMU) and BAO database for MFD. In
this research paper, the proposed methodology gives a better result than the existing
technique. Finally, our accuracy raised up to 90% approximately.

Keywords Face detection · SVM · DWT · Edge Histogram · Correlogram

Median filter

1 Introduction

In face detection, the different type of facial expression recognition (FER) will play
a major role to identify the human face, for example, ID, security framework, human
asset information, and participation administration. The initial step to be taken for
recognizing the face is face detection to find a face in a picture. Face detection is
an innovation to decide whether a static picture incorporates a face and recognizes
it. The face detection accuracy depends on the various kinds of the parameters such
as the size of the pictures, horizontal or vertical turning, profiles and front faces,
facial appearances, and light conditions, which cause reduced recognition rate. In

S. Kumar (B) · S. Singh

Department of Electronics & Communication, PEC University of Technology, Chandigarh, India
e-mail: [Link]@[Link]
S. Singh
e-mail: sukhwindersingh@[Link]
J. Kumar
Department of Electrical Engineering, PEC University of Technology, Chandigarh, India
e-mail: jagdishkumar@[Link]

© Springer Nature Singapore Pte Ltd. 2019 253

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
254 S. Kumar et al.

like manner, as a result of the trouble in and the significance of FD, this territory of
research is viewed as a free segment instead of a preparatory advance of FR. As of
late, the regions of FD application have kept expanding [1–3]. Picture resizing will
be performed in this exploration. The viability of the FD calculation is influenced
by the measure of the information picture [1, 4]. A huge picture size will devour all
the more handling time yet will keep up the greater part of the imperative data and
give a better outcome, while a little picture will devour quick time yet may cause low
execution as the face data. Consequently, this calculation begins with a pre-preparing
stage where pictures will be resized to a settled size. This progression can upgrade
the execution of a standard handling time. The way toward expanding the pixels’
size will enhance the location exactness. In any case, decreasing the pixels’ size will
accelerate the recognition time; however, the precision will be disintegrated [5, 6].

2 Background

Sr. No. Authors’ name Year Remarks

1 Zhou et al. [7] 2018 • Identity-specific localized metric
learning technique to identify the
real-world multiple faces
• Three datasets used: class 4, class 8,
and class 16
• Comparing the result with Euclidean
metric, LMNN, ITML, etc.
2 Shah et al. [6] 2017 • Threefold classification used
• Six natural expressions used for
multiclass data
• Time-consuming process during
training
3 Mahajan et al. [8] 2017 • Quahog feature used for robust face
detection
• The proposed methodology gives huge
changes on contortive images
4 Nasr et al. [9] 2017 • Viola–Jones methodology used for face
detection
• SURF feature is extracted with
Bag-of-Words technique
• More accurate and fast recognition
5 Tao et al. [10] 2017 • Proposed a locality-sensitive SVM with
kernel combination for feature
extraction
• Multiple-cascaded neural network to
joint for local features
(continued)

[Link]@[Link]
Multiple Face Detection Using Hybrid Features … 255

(continued)
Sr. No. Authors’ name Year Remarks
6 Julina et al. [11] 2017 • Histogram of gradient for feature
extraction and SVM for classification
• Database used AT&T due to its various
challenges and poses
• Accuracy improved and got the
minimal false-positive rate
7 Naik et al. [12] 2015 • Used a fusion RGB and depth image
technique to identify the real-time face
• Time process of face recognition
reduces
• Accuracy also improved with speed
8 Yadav et al. [13] 2015 • Skin color segmentation and facial
feature for face detection
• Increase the accuracy
• Two datasets used: BAO and CMU
9 Raghavendra et al. [14] 2014 • Light-field camera used for face
recognition in the real world
• Data collected by the conventional
camera also
• In future, the author wants to work on
parallax effect and 3D reconstruction

3 Problem Statement

The fundamental issue to be explained is to actualize a calculation for MFD in a

picture. At a first look, the undertaking of MFD may not appear to be so overpowering
particularly considering how simple it is tackled by a human. However, there is a
conspicuous difference to how troublesome it really is to make a PC effectively
unravel this errand.

4 Proposed Methodology

The proposed methodology contains various steps, i.e., face detector, noise removal,
and feature extraction and classification which are explained in Fig. 1.

[Link]@[Link]
256 S. Kumar et al.

Fig. 1 Block diagram of

proposed methodology Input Image

Face Detector

Crop Face Size

(75X75)

Median Filtering Load

Crop
Image

EHD, AC and DWT for

Feature Extraction Constructing
classifier for
two classes

Memory

Face Classification

4.1 Face Detection (FD)

The essential rule of the Viola–Jones FD calculation is to check the locator commonly
through a similar picture—each time with another size. A picture contains at least
one faces clearly an inordinate substantial measure of the assessed sub-windows
would even now be negatives (non-faces). Each phase of the course was prepared to
utilize a positive set, a negative set, and for execution estimation an assessment set.
For each stage, the positive set and the assessment set were the same (informational
collection A) while the negative set was particularly intended for precisely that stage.
As depicted before false positives are favored over false negatives, and since the
AdaBoost calculation goes for limiting false negatives it needs a touch of tweaking.
In their paper, Viola–Jones depicts the methodology that step by step brings down
the classifier limit of a phase to the point that a given execution necessity is met [15].
Figure 2 shows (a) original image, (b) face-detected image using Cascade classifier,
and (c) face-registered image in the folder.

[Link]@[Link]
Multiple Face Detection Using Hybrid Features … 257

Fig. 2 a Original image, b

detected face, and c crop
detected image

(a) Test Image (b) Face Detection (c) Face

Registration

4.2 Feature Extraction

• Edge Histogram Descriptor: Different types of edges will provide the complete
information regarding the human face. To investigate this component, one path is
to utilize histogram. Edge Histogram Descriptor (EHD) gives information about
neighborhood pixels of edges over the whole picture. EHD essentially speaks to
five sorts of edges in a neighborhood which is known as sub-picture. Figure 3 shows
the five sorts of edges and all the edges are used to process the EHD. To register
80-receptacle neighborhood histogram, the primary picture is partitioned into 4 ×
4 equivalent estimated non-covering pieces. The nearby histogram is registered
for each of these sub-pictures for each edge sort [16].

Fig. 3 Various types of edges: a vertical, b horizontal, c 45°, d 135°, and e non-directional

[Link]@[Link]
258 S. Kumar et al.

• Discrete Wavelet Features: The 2D-DWT can be characterized effectively as far

as the 1D-Haar. 2D-DWT comprises two stages. The primary stage is applying the
1D-DWT on the 2D input signal in row-/column-based order; while, the second
stage is applying the 1D-DWT on the yield of the main stage in column-/row-
based order. The 2D-DWT can be characterized in view of a falling–sifting plan.
In such a case, it can be characterized as a two-phase separating process i.e. low
pass as well as high pass, and each channel is trailed by another two channels.
Figure 3 demonstrates a case of an info picture and its relating DWT up to level
2. It additionally demonstrates the relating yields from the channels [17] (Figs. 4,
5, 6, and 7).
• Correlogram Color Descriptor: Smith and Chang [18] gave a proposed method-
ology to remove the regions of similar color (or dark level) in the pictures by
requantizing the color space of the pictures. In our approach, we utilize this stan-
dard to diminish the quantity of the levels G in the pictures. This is finished by
requantizing the pictures previously processing of the Correlograms. Utilizing this
technique, we accomplish two critical advantages: (1) The decline of G helps the
computational cost, and (2) quantization of the picture into sets of comparative
levels sums up the picture content, which enhances the recovery, comes about.

Fig. 4 2D-DWT scheme

[Link]@[Link]
Multiple Face Detection Using Hybrid Features … 259

Algorithm I

Step 1: Bit by bit brings down the threshold in proper strides until
the point when a given genuine positive rate is
accomplished.
Step 2: Run the negative cases through the course utilizing the
threshold decided in [Link] false positive rate is above 50 %:
Step 3 (a): Increment the measure of highlights by a proper number
and go back to [Link]:
Step 3 (b): Utilize the present threshold and measure of highlights
for this stage. Create another negative preparing set and
another stage containing an indistinguishable measure of
highlights from the present stage. Presently concentrating on
the new stage go back to 1.

Fig. 5 Algorithm for preprocessing steps

Fig. 6 Algorithm for

median filter steps

4.3 Classification Using SVM

SVM is one of the popular and more accurate classifiers for image classification and
recognition of the face. The SVM classifier is more accurate for two classes, but error
will increase for multiclass classification; then multiclass SVM classifier will reduce
the problems which were occurring in binary [19]. SVM is based on hyper-plane, and
this hyper-plane gives the information between two classes; numerous hyper-planes

[Link]@[Link]
260 S. Kumar et al.

Algorithm III for Feature Extraction

Step 1: After removing the noise image is converted into 4 sub images
DWT and each sub images are high as well as low frequency
coefficients.
Step 2: Images are further split with DWT to get the more accurate data in
limited no of features. In the proposed methodology, decomposition
is done up to 4 levels.
Step 3: All the 4 level pixels of sub images gives additional as well as more
accurate information.
Step 4: The features are extracted from images by mathematical formula of

Y: pixels value of image, N: size of image

Step 5: For every filtered image, we have to calculate the histogram using
EHD with the help of different type of edges i.e. vertical edges,
horizontal edges, 45º diagonal, 135º diagonal and non directional.
Sobel operator used to determine the max value of gradient.

: Filtered image
: Edges of images determined by canny method. Then
multiply both:

: Filtered image, : canny illustration

: Histogram of image
Step 6: Assume [D] represents a group of static distances image, I is
represented the AC for close couple at a distance d.

p1 and p2 are the probability of the pixels, d: distance between each

other, are of a comparative level gi. The partition measure between
the histograms, AC, and it is the L1-standard that is computationally
light system and used as a part of [13] and [14].

Fig. 7 Algorithm for feature extraction steps

can be attracted to isolate the two classes of training information, and SVM chooses
ideal hyper-plane. The hyper-plane linear equation can be represented as:

[Link]@[Link]
Multiple Face Detection Using Hybrid Features … 261

hy + w  0

y denotes hyper-plane point,

h represents the hyper-plane,
w represents the bias,
(|w|/h) hyper-plane is perpendicular to origin,
h Euclidean mean.

5 Result Analyses

Two databases, specifically BAO [20] database and CMU [21] database, are utilized
to assess the proposed framework. The BAO database consists of 274 pictures with
single countenance and various faces in the diverse foundation. Another face database
from the CMU contains 60 pictures in the database with various sorts of pictures:
Four side-view pictures and three frontal pictures are utilized. Every one of the
reenactments is executed on this PC by designing programming, MATLAB 14b keep
running on Windows 8 Home Edition OS. The proposed approach was assessed by
utilizing two comprehensively utilized face databases. The proposed approach was
assessed by utilizing two comprehensively utilized face databases. In the analysis,
proposed methodology gives better results on BAO database [22] as well as CMU-1
database. SVM classifier is used to isolate multiple or single face in a given image. A
total of 110 features are used for verification and threefold cross-validation technique
used to find the correct detection. Finally, this methodology gives better result as
compared to existing algorithm as shown in Tables 1 and 2.

• Calculate detection rate (CDR) and false-positive rate (FPR) of the system using
this formula:

CDR(Re call)  TP/TP + FN

Table 1 Accuracy comparison between the proposed methodology and existing approach on BAO
database
Method CDR (%) FPR (%)
Reference 84.2 4.2
Proposed 89.09 10.90

Table 2 Result obtained using the proposed system for CMU-1 database
Method CDR (%) FPR (%) Time (s)
Proposed 88.67 11.32 3.56

[Link]@[Link]
262 S. Kumar et al.

Fig. 8 GUI of proposed system on CMU dataset, a test image, b face detection

FPR(MissRate)  1 − CDR

FN represents the false negative, and FP is false positive;

TP represents true positive, and TN is true negative.
According to Table 1, the accuracy or Correct Detection Rate of the proposed
method is much higher than that of reference system on BAO dataset. Experimental
results show that detection rate can reach approximate 90%, whereas Ref gives
84.2%. This methodology applies on CMU database also, and accuracy achieved
88.67% with fast computational time as shown in Table 2 (Figs. 8, 9, 10, and 11).

6 Conclusion

In this research, we focused on a multiple face detection technique which uses the Cor-
relogram, DWT, and EH for face recognition. The median filter is used for removing
the noise in given image. For MFD, the combination of DWT, EH and Correlogram
of a face is used as the feature vector and an SVM classifier is trained to classify
the input face as multiple or single. The proposed algorithm has been tested on the
standard BAO and CMU datasets, and the accuracy reached up to 90%. From the
experimentation, we noted that the SVM classifier recognizes the facial images more
accurately from the previous classifier. We got an average recognition accuracy of
89% for SVM.

[Link]@[Link]
Multiple Face Detection Using Hybrid Features … 263

Fig. 9 GUI of proposed system on CMU dataset, a test image, b face detection

Fig. 10 GUI of proposed system on BAO dataset, a test image, b face detection

[Link]@[Link]
264 S. Kumar et al.

