The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

[Link]

ORIGINAL ARTICLE

Improving operations through a lean AI paradigm: a view

to an AI‑aided lean manufacturing via versatile convolutional neural
network
Mohammad Shahin1 · Mazdak Maghanaki1 · Ali Hosseinzadeh1 · F. Frank Chen1

Received: 13 February 2024 / Accepted: 24 May 2024 / Published online: 2 July 2024
© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024

Abstract
Integrating Lean Manufacturing tools with artificial intelligence (AI) is emerging as a revolutionary approach to optimize
production processes, reduce waste, and enhance efficiency. Traditional Lean practices focus on waste reduction and process
improvement, often relying on human expertise for problem identification and resolution. AI algorithms, on the other hand,
excel in pattern recognition, data analysis, and decision-making. Lean tools and AI can offer more precise, data-driven
solutions for common manufacturing challenges when integrated. AI algorithms can automate and refine Lean techniques
like value stream mapping, Kanban, and 5S by providing real-time, actionable insights drawn from big data. This fusion
of Lean and AI aids in predictive maintenance, quality control, and optimization, enhancing the efficiency and responsive-
ness of the manufacturing process. Moreover, AI’s capability for machine learning allows the system to adapt and improve
autonomously over time, further aligning with Lean’s continuous improvement ethos. Seven case studies were conducted to
show how this alignment might aid Lean Manufacturing. However, successful implementation necessitates overcoming data
quality and algorithmic bias challenges. Despite these hurdles, integrating Lean tools and AI can redefine best practices in
manufacturing, setting new standards for operational excellence.

Keywords Lean Manufacturing · Artificial intelligence · Machine learning · Industry 4.0 · Computer vision · Convolutional
neural network

1 Introduction is not just on efficiency but also on operational effective-

ness. Intelligent automation stands at the forefront of this
More than two decades ago, the integration of automation new wave powered by artificial intelligence (AI). While
into Lean Manufacturing, six sigma methodologies was in its AI’s impact is often associated with automated vehicles
initial phase. The focus was on dividing processes, eliminat- and robotics, its role in streamlining internal processes is
ing waste, and streamlining operations. The ultimate goals equally transformative. This new form of intelligent auto-
were cost reduction and waste minimization. As technology mation extends beyond cost and waste reduction to include
began permeating business operations, more sophisticated predictive insights and smart decision-making. This merging
forms of automation were introduced. Many manual process was believed to propel businesses towards enhanced effi-
steps could now be bypassed or expedited through technol- ciency and competitiveness. This paper is divided into the
ogy, leading to the same core outcomes of cost and waste following seven sections:
reduction. Today, while technology and business processes
remain constant, the evolving nature of data is driving what • Section 1: Introduction section
can be termed intelligent automation. The emphasis now • Section 2: The rise of intelligent automation discusses
AI’s components and inner workings
• Section 3: The integration of Lean Manufacturing and AI
* F. Frank Chen section that discusses how AI can impact each Lean tool
[Link]@[Link] • Section 4: Convolutional neural network (CNN) section
1
Mechanical Engineering Department, The University that discusses the history and development of CNNs
of Texas at San Antonio, San Antonio, USA

