This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2023.0322000

Augmentation of Fingerprints for Indoor

BLE Localization Using Conditional
GANs
SUHARDI AZLIY JUNOH AND JAE-YOUNG PYUN
Wireless and Mobile Communication System Lab, Department of Information and Communication Engineering, Chosun University, Gwangju, South Korea
(e-mails: suhardi@[Link])
Corresponding author: Jae-Young Pyun (e-mail: jypyun@[Link]).
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT).(No.
NRF-2022R1A2B5B01002385).

ABSTRACT
Location estimation in indoor environments using radiofrequency (RF) has garnered considerable attention
in recent years owing to the widespread adoption of mobile devices. RF-based fingerprinting—a direct
approach that allows location estimation based on observed signals—relies on manual surveys during
the offline phase to create a radio map with coordinates and RF measurements at multiple locations.
The accuracy of RF fingerprint-based localization is related to the number of reference points. However,
conventional site survey procedures tend to incur substantial expenses. To alleviate the workload of site
surveys and address the challenge of incomplete fingerprint databases, we propose a data-augmentation
method to complement existing fingerprint data. Our approach leverages a conditional generative adversarial
network with long short-term memory (CGAN-LSTM) prediction model to effectively learn the intricate
patterns inherent in the initial training data and generate high-quality synthetic data that align with the un-
derlying data distribution. In an experimental evaluation conducted on a real testbed, our data augmentation
framework increased the average localization accuracy by 15.74% compared with fingerprinting without
data augmentation. Furthermore, experiments conducted in two typical indoor environments using sparse
data highlighted the significant performance enhancement of the proposed approach in reducing localization
error and was comparable to state-of-the-art data-augmentation methods.

INDEX TERMS Bluetooth low energy (BLE), fingerprint, data augmentation, generative adversarial
network (GAN), location estimation.

I. INTRODUCTION Wi-Fi access points (APs) for localization, BLE-based meth-

Recently, the scope of location-based services (LBSs) has ods provide enhanced deployment flexibility. BLE beacons
expanded from outdoor to indoor settings. This expansion can be easily installed in less-visited areas and operate on
is driven by the recognition that individuals spend approxi- batteries. The advantages of Bluetooth technology include
mately 80% of their time indoors [1], [2], which has led to low energy consumption, cost-effectiveness, ease of deploy-
the emergence of indoor mobile applications, such as indoor ment, and the capability to achieve accurate localization.
navigation and smart building solutions. Although global Thus, Bluetooth has emerged as a competitive technology in
navigation satellite systems are widely used for outdoor lo- various domains, including the Internet of Things (IoT) [4],
cation tracking, their indoor effectiveness is limited because and has the potential to gain a larger market share in indoor
of signal attenuation. Consequently, research efforts have localization.
focused on leveraging radiofrequency (RF) signals, including RF-based indoor positioning systems (IPSs) employ sev-
Wi-Fi and Bluetooth low energy (BLE), for indoor LBS, eral methods, including the angle of arrival, the time of
given the widespread deployment of wireless devices and RF arrival, the time difference of arrival, and fingerprint-based
sensors in indoor environments [3]. approaches [5]. RF-based fingerprint indoor positioning tech-
Compared with Wi-Fi-based approaches, which depend on nology is widely utilized owing to its simplicity and minimal

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

hardware requirements, offering user convenience through extend coverage to new areas, increasing the position-
the extensive use of wireless devices such as smartphones. ing accuracy—particularly in unsurveyed locations.
One of the primary challenges faced by RF-based fin- 2) This paper introduces a CGAN-based long short-term
gerprint positioning systems is the extensive scale of sur- memory (LSTM) network to find the best RSS pre-
veying needed to collect sufficient received signal strength diction and augment the fingerprints. The proposed
(RSS) data at multiple reference points (RPs) to construct model was compared with other CGAN deep learning
an indoor radio map. In large-scale deployments, survey prediction models.
efforts are expensive. In addition, the radio map changes 3) Furthermore, a comparative analysis was performed to
over time, necessitating periodic calibration. As the area and evaluate the proposed algorithm against three state-of-
RSS measurement time increase, the dataset size required for the-art algorithms in two distinct scenarios, encom-
calibration increases. Despite the time-consuming and labor- passing rooms of varying dimensions and interfer-
intensive nature of site surveying, fingerprinting-based ap- ence levels. The experimental results indicated that the
proaches continue to be popular owing to their applicability proposed approach achieved satisfactory localization
to IPSs [6], [7]. performance.
To address the challenge of data collection costs, re- The remainder of the paper is structured as follows. Sec-
searchers have explored various approaches, such as crowd- tion II provides an overview of the related work. Section
sourcing, interpolation, signal propagation models, and si- III presents the design of the proposed scheme. Section IV
multaneous localization and mapping (SLAM). Their lo- presents the experimental results and comparisons with those
calization accuracies are significantly affected by both the of state-of-the-art methods. Finally, we summarize the study
number of RSS values per fingerprint and the fingerprint in Section V and outline future work.
density within a specific area [8]. The positioning accuracy
can be increased by increasing the number of RPs collected II. RELATED WORK
in the offline phase [9]. However, this approach incurs sub- Compared with outdoor localization, indoor localization us-
stantial costs associated with the offline data collection. To ing RSS-based fingerprints presents more significant chal-
mitigate this challenge, fingerprint augmentation is an effec- lenges owing to the unpredictable nature of environmental
tive solution to reduce costs while maintaining positioning factors. When utilizing RSS-based fingerprints, we must
accuracy [10]. Yet, these existing methods have limitations consider two critical components: 1) the selection of RF tech-
when it comes to generating diverse synthetic data. More- nologies for fingerprinting and 2) the method employed for
over, augmenting Bluetooth fingerprint data faces significant fingerprint data collection. This section summarizes related
challenges due to noise, device variability, and environmental studies focusing on three key aspects: general RF indoor
changes [11]. localization technologies, fingerprinting systems, and data-
The Conditional Generative Adversarial Network (CGAN) augmentation methods.
stands as one of the most popular GAN methods, proficient in
A. GENERAL RF INDOOR LOCALIZATION
generating synthetic data under specific conditions or scenar-
TECHNOLOGIES
ios within various environments. This synthetic data exhibits
remarkable diversity and closely mimics real-world data, In recent years, various smartphone-oriented IPSs have been
owing to its innate capability to adapt to various conditional investigated, including commonly used technologies such as
settings [12], [13]. Additionally, the CGAN approach proves Wi-Fi, Bluetooth, ultra-wideband (UWB), ZigBee, and cel-
highly effective in addressing this challenge by enabling lular networks, each of which has strengths and limitations.
data generation based on specific class labels, facilitating the In addition to the localization algorithm, the selection
targeted generation of data for a particular type. The CGAN of wireless technology plays a crucial role in designing an
framework requires the generator and discriminator to be effective localization system. Among the various wireless
conditioned on auxiliary information, such as class labels. technologies, Wi-Fi has gained significant popularity and is
This conditioning acts as an extension to the latent space, widely employed in localization systems [20]. Furthermore,
enabling the generation and discrimination of synthesized with recent advancements in Bluetooth technology, many
data [14]. systems leverage BLE beacons for indoor localization [21],
[22]. Another promising communication technology for IPS
In this study, we considered a scenario in which only a
is UWB [23], which offers advantages such as low power
small amount of labeled data was available, and data aug-
consumption, high data rates (up to 1 Gb), and remarkable
mentation was used to interpolate the missing fingerprint data
accuracy. The results of the survey conducted in [24] high-
and extend the fingerprint database. The main contributions
light a remarkable surge in BLE-based research during 2021
of this study are summarized as follows.
and 2022, surpassing research in the domains of Wi-Fi and
1) A CGAN-based localization system is proposed for UWB in comparison to previous years.
generating supplementary data by leveraging exclu- Although less prevalent than Wi-Fi and Bluetooth, ZigBee
sively labeled data. This approach facilitates generat- has attracted attention for localization applications owing
ing RSS measurements and corresponding positions to to its low power consumption and widespread use in IoT
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

TABLE 1. RF technologies for indoor localization.