Fig. 11 GUI of proposed system on BAO dataset, a test image, b face detection

References

1. Kang, S., Choi, B., Jo, D.: Faces detection method based on skin color modeling. J. Syst. Archit.
64, 100–109 (2016)
2. Kumar, S., Singh, S., Kumar, J.: A study on face recognition techniques with age and gender
classification. In: IEEE International Conference on Computing, Communication and Automa-
tion (ICCCA), 5, 6 May 2017
3. Kumar, S., Singh, S., Kumar, J.: A comparative study on face spoofing attacks. In: IEEE
International Conference on Computing, Communication and Automation (ICCCA), 5, 6 May
2017
4. Kumar, S., Singh, S., Kumar, J.: Automatic face detection using a genetic algorithm for various
challenges. Int. J. Sci. Res. Mod. Educ. 2(1), 197–203 (2017)
5. See, Y.C., Noor, N.M., Lai, A.C.: Hybrid face detection with skin segmentation and edge detec-
tion. In: 2013 IEEE International Conference on Signal and Image Processing Applications
(ICSIPA), pp. 406–411 (2013)
6. Shah, J.H., Sharif, M., Yasmin, M., Fernandes, S.L.: Facial expressions classification and false
label reduction using LDA and threefold SVM. Pattern Recognit. Lett. 1–17 (2017)
7. Zhou, X., Jin, K., Chen, Q., Xu, M., Shang, Y.: Multiple face tracking and recognition with
identity-specific localized metric learning. Pattern Recognit. 75, 41–50 (2018)
8. Mahajan, J.A., Paithane, A.N.: Face detection on distorted images by using quality HOG
features. In: 2017 International Conference on Inventive Communication and Computational
Technologies (ICICCT), pp. 439–444 (2017)
9. Nasr, S., Bouallegue, K., Shoaib, M., Mekki, H.: Face recognition system using a bag of
features and multi-class SVM for robot applications. In: 2017 International Conference on
Control, Automation and Diagnosis (ICCAD), pp. 263–268 (2017)
10. Tao, Q.Q., Zhan, S., Li, X.H., Kurihara, T.: Robust face detection using local CNN and SVM
based on kernel combination. Neurocomputing 211, 98–105 (2016)

[Link]@[Link]
Multiple Face Detection Using Hybrid Features … 265

11. Julina, J., Kulandai J., Sree Sharmila, T.: Facial recognition using a histogram of gradients and
support vector machines. In: IEEE International Conference on Computer, Communication
and Signal Processing (ICCCSP), pp. 1–5 (2017)
12. Naik, N., Rathna, G.N.: Robust Real-Time Face Recognition and Tracking on GPU Using a
Fusion of RGB and Depth Image. arXiv preprint. arXiv:1504.01883 (2015)
13. Yadav, S., Nain, N.: Fast face detection based on skin segmentation and facial features. In:
2015 11th International Conference on Signal-Image Technology & Internet-Based Systems
(SITIS), pp. 663–668. IEEE (2015)
14. Ramachandra, R., Yang, B., Raja, K.B., Busch, C.: A new perspective—face recognition with
the light-field camera. In: IEEE International Conference on Biometrics (ICB), pp. 1–8. IEEE
(2013)
15. Jensen, O.H.: Implementing the Viola-Jones face detection algorithm. Master’s thesis, Tech-
nical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark (2008)
16. Chaudhary, M.D., Upadhyay, A.B.: Integrating shape and edge histogram descriptor with sta-
tionary wavelet transform for effective content-based image retrieval. In: 2014 International
Conference on Circuit, Power and Computing Technologies (ICCPCT), pp. 1522–1527. IEEE
(2014)
17. Masselos, K., Andreopoulos, Y., Stouraitis, T.: Performance comparison of two-dimensional
discrete wavelet transforms computation schedules on a VLIW digital signal processor. IEE
Proc. Vis. Image Signal Process. 153(2), 173–180 (2006)
18. Smith, J.R., Chang, S.F.: Tools and techniques for color image retrieval. In: Storage and
Retrieval for Still Image and Video Databases IV, vol. 2670, pp. 426–438. International Society
for Optics and Photonics, Washington (1996)
19. Yang, S., Bebis, G., Chu, Y., Zhao, L.: Effective face recognition using a bag of features with
additive kernels. J. Electron. Imag. 25(1), 013025 (2016)
20. [Link]
21. [Link]
22. Huang, Z.C., Chan, P.P., Ng, W.W., Yeung, D.S.: Content-based image retrieval using color
moment and Gabor texture feature. In: 2010 International Conference on Machine Learning
and Cybernetics (ICMLC), vol. 2, pp. 719–724. IEEE (2010)

[Link]@[Link]
A Brief Overview on Generative
Adversarial Networks

Raj Patel

Abstract Generative models dwell into a research domain, which incorporates the
use of mathematical models by which machines could mimic and generate new data.
Generative models are used in tasks such as image or music generation as well as
in capturing and mimicking certain features of a dataset. One such generative model
is generative adversarial network, which makes use of a generator and discriminator
in order to perform such generative tasks. It does this by maintaining a min–max
relationship between the two models. This study gives a brief overview of generative
adversarial networks its optimization function and metrics as well as its implemen-
tation and working. The study also incorporates a case study, which provides a clear
illustration of the actual training, which is to be performed. The performance and the
results are sampled and very well compared. The study also gives insights regarding
the various applications and the variations of generative adversarial network.

Keywords Deep learning · Generative adversarial networks · GAN

Neural networks

1 Introduction

Advancement in deep learning and knowledge-based AI has facilitated computers

in performing complex tasks that were perceived only to be possible by humans.
Whether it would be image recognition, understanding natural language, speech
recognition or to learn and play games, computers are capable of performing such
tasks with the highest level of accuracy. This has been possible by incorporating
several algorithms and mathematical models. One such set of tasks is generative
tasks, in which new data is generated, similar to existing data. This generated data
can be used as generative output for generative tasks or in places having missing data,

R. Patel (B)
Department of Information Technology, Dwarkadas J. Sanghvi College of Engineering, Vile
Parle, Mumbai, India
e-mail: raj1471997@[Link]

© Springer Nature Singapore Pte Ltd. 2019 267

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
268 R. Patel

which may be useful for other machine learning tasks. The models that facilitate
such generative tasks are known as generative models. One such generative model
is generative adversarial network.
Ian Goodfellow introduced generative adversarial network or GAN in 2014 [1].
GANs work on a zero-sum game framework [2]. It consists of a generator and a
discriminator, and the generator performs the task of generating new items, whereas
the discriminator performs the task of distinguishing between real and fake data
samples. The data samples that generated by the generator would be fake since it
is trying to mimic the real data. The main goal of the generator is to make these
data items as real or as true to the original data items as possible in order to fool the
discriminator. On the other hand, the goal of the discriminator is to flag these fake
data items apart from the original ones, as best as possible. So here, the discriminator
works as an adversary or judge, judging the real and the fake data items.

1.1 Background

GANs were not the first to tackle the challenge in performing generative tasks, and
there have been several studies made over the years. Variational autoencoders (VAE)
and restricted Boltzmann machine (RBM) both provide models and optimization
techniques, which incorporated can perform generative tasks. VAEs are the prob-
abilistic graphical model whose explicit goal is latent modeling and marginalizing
out certain variables [3]. In the study (Pu [4]), the authors develop an autoencoder to
model images, as well as to associate labels or captions in a semi-supervised setting.
The study employs a convolutional neural network (CNN) inside the encoder for
the posterior distribution of the parameters of the image generative model [4]. RBM
works on the principle of unsupervised feature extractor, and in the process, it learns
the joint probability, which facilitates generation of new data. Ruslan and Geoffrey
(Salakhutdinov [5]) formulate a new algorithm with multiple layers of hidden ran-
dom variables known as deep Boltzmann machines to learn good generative models.
Both VAE and RBM use Markov chains for learning. GANs avoid the use of Markov
chains due to its high computational cost [6].
Section 2 of this study will discuss the optimization function and its metrics.
Sections 3 and 4 will dwell on the implementation and working of GAN with the
help of a case study. Section 5 will discuss its various applications and different
variations of the GAN, and at last, Sect. 6 will summarize and conclude the study.

2 Objective Function and Metrics

The optimization function for training GAN is represented using a set of two functions
G and D in a min–max relationship where G is the generative function and D is the
discriminative function. The relationship established is illustrated in Eq. (1).

[Link]@[Link]
A Brief Overview on Generative Adversarial Networks 269
   
min max V (D, G)  E x∼ pdata (x) log(D(x)) + E z∼ pdata (z) log(1 − D(G(z))) (1)
G D

• z represents the initial random noise or input given to the generator.

• x represents the actual data from the dataset to be fed to the discriminator.
• G(z) represents the output of the generator. G is a differentiable function, which
tries to map the input z based on its distribution pg .
• D(x) is the discriminating function which gives out a probabilistic value on how
likely the data item x represents or belongs from the actual dataset.
• D(G(z)) suggests the probability on how likely the output of the generator is
similar to the actual data. This value can be interpreted as the similarity index for
the generator.

The discriminator is trained in order to maximize the probability of assigning

the correct label to both training examples and the fake samples generated by the
generator [2]. Simultaneously, generator is trained in order to minimize log(1 −
D(G(z))) [1]. Error estimation of both the models is a form of log loss function.
Initially, when G is poor, D can reject samples with high confidence because they are
clearly different from the training data. In this case, log(1 − D(G(z))) saturates [1].
Rather than training G to minimize log(1 − D(G(z))), we can train G to maximize
log(D(G(z))). In other words, D and G play the following two-player minimax game
with value function V (G, D) [1]. Ideally, both error estimation of both the models
should converge to zero but in a minimax relationship, such convergence will result
in minimum possible loss value of either of the model.

3 Case Study: Introduction and Implementation

Working of GAN can be inferred with consideration of a case study. This case
study implements a deep convolutional GAN, which is a variation of GAN. Deep
convolutional GAN (DCGAN) consists of two primary networks: a deconvolutional
neural network as a generator for generating images and a convolutional neural
network as a discriminator for recognizing images. The output of the generator will
be an image, and the output of the discriminator will be single value suggesting
the probability of high likely the generated image from the generator is to the real
dataset.
The convolutional neural network or the discriminator will start with an input
layer to take in the input as an image. The deconvolutional neural network or the
generator takes in input as random noise. During the training, the weights of both the
networks are updated so that the generator gets better at deluding the discriminator
and the discriminator gets better at identifying fake data.

[Link]@[Link]
270 R. Patel

3.1 Dataset

The dataset used in the case study is the CelebFaces Attributes Dataset (CelebA) [7].
It is a large-scale face attributes dataset with more than 200 K celebrity images, each
with 40 attribute annotations. The images in this dataset cover large pose variations
and background clutter. From the exploratory visualization, different visual variations
are observed in the dataset. These variations range from:
1. Varying gender and age
2. Varying color complexity and ethnicity
3. Varying hairs and face positions.
These above-mentioned variations are clear in Fig. 1.

3.2 Architecture and Implementation

The architecture for the case study is based upon the one proposed in the DCGAN
paper (Radford [8]). The final implemented architecture is illustrated in Fig. 2. This
architecture will generate RGB images of size 64 * 64. The generator consists of an
input layer to take in the noise as an input the next follows a set of reshape layers and
convolutional layers with rectified linear nit (ReLu) activation and upsampling layers
in between. The final layer is a tanh activation layer. The upsampling layers increase
the height of the image and hence increasing the resolution after each convolutional.

Fig. 1 Image samples from the dataset

[Link]@[Link]
A Brief Overview on Generative Adversarial Networks 271

Fig. 2 Architectural implementation of DCGAN

The discriminator is quite reverse of that of the generator. It starts with an input
layer, which takes in an image of size 64 * 64 * 3; it then follows with a set of
convolutional layers with increasing number of filters. Between the convolutional
layers are leaky rectified linear unit (leaky ReLu) activation layers along with batch
normalization and dropout. The last layers are the flatten layer followed by the dense
layer. Batch normalization helps to prevent arbitrary large weights in the intermediate
layers as batch normalization normalizes the intermediate layers, thus helping to
converge well. Dropout layers are quite effective in preventing the discriminator
from overfitting during training. This forces the network from learning redundant
information.