Vol.:(0123456789)
5344 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

• Section 5: Case studies section that showcases the appli- shift brings forth a compelling phenomenon known as the
cation of CNNs in different areas of enterprise operations “data flywheel” effect [13]; see Fig. 1 for an illustration.
• Section 6: Opportunities, challenges, risks, and cost The concept revolves around acquiring more data through
• Section 7: Current and future trends in AI internal and external customers’ feedback, which allows for
• Section 8: Conclusions and final thoughts. the development of superior algorithms and, ultimately, a
more exceptional product. As this cycle repeats, companies
The fervor surrounding AI and machine learning (ML) with a swifter learning capability will continuously enhance
has surged, captivating diverse industries such as marketing their offerings, attracting an ever-growing customer base and
[1], healthcare [2], manufacturing [3], transportation [4], accumulating abundant valuable data. Consequently, their
finance [5], and retail [6]. They are harnessing AI advance- learning capacity accelerates further, creating a self-perpetu-
ments and applications to enhance productivity and perfor- ating loop of progress and success. In the ever-evolving busi-
mance. Economists have acclaimed AI as a pivotal facilitator ness landscape, companies must shift their focus from solely
of the “Fourth Industrial Revolution” [7]. Given this tremen- relying on forecasting and planning to embracing discovery
dous anticipation, business leaders are now contemplating and adaptation [14]. The dynamic and uncertain nature of
how AI can ensure their survival during this revolution and the business environment demands an agile approach, where
propel them ahead of their peers [8]. Current predictions the ability to explore new possibilities and swiftly adapt to
suggest that AI will contribute a staggering $15.7 trillion to changing circumstances becomes paramount [15]. Success
the global economy by 2030, presenting a colossal opportu- lies in predicting the future and actively uncovering oppor-
nity for businesses to reap the benefits of investing in AI [9]. tunities, and seamlessly adjusting strategies to navigate the
In fact, the MIT Sloan Management Review’s 2017 Artificial unpredictable terrain [16]. By prioritizing discovery and
Intelligence Global Executive Study and Research Project adaptation, companies can unlock the true potential for
[10] revealed that a striking 85% of executives believe that growth and resilience in the face of uncertainty [17].
AI will confer a competitive advantage to their businesses,
enabling them to gain or sustain supremacy in their respec-
tive fields. 2 The rise of intelligent automation
The amalgamation of AI, sensors, and digital platforms
has already opened doors to enhanced learning opportunities 2.1 Artificial intelligence (AI)
[11]. However, the differentiating factor that will separate
thriving manufacturing industries from their counterparts is AI is a broad term encompassing various technologies
their ability to compete on learning speed. Enterprises that capable of simulating human intelligence. It represents
wholeheartedly embrace AI will gain a remarkable advan- the capacity of computers or machines to think and learn,
tage, as they can rapidly test, learn, and iterate, raising the aiming to create “smart” systems that can function inde-
bar for learning in the competitive arena [12]. This paradigm pendently without relying on explicit commands. John

Fig. 1  The data flywheel effect

The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5345

McCarthy coined the term in 1955 [18]. It involves the con- construction of models for data analysis. The underlying
struction of machines or computer programs that can ana- principle of ML is that a computer can learn from the data
lyze data, learn from diverse signals, and make intelligent it analyzes by identifying patterns, ultimately enabling it to
decisions to accomplish specific goals or desired outcomes. make decisions independent of human intervention [27]. It
AI focuses on the scientific and engineering aspects of ena- operated by initially having the system employ probability
bling computers to behave in ways previously believed to to make decisions or predictions based on available data,
require human intelligence [19]. However, AI goes beyond using feedback loops to ascertain the accuracy of its predic-
mere technical definitions. It encompasses the concept that tions. The projections become increasingly precise with each
machines and computers can successfully perform tasks that iteration, and the system grows more intelligent and adept
traditionally necessitate human intelligence [20]. The cur- through experience [28]. Figure 2 shows the three pillars of
rent manifestation of AI is witnessed in human-AI interac- AI, and Table 1 describes each one.
tive devices and services [21] like Google Home, Siri, and By analyzing vast amounts of data, quantum computers
Alexa. It is evident in ML-powered video prediction systems give AI systems the invaluable feedback necessary for per-
that fuel streaming platforms like Netflix, Amazon Prime, formance improvement. Moreover, they surpass traditional
and Hulu. Additionally, it includes the algorithms utilized by computers in efficiency, hastening the learning curve for AI
hedge funds to execute micro-trades that generate millions machines. AI machines empowered by insights from quan-
of dollars annually [22]. tum computers can learn from experience and self-correct
[38]. The integration of quantum computers will enable the
2.2 Machine learning (ML) expansion of AI into new realms of manufacturing and usher
in a future where technology becomes significantly more
ML, a remarkable aspect of AI, empowers computer systems intuitive [39]. Figure 3 highlights the importance of data.
to enhance task performance by progressively “learning”
from various statistical techniques. In simpler terms, ML 2.3 Applications of AI
involves crafting algorithms that enable increasingly precise
predictions based on the incremental accumulation of data In the present industrial era, decision-makers often find
[23]. The essence of ML lies in granting machines access themselves immersed in a landscape where a profusion of
to data, allowing them to learn from it autonomously. In user, machine, and/or IoT device behavior and activity is
contrast to ML, AI is a dynamic and evolving concept, con- meticulously tracked and shared. This relentless deluge of
tinuously redefined as its related technologies advance [24]. data displays no indication of decelerating. The understand-
AI encompasses an array of technologies such as ML, deep ing that knowledge holds significant influence rings true,
learning (DL), inference algorithms, natural language pro- and a trove of untapped wisdom resides within internal and
cessing (NLP), neural networks (NN), and computer vision. external data sources [41]. To unlock these valuable insights,
AI is pivotal in its evolution, particularly in automation [25]. industries must meticulously deliberate the type of data
ML thrives when confronted with vast datasets, diligently needed, where to procure it, the means of acquisition, and
scrutinizing and comparing the data to uncover common the construction of appropriate data models for comprehen-
patterns and explore subtle nuances [26]. It automates the sive business analysis [42]. The continuous updating of data