Ref. Technology Coverage Energy Cost Localization Advantage Disadvantage

error
[15] Wi-Fi 30–100 m High Low 2–10 m Ubiquitous, low costs Prone to interference, vulnerable
to AP changes, meticulous Wi-Fi
AP selection for positioning
[16] Bluetooth 7–90 m Very low Low 1–5 m Low energy and cost, low la- Prone to noise, require optimal
tency, ubiquitous, compatibility beacon placement
(Android and iOS)
[17] UWB 10–15 m Medium High 10–30 cm High accuracy in line-of-sight Requires special equipment, sig-
environments, high transmission nal interference, antenna varia-
rate, strong penetration ability tions, low accuracy in complex in-
door environments
[18] ZigBee 10–100 m Medium Low 1–10 m Lower power, secure and reliable Interference, short communication
transmission range, limited data transfer rate
[19] Cellular net- Long Medium High 4m Wide coverage area Cellular signals may not deeply
works range penetrate buildings, leading to
weak signal reception, signal in-
terference

deployments. Alternatively, cellular signals such as 3G, 4G, or floor plans, along with the assumption of a constant walk-
and 5G collected by smartphones offer the advantage of ing speed. The surveyor follows a predesigned path between
cost reduction—by eliminating the need for additional hard- landmarks while maintaining a constant walking speed. To
ware installation—and widespread availability. However, the overcome the limitation of a constant walking speed, the
adoption of 5G technology remains limited owing to a lack pedestrian dead reckoning (PDR) algorithm is employed to
of widespread infrastructure. Each of these technologies has track the movement of pedestrians. Using PDR-based meth-
advantages and disadvantages when utilized in localization ods, fingerprints can be collected while walking, significantly
systems. Table 1 summarizes the approaches for RF tech- reducing the time required for radio-map construction. How-
nologies used in indoor localization, including the accura- ever, it is essential to note that the PDR approach is subject to
cies, advantages, and disadvantages of these methods. drift error, necessitating complex algorithms to compensate
Considering the cost, energy consumption, and deploy- for this error [26].
ment, Wi-Fi and BLE can be good choices. BLE beacons Crowdsourcing offers an efficient and cost-effective ap-
require less energy than Wi-Fi APs because of their IoT de- proach for collecting RF fingerprints using smartphones,
sign, which allows them to operate for years using traditional allowing the generation and updating of RF fingerprint
batteries. Additionally, the advertisement period in BLE is databases through the collective effort of individuals [27].
far shorter than that in Wi-Fi, allowing multiple RSS read- For example, as discussed in [28], the authors have inte-
ings per second. These characteristics make BLE a suitable grated labeled fingerprints with a crowdsourcing system that
substitute for Wi-Fi in fingerprinting applications [25]. Other incorporates indoor floor plans and pedestrian walking traces
reasons for using BLE beacons are their ubiquity and the to construct a comprehensive Wi-Fi radio map. The map-
ease of measuring RSS values using modern smartphones. assisted approach to generate fingerprints was introduced in
Because most smartphones have Bluetooth technology built- [29]. It utilizes crowdsourced fingerprints that are calibrated
in, BLE has become a suitable choice for indoor localization. with supplementary map information. However, constructing
an accurate RF map in dynamic crowdsourcing scenarios is
B. OVERVIEW OF FINGERPRINTING SYSTEM challenging—particularly when users move freely without
In traditional fingerprint-based methods, constructing a radio additional sensors. Although previous studies have focused
map involves collecting a set of RSS values at grid points on offline database creation through crowdsourcing, col-
from the surrounding APs. This process includes annotating lecting data from freely moving users poses challenges, as
fingerprints with location tags during the offline phase to they may include unqualified inertial data that compromise
create an RSS radio map of the area of interest. In the localization accuracy. Moreover, the distributions of crowd-
online phase, the system estimates the user’s location by sourced participants in space and time are often uneven,
matching the observed fingerprint with the fingerprints stored resulting in incomplete signal data collected through these
on the radio map. Owing to its comparable or even superior methods [30].
localization performance without additional hardware or in- As the indoor RF environment undergoes changes over
frastructure costs, fingerprinting localization can be widely time, such as the replacement of faulty BLE beacons, furni-
adopted in real-life scenarios, despite the labor-intensive and ture relocation, and the movement of people, periodic updates
time-consuming features of radio-map construction. to the radio map become essential [31], [32]. An outdated
To minimize the costs associated with site surveys, fin- fingerprint database can lead to a deterioration in localization
gerprints are collected dynamically while walking. This dy- performance [33]. Directly collecting indoor radio maps is a
namic approach to data collection relies on either landmarks time-consuming and costly endeavor, particularly in larger
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

spaces. Additionally, constructing a comprehensive radio both complex and high-dimensional, can pose limitations
map is challenging, especially in dynamic and crowdsourced on the performance of ML approaches. These methods can
scenarios where users move arbitrarily. experience significant performance degradation when RSS
Adapting to these dynamic changes requires frequent values in the fingerprint vector are unstable due to unforeseen
model retraining, resulting in significant computational over- environmental changes [45]. Tree-based models, including
head. Therefore, data augmentation offers a solution to ex- Random Forests and decision trees, exhibit robust predictive
tend the training database, especially in cases where only capabilities. However, their construction demands substantial
limited labeled data is available [34]. By implementing data time and computational resources, particularly when nu-
augmentation, the DL model can learn from the labeled merous trees are involved [46]. As networks become more
data and expand its database. It is well-suited for scenarios complex in indoor settings, there is a demand for more so-
where obtaining ample labeled training data is challeng- phisticated ML methods compared to traditional supervised
ing and expensive [35]. This approach involves generating ML to address localization challenges [47]. In contrast, deep
new synthetic data to complement the real collected data, learning (DL) shows great promise for improving localiza-
enhancing model training, improving localization accuracy, tion accuracy in complex environments where feature ex-
and simultaneously reducing measurement time and human traction is challenging, and data exhibits high dimensionality
effort. These advantages further enhance the practicality and [48]. Moreover, GAN, which leverages DL models, excels
applicability of our proposed approach. in managing complex RSS-based fingerprint environments
through data augmentation for fingerprint reconstruction [49]
C. DATA AUGMENTATION-BASED APPROACH FOR despite being computationally more complex [50].
FINGERPRINT DATABASE CONSTRUCTION
Data augmentation plays a critical role in enhancing the III. SYSTEM MODELING
localization performance by increasing the training data di- A. SYSTEM OVERVIEW
versity and sample size [36]. Although it is widely used Fig. 1 illustrates the proposed positioning system. In the
in image classification tasks [37], its effectiveness has been initial stage, the RSS of BLE is collected using a mobile
demonstrated in various domains, including natural language device. Concurrently, the coordinates of various locations and
processing, speech recognition, object detection, and human their corresponding RSS fingerprints are collected during an
activity recognition [36]. offline site survey. Multiple site surveys are conducted to
Research has focused on data augmentation for fingerprint ensure sufficient data collection. The red circle represents
generation to address the challenge of acquiring sufficient the original data (real data) at a fixed location. To enrich the
radio fingerprints for indoor localization under labor and time fingerprint RSS database, the system utilizes conditional gen-
constraints. Various approaches have been proposed, such as erative adversarial networks (CGANs) to generate synthetic
Gaussian process regression (GPR) [15], [38], [39], inverse RSS fingerprints. The blue circle represents the augmented
distance weighting (IDW) [40], and linear interpolation (LI) data, which has been randomly augmented to cover the
[41]. However, these methods have limitations in accurately entire area. Data augmentation is achieved by implement-
augmenting fingerprinting data for indoor localization be- ing the CGAN architecture, which comprises two essential
cause of their assumptions and limited ability to capture com- components: a discriminator and a generator. In contrast to
plex spatial relationships and nonlinearities in radio signal traditional generative adversarial networks (GANs), CGANs
propagation. incorporate conditional labels to facilitate the targeted gener-
Recently, graph-based data-augmentation techniques [20], ation of fingerprints specific to particular floors or buildings.
[42], [43] have been used to construct large-scale finger- Fig. 2 illustrates the layers of the discriminator and gen-
print data. In [42], radio fingerprints were modeled as graph erator model, along with the number of units and details
signals, and virtual RPs were interpolated into the graph of the activation function. The generator network comprises
to solve the radio-map reconstruction problem. In another LSTM and dense layers with 32 and 16 units, respectively.
approach [20], graph convolutional networks (GCNs) and In CGAN-LSTM, noise and labels are required at the gen-
domain adversarial training are integrated to learn robust erator input, while real data and labels are necessary at the
and domain-invariant features for localization. However, the discriminator input. Both components operate in a mutually
use of graph-based methods for data augmentation in indoor adversarial manner. In contrast, the discriminator layer con-
localization requires prior knowledge of the spatial structure, sists of one LSTM layer and three dense layers with 128
a well-defined graph representation, and the assumption of a units, 64 units, 16 units, and 1 unit, respectively. Each LSTM
consistent spatial structure across different domains, which unit comprises four components: 1) Cell State, 2) Input Gate,
may only occasionally hold true in real-world scenarios. 3) Forget Gate, and 4) Output Gate, as shown in Fig. 3.
Machine learning (ML) approaches like KNN, support The LSTM unit is essential for maintaining long-term depen-
vector machines, and random forests are highly effective for dencies, thereby improving overall outcomes [51]. Both the
regression and classification tasks [44]. However, they may discriminator and the generator employ the rectified linear
encounter challenges when dealing with complex, nonlin- unit (ReLU) activation function for all layers, except the last
ear relationships in data. The RSS fingerprint data, being one, to disregard the negative weighted values. In addition, a
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