3.3 Training and Refinement

The training process of DCGAN is as follows:

1. A random noise generator is used to generate random noise. This random noise
acts as input for the generator. On given input, the generator outputs an image
and this generated image is used to train the discriminator in distinguishing the
fake images from the real ones.
2. The generated image is given as input to the discriminator along with a real
image from the dataset. The label attached to the generated image is zero since it
was generated by the generator; hence, it does not follow any similarity with the
dataset, whereas the label attached to the image from the dataset is one since it is
the real image. The goal here is to train the discriminator to distinguish between
real and the fake image as best as possible. Training is performed by back-
propagation, by updating the weights of each layer based on the error function.
At this stage, weights of discriminator are only updated and this can be done by
setting generator’s trainable argument as false.
3. The stage that is left is to train the generator and for training the generator.
Similar to step two generators generates a new image by feeding it with random
noise. This generated image is again fed into the discriminator, which outputs the
probabilistic value of how likely the generated image is to the dataset. However,

[Link]@[Link]
272 R. Patel

for back-propagation, the label attached to the generated image is one instead of
zero. Since the goal here is to train the generator to generate images that look as
real as the original images in the dataset; thus, eventually the generator will learn
to generate images that look real. Hence, the label attached is one. For training
of generator, a feedback cycle facilitates updating weights of the network. The
discriminator is not trained during this step by setting its trainable argument as
false.
4. These three stages are performed indefinitely until the loss of each model con-
verges to the minimum. After each step, the generator becomes stronger at gen-
erating images identical to the dataset and the discriminator becomes stronger at
distinguishing these images. Moreover, in the end, two very strong models are
obtained which train itself by trying to outperform each other.
Several hyperparameters were tuned in order to obtain adequate results. The hyper-
parameters set for each model were as follows:
Hyperparameters set for the generator:
1. Kernel size: 4. (To set the filter size of convolutional layers to {4, 4} and to
perform fast convolutions on the image.)
2. Momentum: 0.8. (To avoid the problem of falling in local minima whenever one
of the models falls to its minimum value.)
Hyperparameters set for the discriminator:
1. Kernel size: 4. (To set the filter size of convolutional layers to {4, 4} and to
perform fast convolutions on the image.)
2. Strides: 2. (In order to jump two pixels at a time.)
3. Dropout: 0.25 between first two layers and 0.4 between the rests and thus pre-
vention from overfitting.

4 Case Study: Validation and Results

The training was conducted with a batch size of 32 samples. Initially, the learning
rate was kept at 0.02 and the model was tested for 10,000 steps/epochs. The mean
loss of generator and discriminator was found to be 2.4 and 3.43 after the training,
which was quite higher than expected. This was probably due to high learning rate.
Next, the learning rate was reduced for one-tenth of its value, which resulted in a
reduction of the losses.
In the final implementation, the model was trained for 100,000 steps/epochs with
a learning rate of 0.002 (Note: The training was done on a GPU-based workstation
resulting in total training time of 20 h). The losses of each model over the period
are illustrated in Fig. 3; the losses of generator and discriminator are plotted on the
Y -axis and the number of steps on X-axis.
Initially, the loss from the generator is quite high averaging around 1.13 until
20,000 steps but it lowers down as the model trains for more time. On the other hand,

[Link]@[Link]
A Brief Overview on Generative Adversarial Networks 273

Fig. 3 Loss of generator and discriminator over each step

the loss from the discriminator is quite low as compared to the generator. The above
figure illustrates the interesting relationship between generator and discriminator
in GAN. In which, whenever the loss of one model goes down the loss of other
model goes up. This is the min–max relationship, which is to be seen in GAN where
one network tries to outperform the other this is visible near 20,000 epochs where
discriminator loss is close to zero and generator loss is at its highest at around 14.
However, as the training continuous the two models begin to maintain an equilibrium,
which is clear after 40,000–60,000 epoch and after 70,000 epoch. The losses of the
generator and discriminator seem to be the same value, but having a closer inspection
still, the two models are still trying to balance each other where a decrease in one
model’s loss leads to increase in the other. This can be observed in Fig. 4.
The average loss of the generator during 100,000 epochs was 0.844, and the
average loss of discriminator was 0.707. The average accuracy of the discriminator
was 48%, which seems to be an appropriate value for a well-trained GAN.

4.1 Results from the Generator Overtime

The outputs obtained from the GAN can be summarized from Fig. 5. Over time, an
improvement in the quality of images generated by the generator can be concluded.
Visually comparing the generator is able to capture the variations that were observed
in the original dataset and reflect these variations in the produced images; a variety
of different images having various hair color, face complexity, background, and

[Link]@[Link]
274 R. Patel

Fig. 4 Zoomed in image of the graph between 95,000 and 100,000 steps

accessories. Figure 5 shows the results obtained over time starting from the first
epoch to each horizontal slice displaying the output after every 20,000 steps. Some
of the problems that can be seen in the images are:
(1) Blurriness caused in the images due to low resolution.
(2) A minor case of morphing seen in the images likely caused due to less training.

5 Types and Applications of GAN

The case study dwelled into the implementation of the DCGAN architecture which
is an essential architecture used for image generation. It yields good quality images
with proper training. Some of the other application of GANs are in music and video
generation, in image-to-image translation as shown by pix2pix and in text to image
translation. Apart from performing generative tasks, GANs can be used in tasks such
as anomaly or outlier detection. Quite often, this type of tasks requires huge number
of data to make proper distinctions, which in most cases is not available. GANs
solve this problem by making use of the generator in generating new data by which
the discriminator could take advantage in training. GAN can be used in datasets for
outlier detection or in fraud detection. It has also been used in drug discovery and in
cancer tissue detection.
Currently, GANs are somewhat unstable and proper architecture design is essential
for convergence due to which the true potential is not always achieved. However,
these issues are been tackled by its certain variations. Some of the types of GAN are:

[Link]@[Link]
A Brief Overview on Generative Adversarial Networks 275

Fig. 5 Outputs obtained from the DCGAN

• Wasserstein GAN
Wasserstein GAN is an alternative to GAN training [9]. Without proper design,
instability can cause inadequate training resulting in poor results. Wasserstein GAN
tries to eliminate this instability. During training, the loss function is changed to
include Wasserstein distance and by which they get rid of several problems such as
mode collapse and help to improve the stability of the model. The discriminator, in
this case, is trained until convergence, which results in mitigating the need to balance
the generator, hence improving the stability.
• Conditional GAN
In traditional GAN, there is no control on the modes of data being generated.
Conditional GAN solves these problems by conditioning the model with extra infor-
mation in order to direct the generation process [10]. Conditional GAN is an extension
of GAN framework, in which the generator is provided with some additional infor-
mation or meta-data along with random noise regarding the expected outputs to be
generated by the generator. Some of the keen applications of conditional GAN are
in generating images from text annotations given as inputs or in the generation of
tags from a given image .

[Link]@[Link]
276 R. Patel

• InfoGAN
InfoGAN is a modification to GAN objective, which makes it learn interpretable
and meaningful representations [11]. It does this by encoding meaningful features
along with the input noise. Conceptwise it looks quite similar to conditional GANs
but the additional information is facilitated within the noise itself. Additionally, it
maximizes the mutual information between a small subset of the latent variables and
the observation [11]. Applications of InfoGAN can range from tasks that require
manipulation of latent codes if InfoGAN can learn the detangled and interpretable
representations.

6 Conclusion and Future Scope

Generative adversarial networks truly exemplify the notion of using computers in

performing generative tasks. They provide a technique through which generative
models are improved as they advance their training, which results in higher quality
of results. The case study demonstrated the intricacy of training a GAN; it provided
insight regarding the importance of convergence in an adversarial setting. The results
of the case study justify the use of GAN for generating images and depict how the
proper usage of GAN. Several variations of GAN were looked upon, and their specific
applications were illustrated.
In the future, as more research is done in the subject matter GANs will surely be
the most appealing state of the art in deep learning. With its growing popularity in
the corporate sector, it will further be used in several tasks from music and video
generation to creating simulations which could imitate human behavior. The future
is bright, and the use cases of GAN are endless.

References

1. Goodfellow, I.J., Pouget-Abadie, J., Warde-Farley, D.: Generative Adversarial Networks,

pp. 1–6. Arxiv (2014)
2. Goodfellow, I.: NIPS 2016 Tutorial: Generative Adversarial Networks, pp. 6–13. Arxiv (2016)
3. Welling, M., Kingma, D.P.: Auto-Encoding Variational Bayes, pp. 1–8. Arxiv (2013)
4. Gan, Z., Henao, R., Yuan, X., Li, C., Stevens, A., Carin, L., Pu, Y.: Variational autoencoder for
deep learning of images, labels and captions. In: NIPS (2016)
5. Hinton, G., Salakhutdinov, R.: Deep Boltzmann Machines. MLR (2009)
6. Karazeev, A.: Generative Adversarial Networks (GANs): Engine and Applications, 17-
8-2017. Available: [Link]
applications-f96291965b47 (Online)
7. Luo, P., Wang, X., Tang, X., Liu, Z.: Deep learning face attributes in the wild. In: Proceedings
of International Conference on Computer Vision (ICCV) (2015)
8. Metz, L., Chintala, S., Radford, A.: Unsupervised Representation Learning with Deep Convo-
lutional Generative Adversarial Networks, pp. 1–7. Arxiv (2015)
9. Chintala, S., Bottou, L., Arjovsky, M.: Wasserstein GAN, pp. 1–5. Arxiv (2017)

[Link]@[Link]
A Brief Overview on Generative Adversarial Networks 277

10. Osindero, S., Mirza, M.: Conditional Generative Adversarial Nets, pp. 1–5. Arxiv (2014)
11. Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., Abbeel, P., Chen, X.: InfoGAN: Inter-
pretable Representation Learning by Information Maximizing Generative Adversarial NetsIn-
foGAN: Interpretable Representation Learning by Information Maximizing Generative Adver-
sarial Nets, pp. 1–12. Arxiv (2016)

[Link]@[Link]
Prediction of PhotoVoltaic Power
Generation Using Monte Carlo
Simulation

Gautam Seth, K. Abhay Prithvi, Arpit Paruthi, Saksham Jain

and Umang Soni

Abstract In this paper, Monte Carlo simulation models have been used to forecast
the expected amount of energy production from photovoltaic panels. Two MC models
have been proposed with traditional Monte Carlo method as their backbone. Model-1
combines the locally weighted scatterplot smoothing (LOESS) with simple Monte
Carlo simulation. Model-2 is derived from the first model, and weather forecast
data is used as an exogenous input. Also, the effect of weather is considered on
traditional MC simulation and compared with Model-2. Using Model-2, the error
between generated and predicted data is found to be the least. Results have been
generated using Python and are discussed with inference in the manuscript.

Keywords Monte Carlo simulation · Solar energy · Photovoltaic panels

Locally weighted scatterplot smoothing

1 Introduction

Owing to the fact that non-renewable energy sources are depleting day by day, there
has been a sudden shift toward the alternative energy sources like the solar and wind
energy. Events within the past decade have shown that every country is shifting toward

G. Seth · K. A. Prithvi · A. Paruthi · S. Jain (B) · U. Soni

Netaji Subhas Institute of Technology, Delhi, India
e-mail: jain.saksham01@[Link]
G. Seth
e-mail: gautamseth21@[Link]
K. A. Prithvi
e-mail: abhayprithvi97@[Link]
A. Paruthi
e-mail: paruthiarpit@[Link]
U. Soni
e-mail: [Link]@[Link]

© Springer Nature Singapore Pte Ltd. 2019 279

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
280 G. Seth et al.

the use of renewable sources of energy to protect the environment. However, there
has been a rapid growth in the industrial and domestic consumption of electricity.
According to the data released by the government of India, about 418,347 GWh
energy is used by the industrial sector, which contributes to 44% of the total electricity
consumed whereas about 217,405 GWh energy is consumed by the domestic sector,
which contributes to 23% of the total electricity consumed. Since the past decade,
this rate has been increasing at 5–8% per annum [1]. India, being one of the largest
producers of energy, in its basket has a mix of all the sources available including the
renewable sources. The total energy produced in India is about 1,433,392 GWh out
of which about 80% is produced in thermal and hydro plants, and only 5.7% of the
energy is produced using the renewable sources [2]. Looking at the above-mentioned
facts, it can be said that the current energy sector of India is heavily dependent on coal
and fossil fuels. This arouses a concern on how to minimize the usage of fossil fuels
and save the resources for the future generations [3]. These alternative sources require
a huge investment but may lead to very low benefit. Hence, making it important for
the nations to collect and analyze the data to forecast how much a plant operating
on solar or wind energy will be able to generate. Power generation from renewable
sources depends on a number of factors. In the case of solar, factors like irradiation
and weather (includes cloud cover along with other conditions like fog, and dust)
play an important role in the prediction.
A lot of research has been done in predicting the data using ANN and Fuzzy
logic, which can more accurately simulate the system. But the usage of ANN and
Fuzzy logic requires higher computing power and takes longer period of time to
simulate the system [4, 5]. Monte Carlo simulation, on the other hand, requires
lower computational power but the accuracy decreases. Therefore, the two strategies
are proposed to improve the accuracy of the traditional Monte Carlo simulation while
keeping the computational requirements at a minimum.
The rest of the paper is organized as follows: Methodology in Sect. 2, Results in
Sect. 3 and Conclusion in Sect. 4.

2 Methodology

With the advent of powerful computers, the working of a plant/system can be simu-
lated to give its working in real time. Simulation of any system is a powerful tool. It
can be used for analysis of new designs and comparing performance of new designs
to existing designs. But a simulation model of a complex system can only be an
approximation to the actual system [6].

[Link]@[Link]
Prediction of PhotoVoltaic Power Generation … 281

2.1 Monte Carlo Simulation

Monte Carlo simulation is a powerful computerized mathematical technique which

is used to model and analyze real-world systems which employ statistical sampling
which approximates solutions to quantitative problems. It investigates stochastic
permutations of the uncertainty of a system and also quantifies their effect. It uses
probability distribution of possible outcomes of a system to randomly select a variable
and calculate value of the uncertainty through a number of runs [7] (Fig. 1).
To perform Monte Carlo simulation, the first step is to determine the probability
distribution for uncertainty based on history as well as expert opinion. Some of the
commonly used probability distributions are:
• Normal or bell distribution
• Lognormal distribution
• Uniform distribution
• Triangular distribution.