Fig. 2  The pillars of AI

5346 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Table 1  The pillars of AI

Pillar Description

Computational power The incredible progress of computing power, following Moore’s law, has exhibited an exponential growth trajectory [29].
Each passing year brings a doubling of computing speed or a halving of prices. In recent times, ML has extensively lev-
eraged graphics processing units (GPUs) [30]. GPUs prove highly adept at executing the vectorized numerical opera-
tions essential for all ML computations. Another remarkable example is Google’s tensor processing unit (TPU), where
specialized (co)processors are explicitly optimized for machine learning tasks [31]. The remarkable advancements in
quantum computing further fuel this trend, likely leading to its continued and accelerated progress. Quantum comput-
ers empower us to tackle complex problems that transcend the capabilities of classical computers, such as encryption,
optimization, and similar tasks [32].
Data availability Data is the fuel of any AI system. The rapid proliferation and accessibility of data are propelled by the widespread adop-
tion of connected devices, such as mobile technology and social media, as well as Internet of Things (IoT) devices [33].
Astonishingly, over the past 5 years alone, internet users have skyrocketed by more than a billion, with mobile phones
accounting for more than half of the world’s web traffic [34]. A substantial volume of data is essential to harness the
power of ML effectively and attain highly accurate predictions [35].
Utilized algorithms Algorithms, the backbone of calculation, data processing, and automated reasoning [36], have seamlessly integrated
into all aspects of manufacturing. Their presence is particularly noticeable in mining data and daily pattern recognition
applications, such as YouTube algorithms [37].

Fig. 3  Value of big data index

[40]

is equally vital, facilitating the retraining and improvement productivity enhancements are, without a doubt, significant and
of algorithms. Algorithms can shape the data we receive, tangible advantages [46]. Still, modern AI technologies present
how we receive it, and even our reactions to it [43]. Figure 4 us with novel capabilities to interact with clients, orchestrate
shows some of the existing ways where AI can be applied. unique experiences, and even carve out unprecedented business
To unravel the current state and future trajectory of AI avenues, given the firm effectively champions the AI-fueled
and automation, a novel scale is presented to assess enter- blueprint for the future. Figure 5 summarizes the essential ele-
prise automation solutions. Table 2 shows the details of such ments of the framework used when deploying AI systems.
a scale. This blueprint could apply to any enterprise contemplat-
ing developing its AI-powered intelligent system to automate
2.4 Deployment of AI systems manufacturing or service processes entirely. It is also helpful
in assessing a holistic approach towards autonomous opera-
In manufacturing and service industries, AI is not merely a tool tions. The critical takeaway underscores the significance of
for refining current operations [44] but also a catalyst for revo- acknowledging the elements within a business context that
lutionary business models and methods [45]. Optimization and are resistant to change, even amidst a rapidly evolving future
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5347