Real data
Dataset
Classification
Label output from the
Discriminator
discriminator
Model
model
Random (real/fake)
Generator Generated
noise Model data
Label Label

Index BLE fingerprints

1 (MAC11, RSS11), (MAC12, RSS12),…,(MAC1m, RSS1m)
Offline 2 (MAC21, RSS21), (MAC22, RSS22),…,(MAC2m, RSS2m) Reconstructed map
phase (augmented data)
... ...

n (MACn , RSSn ), (MACn , RSSn2),…,(MACnm, RSSnm)

1 1 2

Online Online BLE Localization Position

phase RSS measurement algorithm estimation

FIGURE 1. BLE-based fingerprint IPS.

sigmoid activation function is applied to the output layer. information from various BLE beacons. This data, accom-
panied by their respective Media Access Control (MAC)
B. BLE BEACON-BASED FINGERPRINT LOCALIZATION addresses and coordinates, establishes a radio fingerprint
The RSS vector at distinct locations in indoor settings ex- database for subsequent online localization processes. The
hibits noticeable differences due to varying distances from database, encapsulating information about RPs and their
BLE beacons to smartphones, thereby delineating unique corresponding fingerprints, is then pre-stored for utilization
location fingerprints. The RSS demonstrates a characteristic in the online phase.
where the signal increases as the user approaches the beacon The online phase sees the smartphone acquiring the RSS
and decreases as the user moves away from it. The compre- vector and related MAC addresses from an uncharted posi-
hensive set of RSS fingerprints for all RPs within the indoor tion before uploading to the server. Subsequently, the server,
environment is denoted as the radio map. As depicted in Fig. in turn, compares the present RSS information with the stored
1, the BLE fingerprint positioning algorithm progresses two fingerprints in the database to estimate the location. The
stages—an offline stage followed by online one. This study system employs matching algorithms to compare the query
focuses explicitly on location fingerprinting, employing RSS with the radio map, returning the most similar fingerprint as
values garnered from BLE beacons. the estimated location.
During the offline phase, mobile communication devices, Assume a set of BLEs (B = BLE1 , BLE2 , . . . , BLEM )
such as smartphones with Bluetooth modules, gather RSS and RPs (RPi = RP1 , RP2 , . . . , RPN ), where M represents
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

Output

ReLU ReLU Sigmoid

yt
Generator

Ct-1 Output gate Ct

Label Cell state Forget gate Input gate
tanh Next cell state

Generated output
32 (LSTM)

16 (Dense)
Random
noise σ σ tanh σ
Ht-1 Ht
Hidden state Next hidden state
xt
Input
Real data
Discriminated

128 (LSTM)
16 (Dense)

64 (Dense)

FIGURE 3. LSTM cell.

output

Label
C. CGAN
In 2014, [52] introduced CGAN, which enhanced traditional
GANs by incorporating additional information into the en-
ReLU ReLU
Discriminator coded class labels. The class labels are fed to both the gen-
Sigmoid ReLU
erator and discriminator along with the prior noise and real
data, respectively. This conditioning of class labels allows
FIGURE 2. Discriminator and generator model. CGANs to generate data that are specific to each class. The
cost function is defined as

the total number of BLE beacons, and N indicates the total 

number of deployed RPs in the indoor environment. Let NR min max L(D, G) = Ex∼Pdata (x) log D x|y +
G D
be the number of offline RPs and MB be the number of BLEs.   (3)


The offline fingerprint RSSi at RPi is denoted as: Ex∼Pdata (x) log 1 − D G(z|y) ,

    where the term Ex∼Pdata (x) [log D(x|y)] indicates the discrim-
RSSi = M ACi1 , RSSi1 , ..., M ACiMB , RSSiMB (1) inator’s ability to correctly distinguish real data samples (x)
from generated data samples (G(z|y)) given the correspond-
where M ACij and RSSij represent the MAC address and ing class labels (y). The generator aims to minimize this
RSS value of the BLEj at RPi , respectively. Let Pi = term to ensure that the generated samples are indistinguish-
(xi , yi ) represent the 2D location coordinate of the i-th able from the real samples. Conversely, Ex∼Pdata (x) [log(1 −
RP, and t express the sampling amount of the k-th BLE, D(G(z|y)))] indicates how well the discriminator correctly
respectively. The RSS ofP the m-th BLE in the i-th RP can identifies generated samples as fake. The discriminator seeks
m t m,i
be defined by RP i
= i=1 RPi /t. The fingerprints of to maximize this term to improve its ability to differentiate
RPs in the location space can be represented by RPi = between real and generated data.
1 2 M T
[RP i
, RP i
, ..., RP i
] . Consequently, the fingerprints of RPs The losses of the discriminator and generator are expressed
at the grid point in a given location are denoted by: as follows:
 1 1 1

RP1 RP 2
··· RP N



RP2
RP2
··· 2 
RP L(D, θd ) = Ex∼Pdata (x) log D x|y; θd +
 1 2 N
R= . .. .. ..  (2) (4)
 ..

. . . 


M M M Ez∼Pz (z) log 1 − D G(z|y; θg )
RP 1
RP 2
··· RPN

In the online phase, a fresh RSS measurement denoted as 



RSe = [RS1 e , RS2 e , ..., RSMe ]T is obtained and concurrently L(D, θg ) = Ez∼Pz (z) log 1 − D G(z|y; θg ) (5)
compared with R. Subsequently, the positioning scheme is
employed further to determine the estimated coordinates of In (4), L(D, θd ) represents the discriminator loss, with θd
the localization device. denoting the weighted parameters adjusted during training.
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