2.2 Dataset

The solar irradiance data was obtained from National Solar Radiation Database
(NSRDB) website [8]. It consisted of hourly GHI values for our required location
coordinates (Delhi, India in the given case). The GHI values were then converted to
daily PV generation values using the formula as stated below [9].

E  A ∗ r ∗ Ir ∗ Pr

Fig. 1 Normal distribution

about a non-zero mean

[Link]@[Link]
282 G. Seth et al.

Here, E is the energy generated in kWh, A is the area of PV panels, r is the yield
of panel equal to the ratio of electrical power of one PV panel to area of one panel,
I r is irradiation on the panels, and Pr is the performance ratio which ranges between
0.5 and 0.9 depending on the grade of PV panel (Table 1).
We trained the model using this data for the year 2000–2010. Delhi’s weather
data was also obtained which consisted of hourly weather data. Using weather con-
dition data, approximate cloud cover was calculated. The cloud cover was used as
an exogenous input for the solar PV generation and will be discussed in detail in the
following sections.

2.3 Problem with MC Simulation and Proposed Model

One of the problems faced while using Monte Carlo simulation was its inability to
clearly tell how the uncertainty variable changes. In the case of solar PV generation,
it can be easily observed that, for every year, it follows an approximately fixed trend.
Power generated is higher during the summer months (May, June), and lower during
the winter months (December, January). Therefore, Monte Carlo simulation could
give a good overall estimation for the annual PV generation but was not able to give
a satisfactory prediction for monthly or daily values (Fig. 2).
To overcome this problem, an approach can be proposed that instead of obtaining
the random variable from the probability distribution about a fixed mean (or central
tendency), obtain the random variable using the probability distribution about a mean
which is a variable of an external input. In this case, it was assumed that the mean is

Table 1 Solar panel specification

Area of each panel No. of panels Performance ratio Yield of a panel
20 m2 100 0.75 16.16%

Fig. 2 Confidence interval

as a function of x for normal
distribution about a non-zero
mean

[Link]@[Link]
Prediction of PhotoVoltaic Power Generation … 283

a variable of time (day number, from 1 to 365), keeping in mind the trends observed
in the data.
To obtain the relation between the mean PV generation and time, locally weighted
scatterplot smoothing (LOESS) [10, 11] was used in the proposed model.
Using the relation between PV generation values and time, mean PV genera-
tion value was calculated and Monte Carlo simulations consisting of 10 runs were
performed for each day of the year, by selecting the random variable from the proba-
bility distribution about the calculated mean. The proposed model could predict both
annual PV generation and monthly PV generation with higher degree of accuracy in
comparison with the traditional Monte Carlo simulation.

2.4 Proposed Model Along with Weather Condition

as an Exogenous Input Variable

The solar radiation energy reaching the earth’s surface is the sum of the energy
directly transmitted from the sun and the radiation energy diffused by the sky. When
the sky is clear, directly transmitted radiation affects the total solar radiation while on
a cloudy day, scattered radiation plays the major role. During a partly cloudy day, the
direct radiation’s effect decreases while the scattered radiation’s effect increases [12].
Hence, it can be said that the PV generation depends highly on weather conditions
(hence they have high correlation). Based on the octants of the sky covered by opaque
clouds, the weather conditions can be classified into 9 categories, from 0 to 8, where
0 means 0% cloud coverage and 8 means 100% cloud coverage [13] (Fig. 3; Table 2).

There is a simple relation between the fraction of cloud cover and solar radiation
reaching the earth’s surface:
 
Pα 1 − 0.75F 3 watts/m2

where F = Fraction of cloud cover on a scale of 0.0 (no clouds) to 1.0 (Overcast)
and P is the power generated [14]. Using the above expression, we can obtain the

Fig. 3 Flow chart showing weather as an exogenous input

[Link]@[Link]
284 G. Seth et al.

Table 2 Fraction of cloud Cloud cover Weather conditions

cover based on weather
conditions 0/8 Clear/sunny
1/8 Mostly clear
2/8 Partial fog/partly cloudy
3/8 Light drizzle/light fog
4/8 Shallow fog/light rain
5/8 Light freezing rain
6/8 Haze/light thunder storm
7/8 Heavy fog/thunder storm
8/8 Heavy rain

solar PV generation based on cloud cover as well as the value of PV generation

obtained from the model explained in the previous section. It must be noted that the
value of the mean PV generation in the previous model must be first converted to the
maximum possible power, and then be multiplied by the factor (1–0.75F 3 ), and then
the Monte Carlo runs can be performed. This model further increases the accuracy
of prediction.

3 Results

As stated above, traditional Monte Carlo simulation generates the probability distri-
bution about a random variable can be seen in Fig. 4. Here the predicted values are
nearly constant and do not follow the actual values. This results in monthly predicted
also being approximately constant. This leads to a maximum error of about 46% in
the month of December. The total annual error is about 1.7% (Fig. 5).

Fig. 4 Daily prediction

using trad. MC simulation

[Link]@[Link]
Prediction of PhotoVoltaic Power Generation … 285

Fig. 5 Monthly prediction

using trad. MC simulation

To reduce the error, the random variable is taken about the daily generated values.
This reduces the maximum monthly error from 46% to approximately 6%. Further,
the daily predicted values now follow the actual values, and hence monthly data is
accurately predicted as shown in Figs. 6 and 7.
Table 3 shows the monthly comparison of traditional MC simulation to the pro-
posed Model-1.
Since solar power generation depends on the weather conditions, if the weather
conditions are used as an exogenous input, the accuracy can be further increased. The
weather conditions are converted into cloud cover code as stated earlier. To compare
our proposed model, we find the effects of weather on traditional MC simulation.
The daily and monthly predicted values of Traditional + Weather are shown in Figs. 8
and 9. The outputs of Model-2 are shown in Figs. 10 and 11.

Fig. 6 Daily prediction

using Model-1

[Link]@[Link]
286 G. Seth et al.

Fig. 7 Monthly prediction

using Model-1

Table 3 The monthly comparison of traditional MC simulation to the proposed Model-1

Month Actual MC % error Model-1 % error
generated estimation estimation
(kWh) (kWh) (kWh)
Jan 3060.93 4258.46 39.12 2957.57 3.37
Feb 3365.08 3898.90 15.86 3566.27 5.97
March 4832.60 4305.3 10.91 4919.06 1.78
April 5437.38 4167.1 23.36 5369.64 1.24
May 5678.69 4338.7 23.59 5537.08 2.49
June 4975.20 4157.2 16.44 4943.17 0.64
July 4784.38 4302.7 10.06 4688.60 2.00
Aug 4588.47 4307.9 6.11 4590.30 0.03
Sept 4499.64 4149.5 7.78 4350.85 3.30
Oct 4224.58 4344.2 2.83 4075.98 3.51
Nov 3260.74 4189.8 28.49 3296.75 1.10
Dec 2958.56 4323.8 46.14 2833.58 4.22
Total 51,666.3 50,743.56 1.78 51,128.91 1.04

Table 4 compares the effect of exogenous input of weather on both traditional MC

as well as Model-1.

4 Conclusion

The main objective of this study is to compare the two strategies, which are used to
obtain the forecast of the solar power generation. This study can help us to simulate a
solar power plant model to estimate the amount of solar energy that can be generated

[Link]@[Link]
Prediction of PhotoVoltaic Power Generation … 287

Fig. 8 Daily prediction

using Traditional MC +
Weather

Fig. 9 Monthly prediction

using Traditional MC +
Weather

Fig. 10 Daily prediction

using Model-2

[Link]@[Link]
288 G. Seth et al.

Fig. 11 Monthly prediction

using Model-2

Table 4 Comparison of the effect of exogenous input of weather on both traditional MC as well
as Model-1
Month Actual Traditional + % error Model-2 % error
generated weather estimation
(kWh) (kWh) (kWh)
Jan 3060.93 4707.56 53.79 3274.77 6.98
Feb 3365.08 4239.34 25.98 3858.40 14.65
March 4832.60 4161.15 13.89 4747.75 1.75
April 5437.38 3812.35 29.88 4874.15 10.35
May 5678.69 3895.26 31.40 4979.72 12.30
June 4975.20 4063.54 18.32 4759.18 4.34
July 4784.38 4265.89 10.83 4625.50 3.32
Aug 4588.47 4519.99 1.49 4752.71 3.57
Sept 4499.64 4351.88 3.28 4522.15 0.50
Oct 4224.58 4537.37 7.40 4297.01 1.71
Nov 3260.74 4699.34 44.11 3748.27 14.95
Dec 2958.56 4861.12 64.30 3188.21 7.76
Total 51,666.3 52,114.86 0.86 51,627.86 0.07

annually, which can help in making decisions based on the demand of electricity, so
that the remainder of the required amount that has to be obtained from the electricity
grid can be estimated. Previously, a lot of research has been done in predicting the data
using ANN and Fuzzy logic. There are some exceptions to Monte Carlo simulation as
the use of more complex techniques is limited, therefore only a few have based their
study on predicting the solar power generation using the Monte Carlo simulation
method.
The results that are obtained from the study show that there is a slight deviation
from the actual power generated. The comparison of traditional MC and Model-1

[Link]@[Link]
Prediction of PhotoVoltaic Power Generation … 289

shows that by taking the random variable about a variable point (variable of time),
the accuracy of prediction increases exponentially. This is because we are able to
replicate the trends that are followed by the actual model throughout the year using
a combination of LOESS and MC simulation. Studying the effects of weather as
an exogenous input on both traditional MC and Model-1, we see that, even though
the accuracy of Traditional MC increases, but still, it was less than Model-1. On
comparing Model-1 and Model-2, we see that the annual prediction using Model-2
was much closer to the actual power generated. The annual error using Model-1 was
calculated as 1.04%, whereas Model-2 performed with an accuracy of 0.07%. But
the monthly prediction was better in the case of Model-1 with maximum error as
approx. 6% as compared to the maximum error of 14.65% in the case of Model-2.
As the further result, a better model can be generated with the help of weather
information like wind speed and direction, and dew point. Also, ANN and Fuzzy
logic seem to be quite promising approach if historical data along with the various
weather information are present to generate more effective results.

References

1. ENERGY—Statistical Year Book India. [Link]

185 (2017)
2. Growth of Electricity Sector in India. [Link]
growth_2017.pdf
3. Overview of Renewable Energy Resources of India. [Link]
[Link]?aid=41988
4. Kim, J.G., Kim, D.H., Yoo, W.S., Lee, J.Y., Kim, Y.B.: Daily prediction of solar power genera-
tion based on weather forecast information in Korea. IET Renew. Power Gener. 11, 1268–1273
(2017)
5. Sivaneasan, B., Yu, C.Y., Goh, K.P.: Solar forecasting using ANN with fuzzy logic pre-
processing. Energy Procedia 143, 727–732 (2017)
6. Tüzüntürk, S., Eren Şenaras, A., Sezen, K.: Forecasting Water Demand by Using Monte Carlo
Simulation (2015)
7. Paul, V.K., Basu. C.: A Handbook for Construction Project Planning and Scheduling. Copal
Publishing Group (2017)
8. National Solar Radiation Database. [Link]
9. Calculations of Solar Energy Output. [Link]
OF_SOLAR_ENERGY_OUTPUT
10. Cleveland, W.S.: Robust locally weighted regression and smoothing scatterplots. J. Am. Stat.
Assoc. 829–836 (1979)
11. Cleveland, W.S.: Locally weighted regression: an approach to regression analysis by local
fitting. J. Am. Stat. Assoc. 596–610 (1988)
12. Matuszko, D.: Influence of the extent and genera of cloud cover on solar radiation intensity.
Int. J. Climatol. 32, 2403–2414
13. NOAA’s National Weather Service—Glossary. [Link]
word=sky%20condition
14. Cloud Cover Solar Radiation. [Link]
[Link]

[Link]@[Link]
Storage and Analysis of Synchrophasor
Data for Event Detection in Indian Power
System Using Hadoop Ecosystem

Akhilendra Pratap Singh, G. Hemant Kumar, Subhendu Sekhar Paik

and Diptendu Sinha Roy

Abstract Synchrophasor devices commonly referred to as phasor measurement

units (PMUs) have been rapidly deployed in power grids to get a clearer picture
of the events that take place in power grid at very high sampling rates. Thus PMUs
enable event predictions to become more accurate. However, high data-sampling rate
of PMUs creates huge volumes of data on a GB scale per PMU per day, which in
turn makes the storage and analysis of the collected data difficult; giving rise to a
Big Data problem use case. In India, National Load Dispatch Centre (NLDC) uses
Synchrowave; a centralized software package to collect, store and process, which
runs on a single very high-end system for storage and processing. But the centralized
architecture of current system gives no room for scalability, neither in computation
nor storage, even with costly hardware upgrades. This paper proposes a system model
with all the desired features using open-source frameworks and tools like Apache
Hadoop and HBase.