Fig. 4  Applications of AI

[47]. By discerning these elements and directing resources desired output [50]. They infer a function from the
towards them, one ensures an enduring return on their invest- training data and use that function to map new, unseen
ment. This outlook, which emphasizes strategic foresight instances. In simpler terms, you provide these algo-
and sustainable practices, encourages the understanding that rithms with examples of the inputs and outputs you want
efforts made today in these stable areas can continue to pro- the system to produce; then, it learns the relationship
vide valuable returns a decade into the future [48]. between those inputs and outputs. A typical example is
a spam filter, where an SLA is trained on a set of emails
2.5 Breaking down AI subfields for enterprise labeled as “spam” or “not spam” [51]. It learns the char-
operations acteristics of spam emails and uses this knowledge to
classify new emails as “spam” or “not spam.” SLA can
The field of ML and AI encompasses a wide range of studies. be further categorized into regression and classifica-
A few distinct approaches emerge for developing effective tion algorithms. Regression algorithms predict a con-
“learning” algorithms within this expansive domain [49]. tinuous output, like the price of a house, based on vari-
Figure 6 shows different types of these learning algorithms, ous features. Classification algorithms predict discrete
which are discussed in the subsequent parts of this section. results, like identifying whether an email is spam [52].
SLA can also be used in forecasting trends such as the
2.5.1 Supervised learning algorithms (SLA) stock market index. Popular SLAs include linear regres-
sion, logistic regression (LR), support vector machines
SLAs are ML algorithms that learn from labeled train- (SVM), decision trees (DT), random forests (RF), and
ing data, i.e., data that includes both the input and the neural networks, among others.
5348 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Table 2  AI autonomy scale

Scale Nature

Level 0 Restricted AI. Organizations operate with essential tools and systems that offer little to no automation. These tools primarily function
as storage repositories for organizational data and primary reporting purposes. While they provide some insight through reporting
systems, the automation capabilities are minimal. As a result, organizations rely heavily on manual processes to manage tasks and
operations.
Level 1 Perspective AI. Organizations utilize advanced systems that can adhere to predefined business rules set by decision-makers. These
systems generate actionable recommendations aimed at optimizing organizational outcomes. For instance, dashboards with recom-
mendation systems enable marketers to adjust marketing spending across different channels based on data-driven insights. However,
the final step of implementing the recommended adjustments still requires human intervention. These recommendation systems are
powerful tools, empowering organizations to make informed decisions and drive improved performance.
Level 2 Dynamic AI. Expanding upon the foundation of Level 1’s business rules, Level 2 dynamic automation takes a stride forward by
automatically adjusting marketing campaigns, typically through dedicated applications, without user intervention or approval. While
Level 1 requires users to define the rules, Level 2 automation operates autonomously based on these pre-established guidelines.
However, it is important to note that dynamic market conditions are subject to frequent fluctuations, occurring daily, hourly, or even
minute-by-minute. Although rules-based systems provide a degree of automation, their rigidity limits their adaptability to respond to
evolving market changes swiftly.
Level 3 Perceptive AI. Leveraging the power of ML, systems operate with computational autonomy, observing, learning, and enhancing out-
comes through statistical analysis. This advanced level of automation requires minimal user intervention, with the primary task being
establishing goals or high-level parameters. Once these parameters are set, the system autonomously adapts and optimizes strategies,
continuously improving results without constant user involvement. This level of perceptive autonomy empowers organizations to
achieve optimal outcomes while minimizing manual efforts.
Level 4 Insightful AI. Operating with a profound understanding of systems interactions, products and services, performance data, etc., systems
can exhibit applicable autonomy. By leveraging this contextual understanding, organizations can deliver tailored experiences that
resonate with specific products or services. The goal is to optimize performance for operators by dynamically adapting strategies to
meet unique needs and preferences.
Level 5 Unrestricted AI. Level 5 systems go further by independently generating unsupervised tasks, creative variations, targeting parameters,
and more. Remarkably, these systems operate with minimal to no ongoing intervention from the management team. With this level
of autonomy, the system autonomously explores and innovates, continuously generating new and diverse approaches to optimize
operational performance. By freeing the management team from routine tasks and decision-making, organizations can fully leverage
the system’s independent capabilities to drive exceptional outcomes, unleashing the true potential of the autonomous enterprise.

Fig. 5  The intelligent machine framework

2.5.2 Unsupervised learning algorithms (ULA) tasks, ULA could segment a company’s customer base
into different groups based on purchasing behaviors. The
ULA is another type of ML algorithm that operates on algorithm identifies patterns and similarities in the data
unlabeled data. That means the algorithm is not provided that might not be immediately apparent to human ana-
with the correct results during training and must find struc- lysts. Still, no labels are provided to guide this process—
ture in the input data on its own [53]. These algorithms the algorithm determines the grouping based on the data
are often used for clustering, dimensionality reduction, itself [54]. Common ULAs include K-means clustering,
and anomaly detection tasks. For example, in clustering hierarchical clustering, Density-Based Spatial Clustering
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5349