Algorithm 1: Data Augmentation using CGAN- maximizing the generator loss (L(D, θg )) contributes to the
LSTM generation of highly realistic samples, making it challenging
Input : Random positions, RSS measurements from n for the discriminator to differentiate them from real data
BLE beacons with MAC addresses under the same condition y. Hence, the CGAN training aims
Output: Generated augmented dataset to minimize the discriminator loss (L(D, θd )) while maxi-
1 // Initialization
2 Initialize LSTM-based generator (G) and discriminator (D) mizing the generator loss (L(D, θg )), leading to a balanced
networks training process that results in the generation of high-quality
3 Define the categorical cross-entropy loss function and Adam samples by the generator.
optimizer for G and D, along with other hyperparameters During training, we use different hyperparameters (e.g.,
like learning rate and batch size learning rate, batch size, sequence length of LSTM) to fine-
4 for each entry in the dataset do
5 sum ← 0 tune our proposed model. The values of these parameters are
6 count ← number of RSS values in the current entry selected in a heuristic manner to improve the proposed model
7 for each RSS value, RSSi , in the current entry do during training.
8 sum ← sum + RSSi
9 end D. RSS AUGMENTATION MODELS WITH CGAN FOR
sum
10 RL ← count
11 ▷ Calculate the average RSS value in the current entry, RADIO-MAP CONSTRUCTION
where count denotes the total number of RSS in one Algorithm 1 presents the process of data augmentation us-
row ing a conditional generative adversarial network with long
12 return RL short-term memory (CGAN-LSTM). This algorithm aims
13 for each epoch (ep = 1 to epochtotal ) do to augment a given dataset containing random positions
14 RS = {RSS1 , RSS2 , . . . , RSSn }
15 ▷ Obtain real samples (RS) with their and the corresponding RSS measurements observed from n
corresponding conditional labels in batches BLE beacons with MAC addresses. The goal is to generate
16 N = Generator_N oise() additional data samples that can be used to improve the
17 F L = Random_F ake_Labels() performance of the radio-map-based fingerprinting.
18 F S = G(N, F L) The algorithm begins by initializing the LSTM-based gen-
19 ▷ Generate fake samples based on noise and fake
labels erator (G) and discriminator (D) networks. In addition, the
20 Authenticity_RS = D(RS) categorical cross-entropy loss function and Adam optimizer
21 ▷ Discriminator predicts the authenticity based on loss for both networks are defined. The algorithm sequen-
actual RSS samples tially processes each entry (row) within the dataset to fetch
22 Loss_RS_RL = Loss_f unction(RS, RL) the RSS value of each MAC address. It calculates the actual
23 Authenticity_F S = D(F S)
24 Loss_F S_F L = Loss_f unction(F S, F L), labels (RL) by averaging the RSS values of the respective
L(D, θd ) = RL + F L entries. Within each entry, the algorithm enters an epoch
25 WD = WD − D_optimizer(L(D, θd )). loop, updating the generator and discriminator networks it-
L(D, θg ) = Loss_f unction(F S, RL) eratively. During each epoch, real samples (RS) with their
26 WG = WG − G_optimizer(L(D, θg )) corresponding labels are obtained in batches, and noise (N )
27 if current epoch is equal to epochtotal then
28 Save newly generated augmented data as a new is generated to feed the generator. The generator (G) then
dataset generates fake samples (F S) by processing the noise and
29 end fake labels (F L) using G. Subsequently, the discriminator
30 end
31 end
(D) predicts the authenticity of the real samples (RS), and
the loss is computed according to the real samples (RS) and
their corresponding real labels (RL). Similarly, the authentic-
ity of the fake samples (F S) is predicted by D, and the loss
Ex∼Pdata (x) estimates the average over real data points x (in-
is calculated according to the fake samples (F S) and their
put data) sampled from the distribution Pdata . log D x|y; θd corresponding fake labels (F L).
is the logarithm of the probability assigned by the discrim- The discriminator loss (L(D, θd )) is formulated as the sum
inator D to the input x given the condition y and current of the real loss (RL) and the fake loss (F L). To update
weighted parameters θd . In the last part, which is common the weights of the discriminator (WD ), the discriminator
to both (4) and (5), the goal is to determine the logarithm of loss (L(D, θd )) is used in conjunction with the discrimina-
the probability assigned by the discriminator to the fake input tor optimizer (D_optimizer). Similarly, the generator loss
G(z|y; θg ) under a specific condition y. This logarithmic (L(G, θg )) is computed according to the fake samples (F S)
probability is averaged over all noise samples z from the and real labels (RL). To update the generator weights (WG ),
noise distribution Pz (z). the generator loss (L(G, θg )) is used in conjunction with the
In summary, the training of CGANs is focused on accu- generator optimizer (G_optimizer). The algorithm iterates
rately distinguishing between real and fake data under a spec- through multiple epochs, and when it reaches the specified
ified condition, denoted as y. To achieve this, it is essential epochtotal , the newly generated augmented data are saved as
to minimize the discriminator loss (L(D, θd )). Conversely, a new dataset.
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

IV. EXPERIMENTAL RESULTS AND ANALYSIS spanned 45 m in length and 3 m in width, whereas the
This section presents a comprehensive performance analysis laboratory covered an area of 12 m in length and 10 m in
of the proposed fingerprint augmentation method. First, we width. The laboratory environment included various types of
outline the experimental setup. Next, we discuss the perfor- furniture and obstacles. Individuals moved randomly within
mance of various CGAN models in comparison with state- the corridor, making the experimental settings closely resem-
of-the-art data-augmentation methods. ble real-life scenarios.
In developing a BLE fingerprinting-based positioning sys-
tem, the typical process involves conducting radio scans,
BLE 1
supplemented by ground-truth data collected by a user or
surveyor. In this context, the ground-truth information cor-
BLE 2 BLE 3
responding to the radio scans is known to the user and can
BLE 4
be regarded as labeled data. During the static data collection
BLE 5 BLE 6
BLE 7 phase, the user stands at specific points for 4 to 5 minutes,
BLE 8 BLE 9
BLE 10 and RSS data is collected at each predefined reference point
using a smartphone.
To create the radio map, we divided it into grids, each with
an average edge width of 2.0 m and height ranging from 1.5
to 2.0 m. Each grid point was assigned a unique fingerprint
containing the RSS values measured from n BLE beacons. In
total, 17 BLE beacons were deployed for these experiments,
with 10 placed in the corridor and 7 in the laboratory (as
(a) corridor (Scenario 1)
indicated by the purple circles in Fig. 4).
Each beacon was set to broadcast a signal every 300 ms,
BLE 3 BLE 6
and the transmission power level was set to +4 dBm to
BLE 1 BLE 4 BLE 7
provide good signal coverage. Data collection was performed
BLE 2 BLE 5 for 5 min to gather numerous RSS measurements at each
RP. The purpose of employing an extended scanning period
during the onsite fingerprint survey was to ensure a suffi-
cient number of signals for the construction of a fingerprint
database.

B. LOCATION ESTIMATION WITH WEIGHTED

K-NEAREST NEIGHBORS
(b) laboratory (Scenario 2) In this study, we applied a widely adopted weighted K-
nearest neighbors (WKNN) algorithm to evaluate the simi-
FIGURE 4. Layouts of two indoor environments. larity between RSS vectors [53], [54]. In the online phase,
the user’s location is estimated by evaluating the similarity
between the recently acquired fingerprint and the fingerprints
A. EXPERIMENTAL SETTINGS recorded in the database. The WKNN algorithm computes
We developed an Android application using the interface distances, typically Euclidean distances derived from RSS,
provided by Android Studio to measure the RSS of the by comparing each fingerprint in the database with the
beacon messages from multiple BLE devices. This applica- recorded fingerprint at location i of the point of interest.
tion, implemented in Java, is designed to run on Android Subsequently, it identifies k fingerprints with the smallest
smartphones. Our solution consists of a client application for distances, determining a match with the most probable po-
smartphones and a processing server running on a PC. No- sition of the device. The estimation is achieved through the
tably, the client application is optimized for Android 10 (API calculation of the weighted average of the positions of these
level 30) and can collect sensor readings, including signal k fingerprints, expressed as:
strengths from various beacons, along with their correspond-
ing timestamps. The extracted RSS values are expressed in k
X
dBm, denoting signal power levels in decibels (dB) relative pos
ˆ = wi pos
ˆ i (6)
to one milliwatt (mW). For BLE beacons whose signals were i=1

too weak to be detected by our Android app, we set the RSS 1/di
to a minimum value of -99 dBm by default. wi = Pk , (7)
During the offline stage of our experiment, we collected i=1 1/di

raw RSS data in typical indoor environments comprising a where pos

ˆ represents the estimated location, pos ˆ i represents
corridor and a laboratory (as shown in Fig. 4). The corridor the position at the i-th neighbor, and di signifies the distance
8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS


566(VWLPDWLRQ(UURU+HDWPDSIRU%/(
       
      

\P

    

       
         
    
         
[P 

(a) FCNN



566(VWLPDWLRQ(UURU+HDWPDSIRU%/( 
      
      

\P

    

        
        
    
          
[P

(b) LSTM

566(VWLPDWLRQ(UURU+HDWPDSIRU%/( 
      
      

\P

    

        
        
    
         
[P 

(c) BiLSTM


566(VWLPDWLRQ(UURU+HDWPDSIRU%/(
       
      

\P

    

       
         
    
         
[P 

(d) No augmentation
FIGURE 5. RSS estimation errors for BLE 4 at 40 RP locations in Scenario 1. Three CGAN-based prediction models (FCNN, LSTM, BiLSTM) are compared with
the case without any data augmentation method.