Keywords Phasor measurement unit (PMU) · Synchrowave · Hadoop · HBase

1 Introduction

Electrical power system for generation, transmission and distribution is extremely

valuable to modern society. It is a difficult task for power sector engineers to keep a
sharp vigil on power grid to keep it stable and reliable manually. Automated monitor-
ing and alerting mechanism have been into place for quite sometime with supervisory
control and data acquisition (SCADA), but time resolution for SCADA for detection
and identification of various events in massively large-scale power system that are
interconnected through a wide area measurement system (WAMS). Because of the

A. P. Singh · S. S. Paik · D. Sinha Roy (B)

National Institute of Technology Meghalaya, Shillong, India
e-mail: [Link]@[Link]
G. Hemant Kumar
ICTS-TIFR, Bengaluru, India

© Springer Nature Singapore Pte Ltd. 2019 291

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
292 A. P. Singh et al.

complexity, the nature of power system dynamically changes randomly in real time
due to different types of disturbance events such as generation loss, line trips (line-
ground, lineline, etc.), faults (transition, persistence, symmetric, Asymmetric, etc.)
and blackouts. A wide range of disturbance events in power system like blackouts,
line tripping, faults and generation loss are far from the reliability and stability of
requirements as they suppose to be. The inaccuracy in analysis of disturbances at
different regions of the interconnected smart grid is due to lack of wide-area infras-
tructures [1].
In many distribution systems primarily due to economic constraints, data acquisi-
tion is impeded heavily. However, with the Power Grid Corporation of India Limited’s
(PGCIL) initiative of synchrophasors that has been eradicated [2]. Along with issues
in the current observability of the distribution system, the introductions of newer tech-
nologies have resulted in the potential for previously insignificant or non-existing
faults and disturbances to occur. The newer grid technology introduced phasor mea-
surement units (PMUs) for continuous monitoring of transmission bus lines and
recording the continuous changes in electrical parameters. Those readings are called
PMU samples of electrical parameter changes. It helps to analyze the power grid
status. Power engineers can detect, analyze and find solutions for electrical power
faults, blackouts and grid isolation etc from PMU data, owing to it’s high resolution.
PMUs can be thought as descendants of computer relays that employed micropro-
cessors for accounting for power system parameters in earlier days. Computer relays
became very popular, and several reliability studies [3, 4], confirmed their robust
nature. At the same time, information and communication technology (ICT)-based
smart services such as smart energy management for homes [5] and grid computing-
based power monitoring systems [6] were also becoming prominent eventually lead-
ing to the vision of smart grid. PMUs as smart grid sensors became a very popular
topic and several attempts for PMU design [7] were found along with research per-
taining to their utility as effective smart grid sensors [8]. Reliability analysis of
PMUs has attracted a lot of research attention accounting for different aspects and
with various mathematical tools [8–12].
The objective of this paper is to develop an operative event detection model
by using the methodology of statistical analyzing on PMU datasets stored. With
help of this model, the grid control room engineers present at load despatch centers
can monitor and visualize the continuous changes in power grid states (i.e., grid
parameters like voltage, current and frequency) which are helpful for early detection
of power grid faults/events. Early event detection is much crucial for choosing control
operations to overcome the isolated zone, energy loss, blackouts and power outage.
This paper proposes the design of needed system model by implementing dis-
tributed, non-relational database—HBase—on a distributed storage system—HDF-
S—for structured data with back-end capabilities enabling numbers of features like
linear and modular scalability, strictly consistent reads and writes, automatic and
configurable shading of tables and automatic fail-over support, thus making the pro-
cess of data analysis much simpler, robust and reliable. Moreover, this paper brings
forth an efficient event detection algorithm using MapReduce framework of Hadoop
using PMU datasets.

[Link]@[Link]
Storage and Analysis of Synchrophasor Data for Event Detection … 293

The rest of paper organization as follows: Sect. 2 gives a brief introduction about
the background of the problem and the literature review of different event detec-
tion clarification-related works and a brief background of Indian Power Grid PMU
projects, grid structure, Synchrowave. Section 3 depicts proposed MapReduce-based
data analysis model for power system fault detection with non-relational decentral-
ized storage of PMU data. Section 4 presents implementation, platform setup and
details of experimental work done on real PMU datasets collected from NLDC.
Section 5 concludes the proposed idea with extensive future work.

2 Background and State of Art

The entire Indian Power Grid had been divided into five regions as North Regional
Load Despatch Centre (NRLDC), North East Regional Load Despatch Centre
(NERLDC), East Regional Load Despatch Centre (ERLDC), South Regional Load
Despatch Centre (SRLDC), West Regional Load Despatch Centre (WRLDC). Each
region controls and monitors the substation’s power generation unit within that region
only. The all five interdependent regions are centrally under control of National Load
Dispatch Centre (NLDC) located at New Delhi [13]. NLDC is using Synchrowave
Central software package for recoding all 52 PMUs readings. A phasor measurement
unit (PMU) or a synchrophasor is a device which measures the electrical waves on
an electricity grid, using a common time source for synchronization help of GPS
system facility as the block diagram shown in Fig. 1b. Time synchronization allows
synchronized real-time measurements of multiple remote measurement points on the
grid. In power engineering, these are considered as one of the most important measur-
ing devices in the future of power grid systems. Synchrowave Central software was
configured by Synchrowave Central software provider on central server at NLDC
central power control or control room. Currently, 52 PMUs were located in Indian
grid for electrical dynamic properties monitoring and recording. The entered PMU
data samples are in the format shown in Fig. 2. These are collected and transferred
to the regional phasor data collector (PDC) located at regional load despatch centres
(RLDC) and from there the samples are transferred to the central PDC located at
NLDC [13, 14] as shown in Fig. 1a.
Monitoring and analysis of signals of power system parameters like voltage, volt-
age angle, current, current angle and frequency are typical time series data. A lot of
research work was going on in the direction of improving the process of analysis of
large time series data sample with a fast reliable model of storage architecture. If it is

[Link]@[Link]
294 A. P. Singh et al.

Fig. 1 a Indian power system architecture of PMUs; b Synchrophasor architecture for Synchrowave
Central software

[Link]@[Link]
Storage and Analysis of Synchrophasor Data for Event Detection … 295

Fig. 2 PMU raw data sample recorded by Synchrowave

examined thoroughly, then all such research works can be grouped into three classes
according to their processing and storage frameworks.
The synchrophasor data is being extensively used for real-time and post-event
analysis, moreover, event detection, event analysis and for the purpose of devel-
opment of analytics and tools to monitor and control the event. A lot of research
contribution had been given by many researchers working all across the globe for
fault detection and fault classification, to develop smart auto-alarming control system
models. To fulfill the needs of the modern age, many types of research and studies
are going and some have already developed a smart power grid system. PMU sample
data analyzed through any complex mathematical optimization model to extract the
features of the data and map them with known disturbance events. The models that
can identify disturbance directly from the change in frequency range are designed by
Gaussian distribution method [15]. A few approaches to identify events by using some
mathematical models are proposed to apply convex optimization method known as
ellipsoid model by Dahal et al. [16]. This method is a suitable model for smart grid
dynamic behavior identification is well known as minimum volume enclosing ellip-
soid (MVEE) [17]. In this model, all the PMU measurement samples are enclosed

[Link]@[Link]
296 A. P. Singh et al.

inside an ellipsoid of minimum volume in n-dimensional space where the number

of PMU. The ellipsoid properties like volume, change in volume, center, and semi-
major axis are used to classify the power system events and clustering those observed
events by applying hierarchical clustering approach, support vector machine (SVM),
K nearest neighbor classifier, recursive partitioning and regression tree (RPRT) [14,
18]. Additionally, numerous statistical data analysis models had been developed that
employed window-based pattern recognition for tracking power system events [19,
20] and employed experience-based threshold settings for event detection [21].

3 Proposed System Model

Generally, PMUs are located at sub-grid stations working at voltage level of 110 V
or above. It is such a sensitive hardware device that it can record all types of minor
and major changes observed in power properties caused due to any reason like LL
tripping, LG faults and due to fall of tree upon transmission line are observed at load
transmission lines and grid stations.
Therefore power properties are dynamic parameters. Any dynamic change of
properties result in an alerting generation at control rooms, and this is required to be
as soon as possible by the scheme. In this paper, PUM datasets that were collected
from NLDC historian storage (as discussed in Sect. 2), and NLDC historian stores
a total of 52 PMUs sample records. PMUs are installed over 132, 220 and 400 kV
very high voltage transmission bus lines. At each transmission, bus line data was
recorded at a sampling rate of 25 samples per second [14–18, 22, 23].
In this proposed designed model, a subset of major power properties like voltage,
current, frequency and phasor angle of PMU data sample is considered. All the
data are continuously read from the historian to a common sink folder then from a
common sink folder to HBase table as Fig. 3a.
In order to handle this architecture given in Fig. 3a, the primary handlers proposed
in this paper are Apache Hadoop, Apache HBase and Apache Thrift. Apache Hadoop
provides both parallel processing framework, MapReduce (MR) and distributed file
system, Hadoop Distributed File System (HDFS) facility to process and store a huge
amount of data in decentralized clusters. The clusters can be geographically dis-
tributed across different territories. Hadoop cluster supports commodity hardware
processing units. Hadoop provides two major frameworks such as [24]. Apache
HBase is an open source, distributed non-relational database modeled after Google’s
Big-table. It allows distributed storage system for structured data with back-end capa-
bilities on top of Hadoop and HDFS. It has numerous numbers of features like linear
and modular scalability, strictly consistent reads and writes, automatic and config-
urable shading of tables, automatic fail-over support between RegionServers and
convenient base classes for backing Hadoop MapReduce jobs with Apache HBase
tables. The Apache Thrift is a popular software framework, mostly used for scalable
cross-language service development, which integrates a fully functional software
stack with a core engine to build and provide efficient services and platforms for

[Link]@[Link]
Storage and Analysis of Synchrophasor Data for Event Detection … 297

Fig. 3 a Block diagram of primary power distribution substation: [19]; b GRID-PROPERTIES

attributes structure

seamlessly various programming languages. Instead of programming a huge lines of

code to transport and serialize the objects and invoke remote methods, it helps to get
right down to business logics [25].
GRID-PROPERTIES and PMU-PROPERTIES, two logical HBase tables, have
been designed to store and manage PMU data in HDFS of Apache Hadoop. The
proposed idea of this paper operates in two phases. One sample of has GRID-
PROPERTIES table has been shown in Fig. 3b.
Sink data from historian to HBase grid table: This one is designed to read data
from OS file system (i.e., .csv file) and put those records of data into HBase table.
We designed a normal read-write program by using HBase Client API. This program
just scans the sink folder for .csv then reads the .csv file row by row and inserts (i.e.,
put) into the GRID-PROPERTIES table (Fig. 1a). MapReduce version of statistical
trend analysis program for PMU datasets HBase tables.
MapReduce algorithm for characteristic extraction: This one is designed to scan
the GRID-PROPERTIES (Fig. 4) frame by frame. The technique of overlapping,
moving window approach is used to scan records from Fig. 5. The window size is 12 s
long (i.e., 12 * 25  300 sample records) and overlapping half window of previously
scanned window data (i.e., 6 * 25  150 old sample records). Each time the program
gets 300 sample records of a PMU and performs MapReduce computation to find
out the statistical characteristics of a PMU data trend. It computes mean, maximum,

[Link]@[Link]
298 A. P. Singh et al.

Fig. 4 a Stage one process flow for effective values evaluation using MapReduce; b Stage two
process flow to detect, plot and store events

minimum and standard deviation for each PMU present within the window frame.
The results are getting stored into another HBase table PMU-DATA-TREND, and
the process workflow can be represented as Fig. 4a, b.

4 Implementation and Experimentation

4.1 TestBed Setup

The proposed idea has been realized in a cluster of six homogeneous nodes with
1 TB SATA Hard Disk, 8 GB DDR3 RAM, Intel Core i3-4790 CPU @ 3.60 GHz
8 processor, Ubuntu 16.04 LTS operating system, Apache Hadoop V2.7.1 Apache
HBase V1.4.3. For programming, statistical analysis and result visualization R Studio
version 1.1.447 has been configured and Apache thrift V0.11.0. Apache thrift acts
as the gateway to HBase server and access HBase table data in R code. A package
named rhbase provides necessary method to connect and access HBase table in R.
HBase has been configured on HDFS to use its file system [24–26].

[Link]@[Link]
Storage and Analysis of Synchrophasor Data for Event Detection … 299

Fig. 5 a HBase web UI; b List of HBase tables and GRID-PROPERTIES table details; c GRID-
PROPERTIES table records

[Link]@[Link]
300 A. P. Singh et al.

4.2 Analysis and Observation

The graph for Max–Min voltage gets data from the records of the table PUM-
DATATREND and R Studio IDE was used to plot the graphs as shown in Fig. 5.8.
Here, red bots taken for Max voltage and green dots are taken for Min voltage. Win-
dow frame counts are taken along x-axis and voltages along y-axis. From the graph,
we can observe that at some places difference between Max and Min voltages are
likely more. To analyze more on such observations, another graph was plotted by
taking frame count along x-axis and voltage difference along y-axis as shown in
Fig. 6a. Here one can get clearer and many pick values. Those pick differences may
be due to regular electrical faults that are natural in power grid system, and sometime
such pick difference may be courage event also.
To analyze in clear and in very sensitive form, here, Max–Min method is used to
find the SD of framewise voltage difference and mean. According to the reference of
Max–Min method, the more likely and possible event-time window frame’s voltage
differences always come above the three times of the computed SD. Fault voltage
or may be some small electrical property changes can Couse such type of picks as
shown in Fig. 6b. Another histogram Fig. 7a was plotted to know the number of
window frames which show possibly events. Boxplot also showing some out liars
for event Couse window frames as shown in Fig. 7b, in case there is no events are
there than Boxplot shows no out liars.