Fig. 6  General taxonomy for AI techniques

of Applications with Noise (DBSCAN), principal com- achieve more accurate results with less manually labeled
ponent analysis (PCA), and autoencoders, among others data. Examples of SSLA include certain DT, Bayesian net-
[55]. It is worth noting that ULA can be more challenging works, and some neural network models [61].
than SLA, as it can be harder to verify the accuracy of
the results, and the model may require more data to find 2.5.4 Reinforcement learning (RL)
patterns accurately. However, ULA can be very useful for
exploratory data analysis and dealing with data where we RL is a type of ML where an agent learns to make decisions
do not have labels [56]. by taking actions in an environment to maximize cumulative
reward. It differs from SLA in that correct input/output pairs
2.5.3 Semi‑supervised learning algorithms (SSLA) are never presented, and from ULA, there is no need to find
hidden structures in the data [62]. In RL, an agent interacts with
SSLA falls somewhere in between SLA and ULA. These its environment by producing actions and discovering errors or
algorithms use a combination of a small amount of labeled rewards. Through trial and error, it learns which efforts yield the
data and a large amount of unlabeled data for training [57]. most reward by exploring and exploiting its environment [63].
The idea behind SSLA is that the learning process can be The agent’s knowledge is updated as it learns from both its suc-
improved by using unlabeled data when labeled data is cesses (reinforced through rewards) and its failures (penalized
scarce or expensive to obtain. SSLA attempts to use unla- in some way) [64]. A classic example of RL is a chess game.
beled data along with the labeled data to improve learning The agent decides on a series of moves based on the state of the
accuracy [58]. This is done through methods like pseudo- board (the environment), intending to win the game (maximiz-
labeling (where the model makes an educated guess about ing the reward). The agent will adjust its decisions based on the
the labels of the unlabeled data and then uses those guesses game’s outcome, learning from winning and losing games [65].
to learn more) or through self-training (where the model is RL has been used in various fields, from gaming—where it has
initially trained on the labeled data, then used to classify the been used to train computers to play and win games at a high
unlabeled data, and then retrained with the data it organ- level, to robotics, where RL can help autonomous vehicles or
ized itself) [59]. This approach is instrumental in scenarios drones navigate an environment efficiently [66]. It is also used
where labeling data is time-consuming or expensive, such as in recommendation systems, NLP, and many other areas [67].
computer vision or NLP tasks [60]. SSLA provides a practi- Popular RL algorithms include Q-Learning, Deep Q-Network
cal middle-ground, allowing models to generalize better and (DQN), and Policy Gradients (PG).
5350 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

2.5.5 Deep learning (DL) model. This might include a combination of supervised and
unsupervised techniques, different types of neural networks,
DL is a subset of ML, a neural network with three or more or a variety of statistical and machine learning methods [75].
layers. These neural networks attempt to simulate the behav- Figure 7 shows the relationship between AI, ML, DL, DHL,
ior of the human brain—albeit far from matching its abil- RL, SLA, ULA, and SSLA.
ity—to “learn” from large amounts of data [68]. While a
neural network with a single layer can still make approxi- 2.5.7 Computer vision
mate predictions, additional hidden layers can help optimize
the accuracy [69]. DL drives many AI applications and ser- Computer vision is a field of AI that trains computers to
vices that improve automation, performing analytical and interpret and make decisions based on visual data, such
physical tasks without human intervention [70]. This tech- as images or videos. It aims to replicate the complexity
nology powers car navigation systems, voice-enabled TV of the human vision system and enable machines to pro-
remotes, credit card fraud detection, and many more. DL can cess and understand visual information from the world
learn directly from data and compute independently [71]. similarly to how humans do. This technology is used in
For instance, you do not have to program a DL network to various applications like facial recognition, autonomous
recognize a specific fish species; you can feed the network vehicles, and medical image analysis [76]. The relation-
images of it, and the DL network can learn on its own [72]. ship between computer vision and ML is intrinsic, as most
There are various DL models, including CNNs [73] that are computer vision tasks require some form of ML model
commonly used for image-related tasks, recurrent neural net- to operate. DL algorithms, particularly deep neural net-
works (RNNs) used for sequential data such as time series or works, have become the de facto standard in solving com-
sentences, and transformers that are particularly effective in plex computer vision problems. SLA is often used when
natural language processing tasks. Because of its profound labeled data is available, while ULA may be applied for
capability to learn from large volumes of data, DL has been tasks like clustering or anomaly detection where labeled
instrumental in advancing many areas of AI, including com- data is not necessary. DHL combines various architectures
puter vision, NLP, and speech recognition [74]. for better performance, while SSLA uses both labeled and
unlabeled data for training. RL can also be integrated for
2.5.6 Deep hybrid learning (DHL) tasks requiring sequential decision-making, such as robot
navigation. All these paradigms fall under the broader
The term combines DL techniques with other ML umbrella of AI, with computer vision as a critical compo-
approaches within a single model. It refers to using DL net- nent. Computer vision can automate anomaly and defect
works with multiple layers between the input and output detection with unparalleled accuracy to minimize delays
layers to model and understand complex patterns in datasets by catching issues as they arise. In addition, it can moni-
along with different types of ML algorithms within a single tor hazardous areas, track personal protective equipment