TABLE 2. Comparison of RSS estimation errors (in dBm) for CGAN models with data augmentation and without augmentation in two scenarios.

Test condition Models RSS estimation errors (dBm)

BLE1 BLE 2 BLE 3 BLE 4 BLE 5 BLE 6 BLE 7 BLE 8 BLE 9 BLE 10 Average
Corridor (Scenario 1) CGAN-FCNN 3.635 3.839 3.885 3.401 3.830 4.168 3.652 4.443 4.081 3.468 3.840
CGAN-LSTM 3.162 2.625 2.388 2.191 3.382 3.047 4.088 3.616 2.919 2.946 3.036
CGAN-BiLSTM 2.777 3.377 2.829 2.274 3.532 3.928 3.568 4.251 3.896 3.518 3.395
No augmentation 3.317 4.194 4.075 3.533 4.695 4.269 3.895 4.285 3.692 4.449 4.040
Laboratory (Scenario 2) CGAN-FCNN 3.200 4.125 3.287 2.721 3.632 2.946 3.589 - - - 3.357
CGAN-LSTM 2.860 3.333 2.735 2.716 3.636 2.650 2.726 - - - 2.951
CGAN-BiLSTM 2.943 3.349 2.948 3.065 3.552 2.662 2.973 - - - 3.070
No augmentation 3.066 4.268 3.679 3.726 3.219 3.570 3.914 - - - 3.634

between the measured RSS value at BLE of a point of interest and highest similarity, correspond to both offline and online
and the pre-recorded fingerprint of RSS at location i. (7) cal- RPs.
culates the weights (wi ) assigned to each neighbor based on
the inverse of their respective distances. Smaller distances are C. RSS ESTIMATION ACCURACY
accorded relatively larger weights, ensuring that the nearest To evaluate the accuracy of RSS estimation, we utilized a
neighbors, selected based on the smallest Euclidean distance set of randomly chosen RPs in which the fingerprints were
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS


566(VWLPDWLRQ(UURU+HDWPDSIRU%/( 
566(VWLPDWLRQ(UURU+HDWPDSIRU%/( 
 
      
  
     

     
  
     
\P

\P
  
   
  
    
       
 
   
   
 
             
[P [P
(a) FCNN (b) LSTM

566(VWLPDWLRQ(UURU+HDWPDSIRU%/( 
566(VWLPDWLRQ(UURU+HDWPDSIRU%/( 
  
     
  
      
     
  
      
\P

\P
 
   
 
     
       
 
   
   
 
             
[P [P
(c) BiLSTM (d) No augmentation
FIGURE 6. RSS estimation errors for BLE 4 at 25 RP locations in Scenario 2. Three CGAN-based prediction models (FCNN, LSTM, BiLSTM) are compared with
the case without any data augmentation method.

measured. Subsequently, we compared the predicted RSS 1 and 2, respectively. Additionally, LSTM exhibited reduc-
values (selected from 500 newly generated data samples) at tions of 20.94% and 10.57% in the average RSS estimation
these RPs with the corresponding ground-truth values. The error compared with FCNN and BiLSTM, respectively, in
localization performance was quantified using the following Scenario 1. In Scenario 2, LSTM exhibited error reductions
equation: of 12.09% and 3.88%, respectively. The RSS estimation er-
rors of different deep learning models as a data augmentation
ei = ||Rg − Re || , (8) method showed better performance in RSS prediction than
those without data augmentation approaches. For Scenario 1
where ei represents the error in the RSS estimation error, Rg and Scenario 2, as shown in Table 2, the case without a data
represents the ground-truth RSS value, and Re represents the augmentation method produced the worst RSS estimation
estimated RSS value. errors of 4.040 and 3.634 dBm, respectively, compared to the
The collected RSS values in the experimental environ- case with data augmentation methods.
ments were compared with CGAN-estimated RSS values
based on three different models (fully convolutional neural D. PERFORMANCE EVALUATION AND COMPARISON
network (FCNN), LSTM, and BiLSTM) employed for data OF LOCALIZATION
augmentation, as well as the case without any data augmen- Using data augmentation, we employed CGAN-LSTM to
tation method. The average RSS errors in Scenarios 1 and 2 generate radio maps for three BLE devices (BLE 2, BLE 4,
are presented in Table 2. As an example, for BLE 4, Figs. 5 and BLE 5) in Scenario 2, as illustrated in Fig. 7. The signal
and 6 present the RSS estimation errors at the locations of 40 map reconstructed after data augmentation using the pro-
RPs for Scenario 1 and 25 RPs for Scenario 2, respectively. posed method captured detailed RSS expressions, in contrast
As shown in Table 2, the CGAN-LSTM model outper- to the signal without data augmentation (see Fig. 8), which
formed the other models with regard to RSS prediction, with appeared excessively rough and failed to reflect normal signal
RSS estimation errors of 3.036 and 2.951 dBm for Scenarios characteristics.
10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

 %/(  %/(  %/(

    
   
    
   
  
     
  
     
  
\P

\P

\P
  
     
  
     
  
   


  
   
    
  










































































[P [P [P
(a) BLE 2 (b) BLE 4 (c) BLE 5
FIGURE 7. Heatmaps of three BLEs generated by the CGAN-LSTM data-augmentation model for Scenario 2.


%/(  
%/(  
%/( 

  

   
 

    

\P

\P

\P
 

    

  
  
  

     

                    
[P [P [P
(a) BLE 2 (b) BLE 4 (c) BLE 5
FIGURE 8. Heatmaps of three BLEs without data augmentation for Scenario 2.

1 .0 1 .0

0 .9 0 .9 A v e ra g e e rro r = 1 .9 5 6 m
A v e ra g e e rro r = 1 .0 6 5 m 9 0 % e rro r = 3 .7 0 3 m
0 .8 9 0 % e rro r = 1 .6 4 9 m 0 .8
A v e ra g e e rro r = 2 .0 2 7 m
0 .7 0 .7 9 0 % e rro r = 3 .7 6 6 m
A v e ra g e e rro r = 1 .0 8 5 m
0 .6 9 0 % e rro r = 1 .8 7 1 m 0 .6
A v e ra g e e rro r = 2 .2 2 5 m
9 0 % e rro r = 4 .0 2 3 m
0 .5 A v e ra g e e rro r = 1 .1 4 9 m 0 .5
C D F

C D F

9 0 % e rro r = 1 .9 2 7 m
0 .4 0 .4
A v e ra g e e rro r = 1 .2 3 9 m
N o a u g m e n ta tio n N o a u g m e n ta tio n
0 .3 0 .3 A v e ra g e e rro r = 2 .3 3 3 m
9 0 % e rro r = 2 .1 9 9 m L I 9 0 % e rro r = 4 .4 2 4 m L I
0 .2 ID W 0 .2 ID W
A v e ra g e e rro r = 1 .2 6 4 m G P R A v e ra g e e rro r = 2 .3 3 9 m G P R
0 .1 0 .1
9 0 % e rro r = 2 .2 9 8 m C G A N - L S T M 9 0 % e rro r = 4 .1 5 5 m C G A N - L S T M
0 .0 0 .0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 7 8 9
L o c a liz a tio n e rro r (m ) L o c a liz a tio n e rro r (m )
(a) Scenario 1 (b) Scenario 2

FIGURE 9. CDFs for two scenarios with 500 augmented data points.