5 Conclusion

The experiments conducted depict the efficacy of the proposed methodology PMU
data storage more reliably in a decentralized fashion with security by using low-cost
commodity hardware units (i.e., Hadoop clusters and HDFS). This method is cost-
efficient than current vertical scalable high-end server systems providing more faster
processing. This paper focuses to improve the visualization of computed data point
and further analysis and storage and is necessary at load dispatch centers rather than
data point and reading of PMUs.

[Link]@[Link]
Storage and Analysis of Synchrophasor Data for Event Detection … 301

Fig. 6 a Max–Min voltage; b Max–Min voltage differences

[Link]@[Link]
302 A. P. Singh et al.

Fig. 7 a Max–Min method analysis; b Boxplot of voltage difference

Future work will look at the development of MapReduce-based clustering tech-

niques and machine learning techniques for classification of power fault detected
as well as the unique events detections. Machine learning is a pattern recognition
technique which can help to fetch specific patterns present inside the power signals,

[Link]@[Link]
Storage and Analysis of Synchrophasor Data for Event Detection … 303

every different events or faults develop a unique type of signal pattern [19]. So,
development of a self-learning neural network model for signal patent.

Acknowledgements This work has been carried out by financial assistances from SERB-DST (File
no. EMR/2017/001508). The authors wish to thank POSOCO, New Delhi for providing necessary
synchrophasor data.

References

1. Phadke, A.G.: Synchronized phasor measurements in power systems. IEEE Comput. Appl.
Power 6(2), 10–15 (1993)
2. Aghaei, J., Alizadeh, M.I.: Demand response in smart electricity grids equipped with renewable
energy sources: a review. Renew. Sustain. Energy Rev. 18, 64–72 (2013)
3. Mohanta, D., Roy, D.: Importance and uncertainty analysis in software reliability assessment
of computer relay. Proc. Inst. Mech. Eng. Part O J. Risk Reliab. 225(1), 50–61 (2011)
4. Roy, D.S., Mohanta, D.K., Panda, A.: Software reliability allocation of digital relay for trans-
mission line protection using a combined system hierarchy and fault tree approach. IET Softw.
2(5), 437–445 (2008)
5. Das, H., Roy, D.: A grid computing service for power system monitoring. Int. J. Comput. Appl.
62(20) (2013)
6. Polaki, S.K., Reza, M., Roy, D.S.: A genetic algorithm for optimal power scheduling for
residential energy management. In: 2015 IEEE 15th International Conference on Environment
and Electrical Engineering (EEEIC), IEEE, pp. 2061–2065 (2015)
7. Mondal, S., Murthy, C., Roy, D., Mohanta, D.: Simulation of phasor measurement unit (pmu)
using labview. In: 2014 14th International Conference on Environment and Electrical Engi-
neering (EEEIC), IEEE, pp. 164–168 (2014)
8. Mohanta, D.K., Murthy, C., Sinha Roy, D.: A brief review of phasor measurement units as
sensors for smart grid. Electr. Power Compon. Syst. 44(4), 411–425 (2016)
9. Murthy, C., Mishra, A., Ghosh, D., Roy, D.S., Mohanta, D.K.: Reliability analysis of phasor
measurement unit using hidden Markov model. IEEE Syst. J. 8(4), 1293–1301 (2014)
10. Murthy, C., Roy, D.S., Mohanta, D.K.: Reliability evaluation of phasor measurement unit: a
system of systems approach. Electr. Power Compon. Syst. 43(4), 437–448 (2015)
11. Murthy, C., Varma, K.A., Roy, D.S., Mohanta, D.K.: Reliability evaluation of phasor measure-
ment unit using type-2 fuzzy set theory. IEEE Syst. J. 8(4), 1302–1309 (2014)
12. Roy, D.S., Murthy, C., Mohanta, D.K.: Reliability analysis of phasor measurement unit incor-
porating hardware and software interaction failures. IET Gener. Transm. Distrib. 9(2), 164–171
(2014)
13. POSOCO. [Link]
14. Dahal, O.P., Brahma, S.M., Cao, H.: Comprehensive clustering of disturbance events recorded
by phasor measurement units. IEEE Trans. Power Deliv. 29(3), 1390–1397 (2014)
15. Liu, L., Chai, J., Qi, H., Liu, Y.: Power grid disturbance analysis using frequency information at
the distribution level. In: 2014 IEEE International Conference on Smart Grid Communications
(SmartGrid-Comm), IEEE, pp. 523–528 (2014)
16. Dahal, O.P., Brahma, S.M.: Preliminary work to classify the disturbance events recorded by
phasor measurement units. In: 2012 IEEE Power and Energy Society General Meeting, IEEE,
pp. 1–8 (2012)
17. Todd, M.J., Yıldırım, E.A.: On Khachiyan’s algorithm for the computation of minimum-volume
enclosing ellipsoids. Discret. Appl. Math. 155(13), 1731–1744 (2007)
18. Rusitschka, S., Eger, K., Gerdes, C.: Smart grid data cloud: a model for utilizing cloud com-
puting in the smart grid domain. In: 2010 First IEEE International Conference on Smart Grid
Communications (SmartGridComm), IEEE, pp. 483–488 (2010)

[Link]@[Link]
304 A. P. Singh et al.

19. Allen, A.J.: Analysis of Transmission System Events and Behavior Using Customer Level
Voltage Synchrophasor Data (2013)
20. Bach, F., Cakmak, H.K., Maass, H., Kuehnapfel, U.: Power grid time series data analysis with
pig on a Hadoop cluster compared to multi core systems. In: 2013 21st Euromicro International
Conference on Parallel, Distributed and Network-Based Processing (PDP), IEEE, pp. 208–212
(2013)
21. Song, Y., Zhu, Y., Li, L.: Large scale data storage and processing of insulator leakage current
using HBase and mapreduce. In: 2014 International Conference on Power System Technology
(POW-ERCON), IEEE, pp. 1331–1337 (2014)
22. Allen, A., Singh, M., Muljadi, E., Santoso, S.: PMU Data Event Detection: A User Guide for
Power Engineers. National Renewable Energy Laboratory (2014)
23. Foster, I., Zhao, Y., Raicu, I., Lu, S.: Cloud computing and grid computing 360-degree com-
pared. In: Grid Computing Environments Workshop, 2008. GCE’08, IEEE, pp. 1–10 (2008)
24. Borthakur, D., et al.: HDFS architecture guide. Hadoop Apache Proj. 53, 1–13 (2008)
25. Apache: Apache Thrift. [Link]
26. Apache: Apache HBase. [Link]

[Link]@[Link]
Analysing Trends in Student’s
Performance Across Maharashtra
Through Non-adaptive and Adaptive
Online Assessments Based
on the Underlying Framework
of Classical Test and Item Response
Theory

Chandrani Singh and Ajit Pandey

Abstract In today’s era with the advancement of information and communication

technology, and with the rapid digitization across all arenas, it has become a necessity
to identify and map the right resource to a required job profile in minimum time and
cost. To identify the right resource, organizations across the world have adopted
several online tools, strategies and mechanisms to assess the candidates to check
on their potential and suitability. Technologies have even paved ways for creating
immersive modes where candidates are required to get acclimatized and play the job
role and deduce the fitment for that particular role. The changes brought about by the
information and communication technology have been speedy and pervasive and so
has the penetration of artificial intelligence and machine learning where human and
technology are highly interconnected to achieve the desired outcome. In all segments
and sectors, digital assessments (in any form) play a key role in identifying the right
asset, where a lot is dependent on the impregnated intelligence and machine learning
to help decision-making become easy and accurate. It is said that artificial intelligence
could play a role in the growing field of assessment and learning analytics and also
can be applied in evaluating the quality of curriculum content so that it can be applied
to create unique pathways for individual learning. This in turn will create the desired
efficacy in assessing with precision candidate/candidates for the right opportunity. In
this regard, the authors in their second paper propose to first identify and extract the
performance trends of candidates and item characteristics using non-adaptive and
adaptive assessment techniques to create a rationale for measuring their learnability
and implementation of their learned skills in specific job roles.

C. Singh · A. Pandey (B)

Sinhgad Institute of Management, Pune, India
e-mail: awesomeajit@[Link]
C. Singh
e-mail: [Link]@[Link]

© Springer Nature Singapore Pte Ltd. 2019 305

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
306 C. Singh and A. Pandey

Keywords Performance trends · Adaptive assessment and learning

Immersive modes · Resource fitment · Item response theory · Discrimination
Pseudoguessing · Latent trait

1 Introduction

With the advancement in the learning and assessment space, the implementation of
adaptive form of assessment has come to fore in the higher education. The adaptive
form of assessment was developed between 1973 and 1984 by Weiss, Kingsbury.
Initially, such form of assessments starts with a large base of questions followed
by a selection of individual questions of average complexity for test takers to start
off with the test. The difficulty level of the test increases or decreases with each
successful/unsuccessful attempt by the test taker, respectively. The adaptive assess-
ment requires the following components to be present: question bank, calibrated to
a common measurement scale; an algorithm for item selection, a process to score
the candidate’s responses; test termination condition and score analytics. Adaptive
assessments have been found to take less time than the fixed form tests and provide
enough accurate information about the candidates to come to a viable conclusion be it
assessing them for employment opportunities or for the course/training/certifications
undertaken by them. In the former test, theories such as the classical test theory, there
have been certain limitations with regard to the following: single reliability value for
the entire test and all participants. The scores of the candidates depend on the items
rather than the candidate, and there is always a bias towards the fixed form tests to
have questions of average difficulty. On the other hand, when item response theory
became the underlying base for adaptive assessments, the latent trait models under
this theory helped to look beyond the underlying traits responsible for producing
the test performance for each item. The item response theory comprises of majorly
three components: item response function, item information function and invari-
ance. Item response function can, in other words, be correlated to item characteristic
curve which depicts the item’s value in terms of respondent’s ability as a function
of the probability of recommending an item. For dichotomous items (those scored
correct/incorrect), each item has three parameters according to the three-parameter
model:
(a) the discrimination parameter that provides the information on how well a ques-
tion or an item differentiates between the best and poor scorers. While analysing
on the item discrimination index is required to segregate the best, average and
the poor performers based on the scores received and creating the clusters.
(b) the difficulty parameter, an estimate of the appropriateness of the item to the level
of examinees which ranges from −3 to +3, with 0 being an average examinee
level.
(c) the pseudoguessing parameter, which is a lower asymptote, with a value of k
inverse where k is the number of options.

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra … 307

For an item with discrimination equal to 0.8, difficulty equal to 0.68 and the guess
being equal to 0.20, the graph for that question/item would appear as follows (Fig. 1):
This graph has been constructed based on the performance of 20 students across
ten items undergoing Knox cube test as a pilot study to understand the probability of
answering an item correctly given the latent trait of the candidate. An interpretation
of the above item/question states that with increasing ability of the candidates the
probability of getting a correct answer for the said question shows a steep rise thus
discriminating at 0.8 between weak and strong performers. Also, the probability that
the item could be guessed correctly bears a 20-pc chance amongst the low scorers
since a guess is inversely proportional to the number of options. The difficulty level
is considered to be moderately high, with low performers finding it difficult and high
scorers finding it relatively easy with b at 0.68.
Other than item response function, the second important component of IRT is the
item information function which denotes item’s quality i.e., the ability to differentiate
among respondents.
Invariance, on the other hand, can be estimated by any items with known IRFs
and item characteristics are population independent within a linear transformation.
An item information curve will take a bell shape for the typical reason that the
maximum amount of information is achievable at θ  βi. This is because the amount
of information can be extremely small for ability levels that deviate from βi. Sum-
marizing, the difficulty level of an item, βi when matched to examinee ability, the
estimation of the ability is better recorded. Associated with item information func-
tion, as shown above, is the test information function which shows the sum over the
items on the amount of item information on the ability trait as shown in Fig. 2.
Above is a generalized representation of an item’s information and also the test
information function assuming that a test contains only one item or many items. The
initial rise of the curve and a peak shows that some item/items within the test could
provide considerable information with regard to discrimination, difficulty and guess
between the low and top scorers, whereas a less steeper zone shows less information
disseminated by the items. Considering a more specific case where around 35 kids