Fig. 7  Relationship between the

subfields of AI
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5351

Fig. 8  Examples of computer vision applications

compliance, and better manage machine usage in real-time Recently, transformer-based models have revolutionized
[77]. Figure 8 shows some applications of computer vision the field of NLP, offering state-of-the-art performance in
in manufacturing. many tasks. SSLA and RL approaches are also explored
for problems like dialog systems or learning from limited
2.5.8 Natural language processing (NLP) labeled data. Figure 9 shows the different methods in data
mining.
NLP is a subfield of AI that focuses on the interaction
between computers and human language. It aims to enable 2.6 Breaking down data mining techniques
computers to understand, interpret, and respond to human
language in a meaningful and valuable way [78]. NLP cov- Analytics can be broadly categorized into different branches.
ers a range of tasks, including but not limited to text analy- Data mining (DM) is a process used to extract usable data
sis, sentiment analysis, machine translation, and chatbot from a more extensive raw data set. It involves analyzing data
functionality [79]. Similar to computer vision, NLP exten- patterns in large batches of data using one or more software.
sively uses ML and DL techniques to perform its tasks. It consists of discovering patterns, correlations, trends, and
SLAs are often employed for tasks where labeled data is anomalies (or outliers) in large data sets involving methods
available, such as classification problems. ULA is used at the intersection of ML and statistics. Different branches
for functions such as topic modeling and word clustering. of analytics, such as descriptive, diagnostic, predictive, and

Fig. 9  Data mining techniques

5352 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

prescriptive analytics, are techniques that can be applied within to other data points in the same group than those in other
the broader process of DM. However, data mining also encom- groups. Simply put, the aim is to segregate groups with simi-
passes several other designs and methodologies beyond those lar traits and assign them into clusters. Examples include
four, such as association rule learning, cluster analysis, clas- segmenting customers into different groups for marketing
sification, regression analysis, and anomaly detection [80]. campaigns or grouping documents that cover similar topics
[86].
2.6.1 Descriptive analytics
2.6.7 Classification
This is the most basic form of analytics. It involves process-
ing raw data to make it interpretable. The focus is on sum- This technique is used to classify data into defined cate-
marizing and understanding the past. Techniques include gories. It involves training a model on a dataset where the
basic statistics (means, medians, and standard deviations) classes are pre-labeled for it to learn the boundaries that
and data visualization (charts, graphs, and tables) [81]. divide the classes. The model can then classify new inputs
into these categories. For instance, an email can be classified
2.6.2 Diagnostic analytics as “spam” or “not spam” [87].

This takes descriptive analytics a step further. It involves 2.6.8 Regression analysis

more detailed data exploration to understand the root cause
of a particular outcome. For example, suppose descriptive Regression analysis is a set of statistical processes for esti-
analytics tells you that your enterprise has seen a sudden mating the relationships among variables. It includes many
drop in sales. In that case, diagnostic analytics might help techniques for modeling and analyzing several variables
you figure out why by looking at changes in sources of cus- when focusing on the relationship between a dependent vari-
tomer flow, marketing campaigns, or functionality [82]. able and one or more independent variables. It can be used
for forecasting, time series modeling, and finding the causal
2.6.3 Predictive analytics effect relationship between the variables [88].