The overall cumulative error of our method, along with 1 and 2, respectively, corresponding to improvements of at
comparisons with state-of-the-art methods for each scenario, least 15.74% compared with fingerprinting without the data-
is presented in Fig. 9. To ensure a fair comparison, we augmentation method. Moreover, in Scenario 1, the proposed
augmented the dataset to 500 points, encompassing the method outperformed GPR, IDW, and LI by 1.84%, 7.31%,
entire area for both the proposed approach and state-of- and 14.04%, respectively. Across all scenarios, the proposed
the-art data augmentation methods. The proposed approach approach reduced the 90th-percentile errors to < 3.80 m.
achieved mean accuracies of 1.065 and 1.956 m in Scenarios
We noticed an increase in the localization accuracy for all
VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

 5HDO3RVLWLRQV *HQHUDWHG3RVLWLRQV  5HDO3RVLWLRQV *HQHUDWHG3RVLWLRQV

 
 
\P

\P
 
 
 
 
                   
[P [P
(a) 500 data augmentations (b) 750 data augmentations
 5HDO3RVLWLRQV *HQHUDWHG3RVLWLRQV


\P





         
[P
(c) 1000 data augmentations
FIGURE 10. Generated positions in Scenario 1 with the corresponding RSS for data augmentation.

5HDO3RVLWLRQV *HQHUDWHG3RVLWLRQV 5HDO3RVLWLRQV *HQHUDWHG3RVLWLRQV 5HDO3RVLWLRQV *HQHUDWHG3RVLWLRQV

  

  

  
\P

\P

\P
  

  

  
                    
[P [P [P
(a) 500 data augmentations (b) 750 data augmentations (c) 1000 data augmentations

FIGURE 11. Generated positions in Scenario 2 with the corresponding RSS for data augmentation.

the augmented datasets compared with classical localization E. DISCUSSION

using a dataset limited solely to real data. The augmented Data augmentation plays a crucial role in improving the
data presented in Figs. 10 and 11 consist of newly generated performance of the fingerprinting models for BLE indoor
positions and RSS values of BLE beacons. We generated localization. In this study, we compared the performance of
500, 750, and 1000 data points to cover the entire area in three different data augmentation techniques: CGAN-LSTM,
Scenarios 1 and 2, respectively. The enhancements in the CGAN-BiLSTM, and CGAN-FCNN, in the context of BLE
localization accuracy varied by up to 10% across 250–1000 fingerprinting. Our results indicated that the CGAN-LSTM
data augmentations. The highest localization accuracy was model outperformed the other two models with regard to
achieved when the dataset was augmented with 750 and 500 localization accuracy.
data samples in Scenarios 1 and 2, respectively, resulting in
The superiority of the CGAN-LSTM model is attributed to
average errors of 1.058 and 1.956 m, as shown in Fig. 12. Sur-
its ability to capture both the spatial dependencies of finger-
prisingly, the generation of additional fake data beyond this
print data and the temporal dynamics of signal variations. By
point did not lead to further improvements. Notably, when the
combining the GAN and LSTM architectures, the CGAN-
number of generated samples reached 1000, the localization
LSTM model effectively learns the complex relationships
accuracy reached saturation, and no further improvement was
between the input BLE fingerprints and the corresponding
observed.
locations. The GAN component of the model allows the
generation of realistic and diverse synthetic fingerprints,
which enhances the training data and reduces the degree of
12 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

1 .6
2 .5
2 .3 3 9
1 .4 2 .1 8 8
1 .2 6 4 2 .0 6 8
1 .9 5 6 2 .0 0 4
1 .2 1 .1 7 2 2 .0
1 .1 2 3
1 .0 6 5 1 .0 5 8
1 .0
A v e ra g e e rro r (m )

A v e ra g e e rro r (m )
1 .5
0 .8

0 .6 1 .0

0 .4
0 .5
0 .2

0 .0 0 .0
0 2 5 0 5 0 0 7 5 0 1 0 0 0 0 2 5 0 5 0 0 7 5 0 1 0 0 0
N u m b e r o f g e n e ra te d s a m p le s p e r lo c a tio n N u m b e r o f g e n e ra te d s a m p le s p e r lo c a tio n
(a) Scenario 1 (b) Scenario 2

FIGURE 12. Effect of the amount of generated data on the localization accuracy.

0 .9 0 .9

0 .8
0 .8

0 .7
A c c u ra c y

A c c u ra c y

0 .7
0 .6

0 .6
0 .5

0 .5
T ra in in g a c c u ra c y 0 .4
T ra in in g a c c u ra c y
V a lid a tio n a c u ra c y
V a lid a tio n a c c u ra c y
0 .4 0 .3
0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0
E p o c h E p o c h
(a) Scenario 1 (b) Scenario 2

FIGURE 13. Localization model accuracy.

overfitting. the expected higher localization accuracy similar to CGAN-

LSTM.
Furthermore, the LSTM component of the CGAN-LSTM
In contrast, CGAN-FCNN models, which do not incor-
model allows it to capture the temporal dependencies present
porate the LSTM architecture, may struggle to capture the
in the BLE signal variations. This is particularly important in
temporal dynamics of the BLE signals. The CGAN-FCNN
indoor environments, where signal characteristics can change
model relies solely on the feedforward nature of the fully
over time owing to various factors such as human presence
connected neural network, which may not effectively capture
and environmental conditions. The LSTM component effec-
the sequential nature of the BLE signal variations.
tively learns sequential patterns based on forward contexts
and long-term dependencies in the data. Moreover, LSTM The proposed BLE-based data augmentation using
leads to high localization accuracy because of maintaining CGAN-LSTM outperformed the other data-augmentation
sequential training ability in one direction within a few methods. LI data augmentation assumes linear relationships
number of training parameters for small number of data in between two sample points and cannot capture the intri-
our data set. Although the CGAN-BiLSTM model incor- cate nonlinearities present in indoor environments. Similarly,
porates bidirectional processing to capture both past and IDW data augmentation was affected by the relative weights
future contexts, it may face limitations in modeling long-term and spatial correlations between close points. GPR per-
dependencies. Because of this nature, the training complex- formed as well as the second-best data-augmentation method
ity becomes higher during training in modeling long-term by inferring the posterior RSS mean and variance to build a
dependencies for our sequential dataset; when gradients be- complete radio map.
come too small, the training’s parameter updates very slowly As shown in Fig. 13, the training and validation accuracies
for distant time steps in both the forward and backward increased with an increase in the number of epochs in both
directions. As a result, the training outcome does not achieve scenarios for the CGAN-LSTM model, indicating that the
VOLUME 4, 2016 13