Fig. 1 An item with discrimination 0.8, difficulty 0.68 and guess 0.20

[Link]@[Link]
308 C. Singh and A. Pandey

Fig. 2 The test information function

undergo a Knox cube test, the item characteristic curve and the test information
function justify the strong correlation to the three-parameter logistic model of the
item response theory and the bell shape of the curve. The winstep file specification
shows the details of the 35 candidates below who are children ranging from 4 years
to 11 years of age and had been subjected to the Knox cube test. The Knox cube test
was developed as a non-verbal intelligence test developed by Dr. Howard Andrew
Knox, from Ellis Island. It was first published as a paper in 1914 in the Journal of
the American Medical Association [1]. In the test 4, black 1 cubes were placed in
a row, each cube separated by 4 in. from its neighbours. The test administrator was
instructed to take a smaller cube and tap on the 4 1 cubes in increasing complicated
sequences. The test subject was requested, sometimes only by sign language, to
repeat the sequence. The cubes were numbered 1 through 4, with the sequences in
order as shown:
(a) 1, 2, 3, 4
(b) 1, 2, 3, 4, 3
(c) 1, 2, 3, 4, 2
(d) 1, 3, 2, 4, 3
(e) 1, 3, 4, 2, 1 and so on.
Dr Knox’s inference was that a child of 4 years could imitate the simple sequence
of (1–2–3–4) very easily, whereas (1–3–4–2–3–1) could be copied only by a child of
11 years. Taking the above concept into consideration, the 35 kids were subjected to
a performance test constituting of 18 items basically to determine their responses to
the remembering of correct sequences which would determine their mental ability.
Using the test, these kids were assessed on these 18 tapping patterns with the last

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra … 309

tapping patterns having seven actions. A correct response was coded as 1 and an
incorrect one as a 0 (Fig. 3).
Following the responses by the kids are the components which form the under-
lying framework of item response theory. The item information function depicts an
indication of item quality which tells that based on the level of difficulty of the item
the information is disseminated. The level of difficulty is measured based on the
responses provided by the candidates. For item 6, we see that the item information
function in Fig. 4 provides maximum information at ability level (θ ) ranging between
−1 and 2 and is about 0.25 in the scale of 0–1. Before and after that the amount of
information with regard to the ability is negligible. In other words, the item does not
disseminate information on its own attribute such as the level of complexity or the
difficulty based on the latent trait of the candidates. An item which does not provide
enough information with respect to discrimination or difficulty should be ideally
eliminated from the list of questions or item bank. Under item response theory, there
exists a single latent trait or ability level of a candidate across all item responses and
also the characteristics of one item is not contingent to the others.
This is of considerable importance to both the test constructor and the consumer
since it means that the accuracy with which an examinee’s ability is estimated depends
upon where the examinee’s ability is located on the ability scale. Ability and diffi-
culty being synonymous information is received with precision at ability level of a
candidate which matches the difficulty level of an item. As like the item informa-
tion function is the test information function the formula for which is given by the
following:


N
1I (θ )  Ii(θ) − Equation1 (1)
i1

where I(θ ) is the summation of test information at an ability level of θ , Ii(θ ) is the
amount of information for item i at ability level θ and N is the number of items
in the test. It is extremely useful to have test information function as it tells the

Fig. 3 The winstep text file representing the kid’s responses to the items

[Link]@[Link]
310 C. Singh and A. Pandey

Fig. 3 (continued)

test’s performance of determining the ability of the candidates over a range of ability
scores and more the duration of the test better is the precision of measurement. It is
to be further stated that much of item information with precision can be derived from

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra … 311

Fig. 4 The item information function

item characteristic curve for the three-parameter model using the following equation
where the ability estimate does not get compromised with the guess value:

Ii(θ )  a2[Qi(θ )/Pi(θ )][Pi(θ ) − c2]/[1 − c2] (2)

where

Pi(θ )  c + (1 − c) (1/(1 + EXP(−L)))

L  a(θ − b) Qi  1.0 − Pi(θ ).

The computation of the item information at ability level θ for a  0.8, b  0.68
and c  0.2 for 6th item from the set of 18 items is given as follows (Table 1):
From the above result extracted from the group of 18 items, the item information
provided by the 6th item is moderate (Fig. 5).

Table 1 The computation of the item information at ability level θ

θ L EXP(−L) 1/1 + P6 (−3) Q6 (−3) Q6/P6
EXP(−L)
−3 −2.94 18.91 0.05 0.040 0.96 24
((P 6 (3) − I 6 (−3)
c)2 )
0.0256 0.61
The bold results are moderate in nature as compared to others

[Link]@[Link]
312 C. Singh and A. Pandey

Fig. 5 Test characteristic

curve

Also, once information of the item is extracted the test characteristic curve gives
the information on how well the test performs in estimating the ability of candidates
over a given range of scores.
The test characteristic curve generated on the above data of 35 kids shows the
sum of the item characteristic curve of the respective items/questions attempted by
the kids and predicts the scores of the examinee at the given ability level given by
the equation:

TCC(θ )  Pi(θ ) (3)

If we see the above curve, it shows that there is a significant difference shown by a
steep rise in the scores earned on total number of questions between the high and the
low-performing kids. The immediate inference is that as the ability trait increases a
little so does the probability of answering the question rises abruptly but then the test
information should show with precision the probability of correct attempts across all
score ranges.
The underlying benefit of assessing items/questions across the parameters of dif-
ficulty, discrimination and guess provides a discrete judgement of a candidate who is
being assessed. The added edge of item response theory over classical test theory is
that while classical test theory relies on the performance of the test as a whole, IRT
chooses to assess the performance of each item based on the candidate’s attempt to
answer it. In other words, we can say that it is a two-way handshake between item and
candidate and as a candidate is assessed so also an item is assessed. In CTT an item’s
difficulty, discrimination is based on the test takers and person’s ability parameter
is dependent on the test. But with IRT instead of a single test score model there are

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra … 313

item score models and a person’s ability does not vary with the test he or she takes,
neither an item’s difficulty nor discrimination changes with the attempts made by the
candidate. Hence to have a fair judgement of prospective job aspirants, the adaptive
assessment technique which has a strong correlation with the item response theory
was implemented by the authors in this paper. The authors also extracted zonewise
performance trends of candidates using the classical test framework to have a broader
picture of the level of the candidates.
The Development Environment
For the implementation of non-adaptive and adaptive assessment, an application and
an interface were designed and developed and an environment was set up for the
assessment of the candidates (Table 2).
The development environment constitutes of a spring and hibernate framework at
the application layer, with the user interface developed in JavaScript and angular JS
and MySQL at the database layer. The entire application and the database are hosted
on the cloud across the servers. The assessment application is interlinked with the
job application platform wherein a candidate can apply for jobs and get assessed.
The platform on which the candidates apply for the job is shown in Fig. 7. This
platform is integrated with analytics dashboard for recruiters and candidates as the
stakeholders.

2 The Methodology

The data for this study are the responses received of 2016 and 2017 fresher candi-
dates both pursuing graduate or post-graduate professional course and aspiring to get
employed in some organization. The assessment comprising of items are of multiple
choice categories, and this assessment of a particular organization had been con-
ducted across the various zones of Maharashtra for the same category of candidates.
The population of this study constitutes a random sampling of 450 student data that
was selected.

2.1 The Instrument

The researchers collected the data for this study from the candidate base who par-
ticipated in the assessment conducted by an organization across Maharashtra. The
candidate base used for experimentation was around 450 in number from each zone
of Maharashtra. The total numbers of questions considered were 85, and the duration
of the test was around 90 minutes. The assessment constitutes of objective questions
with four options and the guessing parameter can be computed as the inverse of the
numbers. If the numbers of options are four, then the guessing parameter can be

[Link]@[Link]
 Table 2 Snapshot of the database
314

S. No. online_test_ Question Total no. of No. of candidates No. of candidates

question_master_ candidate attempt gives the right gives the wrong
id the question answer answer
1 2490 Out of 7 consonants and 4 vowels, how many words of 3 368 93 275
consonants and 2 vowels can be formed?
2 2491 In how many ways can the letters of the word ‘LEADER’ be 371 139 232
arranged?
3 2492 In a group of 6 boys and 4 girls, four children are to be selected. In 364 82 282
how many different ways can they be selected such th
4 2493 In a shower, 5 cm of rain falls. The volume of water that falls on 358 156 202
1.5 ha of ground is:
5 2494 A hall is 15 m long and 12 m broad. If the sum of the areas of the 363 122 241
floor and the ceiling is equal to the sum of the areas of f
6 2495 66 cubic centimetres of silver is drawn into a wire 1 mm in 357 59 298
diameter. The length of the wire in metres will be
7 2496 A hollow iron pipe is 21 cm long and its external diameter is 8 cm. 347 145 202
If the thickness of the pipe is 1 cm and iron weighs 8 g/
8 2497 A boat having a length 3 m and breadth 2 m is floating on a lake. 358 189 169

[Link]@[Link]
The boat sinks by 1 cm when a man gets on it. The mass
9 2498 A boat can travel with a speed of 13 km/h in still water. If the speed 361 208 153
of the stream is 4 km/h, find the time taken by the
10 2499 A man’s speed with the current is 15 km/h and the speed of the 367 111 256
current is 2.5 km/h. The man’s speed against the current
(continued)
C. Singh and A. Pandey
 Table 2 (continued)
S. No. online_test_ Question Total no. of No. of candidates No. of candidates
question_master_ candidate attempt gives the right gives the wrong
id the question answer answer
11 2500 If one-third of one-fourth of a number is 15, then three-tenth of 375 153 222
that number is
12 2501 I saw a __________ of cows in the field 376 104 272
13 2502 Success in this examination depends _____________ hard work 380 351 29
alone
14 2503 choose the word which is the exact OPPOSITE of ENORMOUS 372 195 177
15 2504 choose the word which is the exact OPPOSITE of EXODUS 364 136 228
16 2505 Tanya is older than EricCliff. is older than TanyaEric. is older than 378 318 60
CliffIf. the first two statements are true, the third statement is not
correct
17 2506 All the trees in the park are flowering treesSome. of the trees in the 375 224 151
park are dogwoodsAll. dogwoods in the park are floweri
18 2507 Mara runs faster than GailLily. runs faster than MaraGail. runs 375 316 59
faster than LilyIf. the first two statements are true, the third
statement is not correct
19 2508 All the tulips in Zoe’s garden are whiteAll. the pansies in Zoe’s 380 106 274

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra …

garden are yellowAll. the flowers in Zoe’s garden are either white
or yellow
20 2509 Form the sentence: 1. left 2. the 3. house 4. he 5. suddenly 374 344 30
21 2510 Form the sentence: 1. of 2. we 3. heard 4. him 5. had 375 274 101
315
316 C. Singh and A. Pandey

computed as 0.25 or 0.20. In the scheme of assessment, the content to be covered for
assessments was already notified to the candidates.
To examine the difficulty, discrimination and guessing parameter using the exam-
inees’ response, of test items, the analyses were performed on only the items that
fitted the 3PL model. The result concludes that around 63 items can fit the three PL
model of the item response theory using chi-square goodness of fit and a sample ICC
for item 12 is shown in Figs. 6 and 7.

Fig. 6 Item fit snapshot using chi-square test

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra … 317

Fig. 7 Candidate’s performance scenario—Zonewise

After the candidates’ application on the platform, the assessment environment for
these candidates enables them to take up both the forms of tests as shown below.

2.1.1 Analytics with CTT and IRT Framework (Item, Candidate,

Performance Trends)

The analytics is based on CTT and IRT framework for the non-adaptive and adaptive
assessments for the candidates across the zones. The classical test theory is based on
the fact that CTT assumes that the true score plus the error gives the observed score.