This type of analytics uses statistical modeling and machine 2.6.9 Text analytics
learning techniques to understand potential future outcomes. For
example, it might involve using regression analysis to predict Text analytics converts unstructured text data into meaning-
sales based on past trends and current market conditions [83]. ful insights or information using algorithms and statistical
methods [89]. It encompasses a wide range of applications
2.6.4 Prescriptive analytics and techniques, including but not limited to sentiment analy-
sis, topic detection, named entity recognition, and keyword
This is the most advanced form of analytics, and it involves extraction. Text analytics aims to allow businesses, research-
recommending actions to optimize outcomes. It uses sophisti- ers, or individuals to derive valuable conclusions and make
cated techniques, including optimization algorithms and simu- informed decisions based on textual data [90].
lations. For example, a prescriptive analytics model might
recommend how to allocate a marketing budget across differ-
ent channels to maximize customer engagement or sales [84]. 3 The integration of lean practices and AI

2.6.5 Association rule learning Lean Manufacturing’s core principle seeks to augment the
quality of goods or services by curtailing wastage, con-
This is a method for discovering interesting relationships sequently increasing value. This approach has become
or associations among large data items. A famous example integral in meeting customer demands and maintaining
is the “market basket analysis.” In retail, this technique can relevance amidst the steep competition prevalent in global
be used to analyze what products are often bought together, markets. Japan’s Toyota introduced this innovative concept
which could influence product placements, sales, and recom- known as the Toyota Production System (TPS), which now
mendations [85]. thrives in numerous industries. At its heart, Lean Manu-
facturing is all about offering premium products while
2.6.6 Cluster analysis minimizing manufacturing costs [91]. Lean foremost
objective is to cut down on waste in a company, viewed
Clustering is dividing the data points into several groups as a superfluous resource draining that adds no real value.
such that data points in the same groups are more similar The tools of Lean Manufacturing cleverly slice operational
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5353

costs by slashing waste, elevating product quality, and Lean Manufacturing, effectively assisting in waste reduc-
boosting efficiency. The application of the Lean Manufac- tion [104].
turing methodology has found widespread acceptance in
industries as a tried and tested means to maximize produc- 3.1 Lean Manufacturing core principles
tivity, enhance quality, and pare down organizational costs
[92]. Studies have corroborated that the Lean approach Lean Manufacturing originates in cost-saving manufacturing
positively impacts productivity, leads to reduced manu- methodologies with a primary waste reduction objective.
facturing costs, curbs unproductive downtime, optimizes While companies across various industries have embraced
resource utilization, and escalates profitability. Lean Man- lean principles in their processes, their most significant
ufacturing propels an organization’s competitive standing applicability lies in physical manufacturing, engineering,
in the market. Firms are increasingly keen to embrace this and related sectors. Irrespective of your organization’s size
approach to solidify their economic footing and viability. or industry, integrating Lean principles can prove highly
A noticeable trend of steady organizational improvement beneficial in optimizing day-to-day operations, streamlining
has been observed in the wake of Lean Manufacturing’s workflows, and maximizing efficiency. These five principles
implementation [93]. The induction of Lean Manufactur- form the basis of Lean Manufacturing and provide a frame-
ing instruments like Total Productive Management (TPM), work for organizations to minimize waste, improve effi-
Overall Equipment Effectiveness (OEE), and Jidoka had ciency, and create more value for customers. Table 3 shows
a profound effect on a company’s landscape and its social how AI algorithms can help in carrying out these principles.
and economic sustainability [94]. In terms of operational By applying AI in these ways, manufacturers can imple-
benefits, Lean Manufacturing stands out in aspects such ment Lean Manufacturing principles more effectively,
as quality, delivery, cost, and volume flexibility [95]. Evi- resulting in more efficient operations, less waste, and higher
dence has indicated that amalgamating digital technol- value for customers. However, the choice of a specific algo-
ogy with Lean Manufacturing can trim production costs, rithm would depend on the business problem at hand, the
irrespective of varying environmental conditions. This nature and quality of the data available, the computational
combination also enhances the operational performance resources, and the team’s technical expertise. Figure 10
and efficiency of production management. Implementing shows an illustrative summary of the relationship between
Lean practices has consistently improved an enterprise’s AI and the five core principles of Lean.
financial health by fostering a safe approach to shop floor
management [96]. For instance, railway transport enter- 3.2 Waste in Lean Manufacturing
prises have profited from the introduction of Lean Manu-
facturing by optimizing their labor and material resources The core idea behind Lean Manufacturing is to eliminate
[97]. Implementing Lean tools has expedited operations, non-value-added activities (waste) to improve customer
improved quality, and broadened the customer base [98]. value [105]. Manufacturing encompasses a diverse array of
Similarly, Lean Manufacturing’s integration into enter- waste forms. The 3M concept is derived from Japan to iden-
prise management has led to greater efficiency and qual- tify and eliminate waste in processes and services. It refers
ity in production [99]. It has effectively curtailed waste to Muda, Mura, and Muri. The 3Ms are vital to the TPS and
at workstations, thereby cutting down on operational other Lean methodologies and tools. All three of these are
expenses [100]. Lean Manufacturing has also proven suc- considered types of waste and are targets for elimination
cessful in avoiding order returns, reducing waste genera- in Lean Manufacturing. The goal is to create a workplace
tion, reducing production costs, and enhancing operational that is efficient, safe, and capable of delivering high-quality
performance [95]. A noteworthy observation is that com- products or services to meet customer demand [106].
panies that had embraced Lean Manufacturing before the
COVID-19 pandemic were better equipped to weather the 3.2.1 Muda
crisis [101]. Despite the proven advantages of Lean Manu-
facturing, certain obstacles hinder its widespread adop- Originating from Japan, the term Muda encapsulates the
tion, such as accountability issues, knowledge and training multi-faceted nature of waste produced in manufacturing,
gaps, budgetary constraints, and unfavorable psychology also known by the acronym DOWNTIME, spanning defects,
[102]. However, on a more positive note, research sug- overproduction, waiting, non-utilized talent, transportation,
gests that there is still immense potential to reap benefits inventory, motion, and excess processing. Statistical data
by applying Lean tools across various enterprises [103]. conclusively indicate defects as the most frequently encoun-
Embracing sustainability in manufacturing necessitates tered issue in manufacturing [107]. Moreover, disordered
the evolution of novel and innovative technologies. One workspaces and non-standardized production procedures
potential path to this might involve the introduction of can exacerbate waste production [108, 109]. Table 4 shows
5354 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