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

model achieved effective learning and generalization to new [3] M. H. Azaddel, M. A. Nourian, K. ShahHosseini, S. A. Junoh, and
data while avoiding overfitting. A. Akbari, “Spotter: A novel asynchronous and independent wifi and ble
fusion method based on particle filter for indoor positioning,” Internet of
The average localization error can be reduced by aug- Things, p. 100967, 2023.
menting RPs during offline fingerprinting construction of the [4] Y. Zhuang, C. Zhang, J. Huai, Y. Li, L. Chen, and R. Chen, “Bluetooth
radio map. By increasing the number of augmented RPs, the localization technology: Principles, applications, and future trends,” IEEE
Internet of Things Journal, vol. 9, no. 23, pp. 23 506–23 524, 2022.
localization performance was improved, i.e., the localization [5] F. Zafari, A. Gkelias, and K. K. Leung, “A survey of indoor localization
accuracy was increased. However, adding too many points systems and technologies,” IEEE Communications Surveys & Tutorials,
may introduce interference nodes and reduce the localization vol. 21, no. 3, pp. 2568–2599, 2019.
[6] C. Laoudias, A. Moreira, S. Kim, S. Lee, L. Wirola, and C. Fischione,
accuracy, as shown in Fig. 12. This suggests that an optimal “A survey of enabling technologies for network localization, tracking, and
number of RPs may exist for a specific indoor environment. navigation,” IEEE Communications Surveys & Tutorials, vol. 20, no. 4,
Overall, the ability of the proposed CGAN-LSTM model pp. 3607–3644, 2018.
to capture both spatial and temporal dependencies and gen- [7] J. Torres-Sospedra, D. P. Q. Gaibor, J. Nurmi, Y. Koucheryavy, E. S.
Lohan, and J. Huerta, “Scalable and efficient clustering for fingerprint-
erate diverse training data through the GAN component based positioning,” IEEE Internet of Things Journal, vol. 10, no. 4, pp.
contributed to its superior performance in BLE fingerprinting 3484–3499, 2022.
data augmentation. These findings highlight the importance [8] J. Yang, “Indoor localization system using dual-frequency bands and
interpolation algorithm,” IEEE Internet of Things Journal, vol. 7, no. 11,
of considering architectural choices and integrating relevant pp. 11 183–11 194, 2020.
techniques when designing data-augmentation models for [9] Q. Li, H. Qu, Z. Liu, N. Zhou, W. Sun, S. Sigg, and J. Li, “Af-dcgan:
BLE indoor localization. Amplitude feature deep convolutional gan for fingerprint construction in
indoor localization systems,” IEEE Transactions on Emerging Topics in
Computational Intelligence, vol. 5, no. 3, pp. 468–480, 2019.
V. CONCLUSIONS AND FUTURE WORK [10] W. Sun, M. Xue, H. Yu, H. Tang, and A. Lin, “Augmentation of fingerprints
In this study, we demonstrated that synthetic data can in- for indoor wifi localization based on gaussian process regression,” IEEE
crease the average accuracy of fingerprint-based localization Transactions on Vehicular Technology, vol. 67, no. 11, pp. 10 896–10 905,
2018.
in a deep-learning context, where data collection is time- [11] F. Ma, Y. Li, S. Ni, S.-L. Huang, and L. Zhang, “Data augmentation for
consuming and expensive. Specifically, we propose the use of audio-visual emotion recognition with an efficient multimodal conditional
a CGAN to generate synthetic data. To investigate the char- gan,” Applied Sciences, vol. 12, no. 1, p. 527, 2022.
[12] K. Wang, C. Gou, Y. Duan, Y. Lin, X. Zheng, and F.-Y. Wang, “Generative
acteristics of RSS fingerprints, we assessed CGAN-based adversarial networks: introduction and outlook,” IEEE/CAA Journal of
RSS data augmentation using different deep-learning models. Automatica Sinica, vol. 4, no. 4, pp. 588–598, 2017.
Experimental results indicated that CGAN-LSTM achieved [13] H. Navidan, P. F. Moshiri, M. Nabati, R. Shahbazian, S. A. Ghorashi,
V. Shah-Mansouri, and D. Windridge, “Generative adversarial networks
average estimation errors of 3.036 and 2.951 dBm in Scenar- (gans) in networking: A comprehensive survey & evaluation,” Computer
ios 1 and 2, respectively, outperforming the CGAN-FCNN Networks, vol. 194, p. 108149, 2021.
and CGAN-BiLSTM models. Additionally, the proposed al- [14] M. Patel, X. Wang, and S. Mao, “Data augmentation with conditional gan
for automatic modulation classification,” in Proceedings of the 2nd ACM
gorithm outperformed the non-augmentation approach by at Workshop on wireless security and machine learning, 2020, pp. 31–36.
least 15.74%. Compared with the baseline data-augmentation [15] H. Zou, M. Jin, H. Jiang, L. Xie, and C. J. Spanos, “Winips: Wifi-based
method, the proposed CGAN-LSTM system reduced the lo- non-intrusive indoor positioning system with online radio map construc-
calization error by up to 14.04% in both scenarios. Therefore, tion and adaptation,” IEEE Transactions on Wireless Communications,
vol. 16, no. 12, pp. 8118–8130, 2017.
given the constraints of limited fingerprints, data augmen- [16] R. K. Yadav, B. Bhattarai, H.-S. Gang, and J.-Y. Pyun, “Trusted k nearest
tation is a viable means of increasing the average accuracy bayesian estimation for indoor positioning system,” IEEE Access, vol. 7,
of BLE-based indoor localization while alleviating the bur- pp. 51 484–51 498, 2019.
[17] T. Otim, A. Bahillo, L. E. Díez, P. Lopez-Iturri, and F. Falcone, “Towards
densome demands of site surveys. In the future, we aim to sub-meter level uwb indoor localization using body wearable sensors,”
develop further into the following directions. For example, IEEE Access, vol. 8, pp. 178 886–178 899, 2020.
exploring alternative machine learning approaches for data [18] S. Sadowski and P. Spachos, “Rssi-based indoor localization with the
internet of things,” IEEE access, vol. 6, pp. 30 149–30 161, 2018.
augmentation, including advanced methods for comparative [19] X. Li, “Cellular base station assisted indoor positioning,” IEEE Transac-
analysis, holds promise. Additionally, scrutinizing the per- tions on Aerospace and Electronic Systems, vol. 55, no. 2, pp. 592–606,
formance of lazy learners and their viability as fingerprints 2018.
[20] M. Zhang, Z. Fan, R. Shibasaki, and X. Song, “Domain adversarial
for data augmentation while considering the dynamic indoor
graph convolutional network based on rssi and crowdsensing for indoor
environment that may require periodic fingerprint updates localization,” IEEE Internet of Things Journal, vol. 10, no. 15, pp. 13 662–
could yield valuable insights for the field of indoor localiza- 13 672, 2023.
tion. [21] K. Kim and J. Lee, “Adaptive scheme of denoising autoencoder for
estimating indoor localization based on rssi analytics in ble environment,”
Sensors, vol. 23, no. 12, p. 5544, 2023.
REFERENCES [22] S. A. Junoh, S. Subedi, and J.-Y. Pyun, “Smartphone-based indoor naviga-
[1] J. Torres-Sospedra, P. Richter, A. Moreira, G. M. Mendoza-Silva, E. S. tion system using particle filter and map-constraints,” in The 9th Interna-
Lohan, S. Trilles, M. Matey-Sanz, and J. Huerta, “A comprehensive and tional Conference on Smart Media and Applications, 2020, pp. 354–357.
reproducible comparison of clustering and optimization rules in wi-fi [23] Z. Yu, Z. Chaczko, and J. Shi, “A novel algorithm modelling for uwb
fingerprinting,” IEEE Transactions on Mobile Computing, vol. 21, no. 3, localization accuracy in remote sensing,” Remote Sensing, vol. 14, no. 19,
pp. 769–782, 2020. p. 4902, 2022.
[2] S. A. Junoh, S. Subedi, and J.-Y. Pyun, “Floor map-aware particle filtering [24] A. S. Yaro, F. Maly, and P. Prazak, “A survey of the performance-
based indoor navigation system,” IEEE Access, vol. 9, pp. 114 179– limiting factors of a 2-dimensional rss fingerprinting-based indoor wireless
114 191, 2021. localization system,” Sensors, vol. 23, no. 5, p. 2545, 2023.