X T+E (4)

where T  the true score and E  the error component but the CTT has its own
limitations in terms of analysis of the examinees’ proficiency and quality of test
items. The IRT, on the other hand,takes as explained above the three parameters into
consideration:
a as in item discrimination
b as in item difficulty
c as in pseudo guessing
The IRT model which has been taken into consideration in this paper is the 3.
Parameter logistic models which take the following form:

Pi(θ )  ci + (1 − ci) ∗ 1/(1 + e − Daj(θ − bi)− (5)

[Link]@[Link]
318 C. Singh and A. Pandey

where c, a and b have already been explained before and D being an arbitrary constant
with a value of 1.7 and θ being the ability level of the examinee.
These logistic equations when displayed in the form of graphs produced plots that
are called item characteristic curves as discussed in the introductory section where
the ability of the examinee is denoted by theta (θ ) on the x-axis, while the probability
of an examinee correctly answering the question is denoted by P (θ ) on the y-axis
which is denoted by an s-shaped curve.
The environment till now has facilitated the assessment of around 20,000 students
mostly from the various districts of Maharashtra of which 15,000 candidates have
been assessed on tests whose underlying framework is based on classical test theory,
whereas the rest 5,000 have been subjected to adaptive assessments. Prior to putting
forth the analysis of item/candidates, a broad scenario of the zonewise performance
of the candidates is shown below. An estimate varying between 450 to 500 students
from each zone of Maharashtra had appeared for the non-adaptive assessment, and
the following observations were noted:
The overall performance as clearly known from the graph shows candidates from
Pune zone performing better than Nagpur and Aurangabad.
Using the concept of item response theory and also to test the item’s fitness to the
entire assessment, the discrimination and the difficulty index was also calculated. For
this reason, the entire student segment was segregated into three segments. Around
30 pc each were considered poor, average and good performers. For each question
answered correctly, the discrimination index was computed as shown for the zones
below.
To calculate the discrimination and difficulty index, the following steps were
observed. Candidates with highest overall scores were allowed to occupy the top
positions in the excel file followed by candidates with average and lower scores. For
the difficulty index, the computation of each item/question was done as follows:
Difficulty Index (Q1)  Total no of students who answered from the top segment
of students + total no of students who answered from the bottom segment/total no of
students who gave the exam. For Q1, it was b  0.155.
The discrimination index was obtained by subtracting the number of candidates
in the lower segment who got the item/question correct from the number of students
in the upper segment who got the item correct. Then, divided by the number of
students in each group. In this case, it would be 150 students in each group. For
Q1, it was discrimination index (Q1) was 0.06. In the same way, the discrimination
and difficulty index of the other questions has been computed for the students who
attended the assessments. If we observe the discrimination and difficulty index of
the 50th question, we see that difficulty being high the question does not do very
well discriminate the top and the poor performers. Reason being that a fair number
of poor performers could answer the question correctly depicting they are prepared
for difficult questions. It was also seen that questions those were less difficult had
higher discrimination power but were less difficult where the question as in question
number 28 had the discrimination and the difficulty in the same range. Around 40 pc
of the questions of the assessment conducted have high-to-moderate discrimination

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra … 319

Fig. 8 Scree plot showing contribution of factors

Fig. 9 Item characteristic

curve for item 12

powers, whereas the difficulty index of 60 pc of the questions ranges from high to
moderate.
An interesting revelation is that while performing a factor analysis on the data set
the authors could identify the contribution of the various factors involved through
the scree plot as shown below (Figs. 8 and 9; Table 3):
The performance trend-line of Aurangabad candidates depicts a real need for
creating unique pathways of learning.
The percentile score trends for the candidates, using classical test framework, from
across the various districts of Maharashtra as shown in Fig. 10 depict that highest
count of candidates from Pune falls in the category of 0–10‰ below, whereas the
count of candidates from Nagpur is the least when we consider the percentile range
of (0–10) and are moderate on their overall performance. The data has been collated
across the three zones taking into consideration strength of 450 students each and
considering the last five performances in the assessment conducted by more than
one organization. Students were chosen from a base of around 20,000 students of
which around 4500 students had attended more than two assessments and around
1500 more than four tests. Using excel random sampling software, the sampling was
conducted (Fig. 11).
Extracting the moving average, for the year 2016, from a base of candidates as
elaborated in the preceding section, who had participated in the assessment across
the zones of Maharashtra, the forecast depicts that the average performance scores
would increase across the candidate base for the first three segments and then would
take a downward trend for the last segment of candidate’s assuming that performance

[Link]@[Link]
 Table 3 Discrimination and difficulty index of items/questions
320

S. No online_test_que Total No. of candidates Total no. of No. of candidates Discrimination Difficulty (b)
stion_master_id gives the right candidate attempt gives the right (a)
answer the question answer
1 2495 122 23 116 14 0.067834935 0.155462185
2 2498 120 87 121 52 0.295247934 0.576763485
3 2497 121 77 119 59 0.140565317 0.566666667
4 2537 123 33 122 19 0.112554978 0.212244898
5 2525 124 79 117 38 0.312310449 0.485477178
6 2494 123 48 117 28 0.150927663 0.316666667
7 2496 121 57 113 39 0.125941637 0.41025641
8 2499 124 51 121 23 0.221207678 0.302040816
9 2506 124 83 126 67 0.137608807 0.6
10 2508 125 52 128 23 0.2363125 0.296442688
11 2569 118 37 107 22 0.107951845 0.262222222
12 2573 120 95 108 65 0.189814815 0.701754386
13 2514 125 95 124 52 0.340645161 0.590361446
14 2513 123 37 125 26 0.092813008 0.254032258

[Link]@[Link]
15 2515 122 74 122 57 0.139344262 0.536885246
16 2503 124 81 125 41 0.325225806 0.489959839
17 2504 125 52 118 40 0.077016949 0.378600823
18 2517 125 119 126 102 0.14247619 0.880478088
19 2539 124 44 121 37 0.049053586 0.330612245
20 2516 124 104 126 79 0.21172555 0.732
(continued)
C. Singh and A. Pandey
 Table 3 (continued)
S. No online_test_que Total No. of candidates Total no. of No. of candidates Discrimination Difficulty (b)
stion_master_id gives the right candidate attempt gives the right (a)
answer the question answer
21 2558 121 54 120 33 0.171280992 0.360995851
22 2521 123 95 116 61 0.246495655 0.652719665
23 2520 120 39 115 24 0.116304348 0.268085106
24 2509 125 124 126 111 0.111047619 0.93625498
25 2510 125 107 124 76 0.243096774 0.734939759
26 2527 122 54 125 30 0.202622951 0.340080972
27 2523 124 57 128 32 0.209677419 0.353174603
28 2501 125 48 126 22 0.209396825 0.278884462
29 2522 124 68 124 52 0.129032258 0.483870968
30 2534 123 72 120 37 0.27703252 0.448559671
31 2500 124 70 125 30 0.324516129 0.401606426
32 2530 125 73 126 37 0.290349206 0.438247012
33 2556 124 51 125 37 0.115290323 0.353413655
34 2492 123 41 120 22 0.15 0.259259259

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra …

35 2528 122 45 119 28 0.133558341 0.302904564

36 2493 120 54 123 54 0.01097561 0.444444444
37 2491 124 70 122 33 0.294024326 0.418699187
38 2512 125 109 128 77 0.2704375 0.735177866
39 2507 125 115 126 98 0.142222222 0.848605578
40 2545 124 100 126 71 0.242959549 0.684
41 2529 123 59 122 42 0.135412502 0.412244898
(continued)
321
 Table 3 (continued)
322

S. No online_test_que Total No. of candidates Total no. of No. of candidates Discrimination Difficulty (b)
stion_master_id gives the right candidate attempt gives the right (a)
answer the question answer
42 2519 125 82 129 54 0.237395349 0.535433071
43 2490 125 37 122 26 0.082885246 0.255060729
44 2518 125 54 125 36 0.144 0.36
45 2542 124 63 123 36 0.215381589 0.400809717
46 2541 124 115 126 79 0.300435228 0.776
47 2533 124 77 129 46 0.264378595 0.486166008
48 2532 125 87 125 48 0.312 0.54
49 2531 124 86 122 68 0.136171338 0.62601626
50 2502 125 119 129 114 0.06827907 0.917322835
51 2505 125 120 129 89 0.270077519 0.822834646
52 2540 123 55 124 49 0.051993181 0.421052632
53 2526 122 30 122 11 0.155737705 0.168032787
54 2548 125 76 124 51 0.196709677 0.510040161
55 2524 123 12 120 9 0.022560976 0.086419753

[Link]@[Link]
56 2555 124 61 125 46 0.123935484 0.429718876
57 2535 123 42 121 30 0.09352953 0.295081967
58 2554 125 107 122 77 0.224852459 0.744939271
59 2561 120 32 123 21 0.095934959 0.218106996
60 2560 123 15 124 9 0.049370574 0.097165992
61 2552 125 35 123 18 0.133658537 0.213709677
62 2511 125 107 128 72 0.2935 0.707509881
(continued)
C. Singh and A. Pandey
 Table 3 (continued)
S. No online_test_que Total No. of candidates Total no. of No. of candidates Discrimination Difficulty (b)
stion_master_id gives the right candidate attempt gives the right (a)
answer the question answer
63 2574 120 19 106 5 0.111163522 0.10619469
64 2546 125 106 124 74 0.251225806 0.722891566
65 2547 124 16 122 14 0.01427816 0.12195122
66 2551 125 56 126 61 −0.036126984 0.466135458
67 2550 124 58 120 42 0.117741935 0.409836066
68 2567 119 40 105 30 0.050420168 0.3125
69 2559 123 64 124 53 0.092905848 0.473684211
70 2566 119 63 107 22 0.323804288 0.376106195
71 2562 125 38 127 17 0.170141732 0.218253968
72 2570 121 21 106 15 0.032044285 0.158590308
73 2568 119 28 107 22 0.029686641 0.221238938
74 2557 123 49 125 18 0.254373984 0.27016129
75 2572 120 48 103 34 0.069902913 0.367713004
76 2571 118 52 105 23 0.221630347 0.33632287

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra …

77 2565 116 35 108 24 0.079501916 0.263392857

78 2538 124 95 121 59 0.278525726 0.628571429
79 2536 123 35 124 29 0.050681878 0.259109312
80 2564 125 72 124 53 0.148580645 0.502008032
81 2553 124 62 125 39 0.188 0.40562249
82 2549 125 91 126 55 0.291492063 0.581673307
83 2563 123 63 125 45 0.152195122 0.435483871
84 2544 124 75 124 66 0.072580645 0.568548387
85 2543 122 57 125 51 0.059213115 0.437246964
323
324 C. Singh and A. Pandey

Fig. 10 Zonewise percentile trends

Fig. 11 Performance trend of candidates—moving average

across three student segments had been made for a group of 150 students each across
the zones of Maharashtra for based on the previous five assessments.

3 Conclusion and Future Work

The findings from this study indicate that for assessing the candidates it is highly
recommended that the items should be fairly good in their characteristics and also
should possess the necessary value to make the overall test performance well. If the
quality or the standard of the test items are improved so will increase the precision
with which the candidate can be measured or assessed by educators or recruiters.
It is crucial that the candidate’s response to test item is proportional to the overall
development of the test and measures the facets of learning with the greatest precision
and accuracy and good quality test items are indicators for those measures. The
growth in computer adaptive testing, in particular, has supported the growing interest
in the use of IRT and IRT has now become a necessity in assessing learning skills and

[Link]@[Link]
Analysing Trends in Student’s Performance Across Maharashtra … 325

learnability and in developing improved measures for assessing change over time.
Educational/employment tests are the main source of information about candidate’s
achievement; hence, the analysis of test data is crucial in determining the test quality
and test information. Hence, the prime task for any educator or recruiter is to choose
quality items for better precision. In the research conducted by the authors, the
adaptive assessment framework will enable recruiters to align items more to job
profiles and candidate’s ability traits so that selection and retention ratio can increase
thus giving lesser turnaround time and more employment satisfaction.

Reference

1. Adedoyin, O.O.: University of Botswana, Botswana T Mokobi Department of Educational Plan-

ning and Research Services, Botswana, USING IRT PSYCHOMETRIC ANALYSIS IN EXAM-
INING THE quality of junior certificate mathematics multiple choice examination test items,
journal homepage: [Link]

[Link]@[Link]
Author Index

A H
Abhay Prithvi, K., 279 Hemant Kumar, G., 291
Abhishek Bhatt, 227, 247
Aishwarya Nair, 109 I
Ajit Pandey, 305 Indrani Das, 43
Akhilendra Pratap Singh, 291
Amreen Ahmad, 227, 247 J
Amudha, T., 235 Jagdish Kumar, 253
Aniket Dabas, 205 Jitendra Kumar Verma, 1
Anurag Singh, 185 Jugendra Singh, 159
Arpit Paruthi, 279
Ashique Mohaimin Ornab, 147 K
Ashwini Kumar, 43 Kumkum Garg, 83
Kunal Sain, 69
B
Bhatia, M. P. S., 195 M
Mansi Saxena, 1
C Meghavi Choksi, 135
Chandrani Singh, 305 Mohapatra, B., 69
Chetna Dabas, 205 Mukesh A. Zaveri, 135
Christy Martina, J., 235
N
D Nighat Rahman, 159
Dayanand Gaur, 123 Nithya, B., 109
Dharmender Singh Kushwaha, 29
Diptendu Sinha Roy, 291 P
Pallavi Yarde, 83
E Prashant Johri, 43
Eva Patel, 29
R
G Rabindra Nath Shaw, 1
Gautam Seth, 279 Raj Patel, 267

© Springer Nature Singapore Pte Ltd. 2019 327

L. C. Jain et al. (eds.), Data and Communication Networks, Advances in Intelligent
Systems and Computing 847, [Link]

[Link]@[Link]
328 Author Index

Rishi Pal Singh, 43 Sufyan Beg, M. M., 219

Ritesh Srivastava, 173, 195 Sukhwinder Singh, 253
Ruqaiya Khanam, 159 Sumit Srivastava, 83

S T
Sachan, V. K., 93 Tanvir Ahmad, 219, 227, 247
Sadaf Siddiqui, 247
Sakia Chowdhury, 147 U
Saksham Jain, 279 Umang Soni, 279
Sakshi Aggarwal, 123
Samir Ahmad Sheikh, 57 V
Sandeep Kumar, 253 Veena Tayal, 173, 195
Sanjoy Das, 43 Veenu Kansal, 13, 21
Sathish Kumar, J., 135 Verma, J. K., 195
Saurabh Kumar, 135 Vivek Kumar Jaiswal, 185
Seevieta Biswas Toa, 147
Shekhar Yadav, 185 W
Shivam Mishra, 93 Waseem Ahmed, 219
Shyam Krishna Nagar, 185
Simranjit Singh, 13, 21 Z
Sindhu Hak Gupta, 57 Zohreen Haseen, 159
Sreelakshmi, A. S., 109
Subhendu Sekhar Paik, 291

[Link]@[Link]