14 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3368449

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

[25] F. J. Aranda, F. Parralejo, F. J. Álvarez, and J. A. Paredes, “Performance [48] A. Nessa, B. Adhikari, F. Hussain, and X. N. Fernando, “A survey of ma-
analysis of fingerprinting indoor positioning methods with ble,” Expert chine learning for indoor positioning,” IEEE access, vol. 8, pp. 214 945–
Systems with Applications, vol. 202, p. 117095, 2022. 214 965, 2020.
[26] A. Mansour, J. Ye, Y. Li, H. Luo, J. Wang, D. Weng, and W. Chen, [49] S. A. Junoh and J.-Y. Pyun, “Enhancing indoor localization with semi-
“Everywhere: A framework for ubiquitous indoor localization,” IEEE crowdsourced fingerprinting and gan-based data augmentation,” IEEE
Internet of Things Journal, 2022. Internet of Things Journal, pp. 1–1, 2023.
[27] Y. Yu, R. Chen, L. Chen, W. Li, Y. Wu, and H. Zhou, “Autonomous 3d [50] F. Parralejo, F. J. Aranda, J. A. Paredes, F. J. Alvarez, and J. Morera,
indoor localization based on crowdsourced wi-fi fingerprinting and mems “Comparative study of different ble fingerprint reconstruction techniques,”
sensors,” IEEE Sensors Journal, vol. 22, no. 6, pp. 5248–5259, 2021. in 2021 International Conference on Indoor Positioning and Indoor Navi-
[28] T. Li, D. Han, Y. Chen, R. Zhang, Y. Zhang, and T. Hedgpeth, “Indoor- gation (IPIN). IEEE, 2021, pp. 1–8.
waze: A crowdsourcing-based context-aware indoor navigation system,” [51] S. M. Sultan, M. Waleed, J.-Y. Pyun, and T.-W. Um, “Energy conservation
IEEE Transactions on Wireless Communications, vol. 19, no. 8, pp. 5461– for internet of things tracking applications using deep reinforcement
5472, 2020. learning,” Sensors, vol. 21, no. 9, p. 3261, 2021.
[29] Y. Yu, W. Shi, R. Chen, L. Chen, S. Bao, and P. Chen, “Map-assisted [52] M. Mirza and S. Osindero, “Conditional generative adversarial nets,” arXiv
seamless localization using crowdsourced trajectories data and bi-lstm preprint arXiv:1411.1784, 2014.
based quality control criteria,” IEEE Sensors Journal, vol. 22, no. 16, pp. [53] X. Du, X. Liao, M. Liu, and Z. Gao, “Crcloc: A crowdsourcing-based
16 481–16 491, 2022. radio map construction method for wifi fingerprinting localization,” IEEE
[30] Y. Zhao, C. Liu, K. Zhu, S. Zhang, and J. Wu, “Gsmac: Gan-based Internet of Things Journal, vol. 9, no. 14, pp. 12 364–12 377, 2021.
signal map construction with active crowdsourcing,” IEEE Transactions [54] Y. Lin, K. Yu, L. Hao, J. Wang, and J. Bu, “An indoor wi-fi localization
on Mobile Computing, vol. 22, no. 4, pp. 2190–2204, 2021. algorithm using ranging model constructed with transformed rssi and bp
[31] B. Shin, J. H. Lee, C. Yu, H. Kyung, and T. Lee, “Simultaneous localization neural network,” IEEE Transactions on Communications, vol. 70, no. 3,
and mapping for pedestrians using radio frequency signals,” IEEE Sensors pp. 2163–2177, 2022.
Journal, vol. 22, no. 13, pp. 13 497–13 507, 2022.
[32] B. Gao, F. Yang, N. Cui, K. Xiong, Y. Lu, and Y. Wang, “A federated
learning framework for fingerprinting-based indoor localization in multi- SUHARDI AZLIY JUNOH received his Bache-
building and multifloor environments,” IEEE Internet of Things Journal, lor of Engineering in Electronics (Telecommuni-
vol. 10, no. 3, pp. 2615–2629, 2022. cations) from Multimedia University and Master
[33] Z. Xing and J. Chen, “Constructing indoor region-based radio map without of Electrical Engineering from UTHM, Malaysia.
location labels,” arXiv preprint arXiv:2308.16759, 2023. He is currently pursuing a PhD in Information and
[34] W. Njima, A. Bazzi, and M. Chafii, “Dnn-based indoor localization under Communication Engineering from Chosun Uni-
limited dataset using gans and semi-supervised learning,” IEEE Access, versity in South Korea. He has also worked at
vol. 10, pp. 69 896–69 909, 2022. Infineon Technologies as an engineer. His current
[35] H. Zou, C.-L. Chen, M. Li, J. Yang, Y. Zhou, L. Xie, and C. J. Spanos, research interests include indoor positioning and
“Adversarial learning-enabled automatic wifi indoor radio map construc- navigation, the Internet of Things, mobile comput-
tion and adaptation with mobile robot,” IEEE Internet of Things Journal, ing, and wireless communication systems.
vol. 7, no. 8, pp. 6946–6954, 2020.
[36] J. Zhang, F. Wu, B. Wei, Q. Zhang, H. Huang, S. W. Shah, and J. Cheng,
“Data augmentation and dense-lstm for human activity recognition using JAE-YOUNG PYUN received a B.S. degree from
wifi signal,” IEEE Internet of Things Journal, vol. 8, no. 6, pp. 4628–4641, Chosun University, M.S. degree from Chonnam
2020. University, and Ph.D. degree from Korea Univer-
[37] X. Chen, H. Li, C. Zhou, X. Liu, D. Wu, and G. Dudek, “Fidora: Robust sity, Korea, in Electronics Engineering in 1997,
wifi-based indoor localization via unsupervised domain adaptation,” IEEE 1999, and 2003, respectively. From 2003 to 2004,
Internet of Things Journal, vol. 9, no. 12, pp. 9872–9888, 2022. he worked with Samsung Electronics, where he
[38] Y. Tao and L. Zhao, “A novel system for wifi radio map automatic adapta-
was involved in the research and development of
tion and indoor positioning,” IEEE Transactions on Vehicular Technology,
mobile phone communication systems. In 2004,
vol. 67, no. 11, pp. 10 683–10 692, 2018.
he joined the Department of Information and
[39] S. A. Junoh, S. Subedi, and J.-Y. Pyun, “Crowdsourcing landmark-assisted
localization with deep learning,” Future Generation Computer Systems, Communication Engineering at Chosun Univer-
vol. 144, pp. 256–270, 2023. sity, where he is currently a professor. In 2010, he was a Visiting Researcher
[40] J. Talvitie, M. Renfors, and E. S. Lohan, “Distance-based interpolation with the University of Washington, USA. He has conducted numerous
and extrapolation methods for rss-based localization with indoor wireless research projects in the fields of IoT protocols and applications. He has
signals,” IEEE transactions on vehicular technology, vol. 64, no. 4, pp. contributed 210 articles to more than 70 international journals. He also
1340–1353, 2015. holds more than 25 patents for wireless communication, Internet of Things
[41] Y. Dong, G. He, T. Arslan, Y. Yang, and Y. Ma, “Crowdsourced indoor protocols, positioning systems, IR-UWB security, and multimedia commu-
positioning with scalable wifi augmentation,” Sensors, vol. 23, no. 8, p. nication. His current research interests include machine learning (ML), deep
4095, 2023. learning (DL), IoT protocol design, IoT applications with indoor positioning
[42] Y. Chen, G. Li, Y. Tan, and G. Zhang, “Graph-based radio fingerprint systems (IPS), real-time location systems (RTLS), and UWB radar. He has
augmentation for deep-learning-based indoor localization,” IEEE Sensors been a member of IEEE, IEICE, IEEK, and KICS since 2004.
Journal, vol. 23, no. 6, pp. 6074–6084, 2023.
[43] Z. Zeng, Y. Xu, Z. Xie, J. Wan, W. Wu, and W. Dai, “Rg-gcn: A random
graph based on graph convolution network for point cloud semantic
segmentation,” Remote Sensing, vol. 14, no. 16, p. 4055, 2022.
[44] X. Guo, N. Ansari, F. Hu, Y. Shao, N. R. Elikplim, and L. Li, “A survey
on fusion-based indoor positioning,” IEEE Communications Surveys &
Tutorials, vol. 22, no. 1, pp. 566–594, 2019.
[45] P. C. Ng, P. Spachos, J. She, and K. Plataniotis, “A kernel method to non-
linear location estimation with rss-based fingerprint,” IEEE Transactions
on Mobile Computing, 2022.
[46] P. Roy and C. Chowdhury, “A survey of machine learning techniques
for indoor localization and navigation systems,” Journal of Intelligent &
Robotic Systems, vol. 101, no. 3, p. 63, 2021.
[47] T. Yang, A. Cabani, and H. Chafouk, “A survey of recent indoor local-
ization scenarios and methodologies,” Sensors, vol. 21, no. 23, p. 8086,
2021.

VOLUME 4, 2016 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see [Link]