how AI might contribute to the reduction process of each

DT or RF, can also be used to model and predict the effectiveness of different improve-
Sentiment analysis and NLP algorithms can help analyze and understand customer feed-

Anomaly detection algorithms such as Isolation Forest or Autoencoder can detect devia-
tions from standard performance, thus indicating areas for improvement. SLA, such as
back and preferences. They can help interpret unstructured data like customer reviews

PdM can be facilitated by regression or time series algorithms such as Auto-Regressive

map the value stream. They can categorize processes based on similarities and differ-
type of Muda.

For demand forecasting, algorithms such as ARIMA or neural network-based methods

Value stream AI can help identify and map the value stream through advanced data analysis. AI can ULA, such as clustering (e.g., K-means, hierarchical clustering), can help identify and

such as Q-Learning, can be employed to optimize scheduling for the smooth flow of

historical demand data to predict future needs, helping align production with actual
Integrated Moving Average (ARIMA) to predict machine failures. RL algorithms,

such as long short-term memory (LSTM) can be used. These algorithms analyze
3.2.2 Mura

Mura is another Japanese word used in the Lean Manufac-

turing context that indicates inconsistency, unevenness, or
imbalance [110]. This could be variability in the production
or feedback forms to understand what customers truly value.

schedule, uneven customer demand, inconsistent work pace,

ences, helping to identify areas of waste and inefficiency.

or any other irregularity that leads to inefficiencies [111].

Lean aims to identify and eliminate Mura in the processes,
resulting in a smoother, more predictable, and efficient pro-
cess flow [112].
AI can help smooth out Mura in the production process.
This could involve anything from improving demand fore-
casting to optimizing production schedules. AI can analyze
patterns in demand, supply, and production data to identify
and address sources of inconsistency. Time series analy-
sis algorithms such as ARIMA or LSTM can be used for
improving demand forecasting. At the same time, RL or
ment strategies.

genetic algorithms (GA) can be used for optimizing pro-

AI algorithms

duction schedules to smooth out production flow.

processes.

demand.

3.2.3 Muri

Muri is a Japanese term used in Lean Manufacturing that

that disrupt flow. AI can also optimize scheduling, ensuring tasks are completed most
nance (PdM) via AI can forecast machine failure, preventing unexpected breakdowns
identify patterns and relationships in data that may not be readily apparent, pinpoint-

success of improvement initiatives over time, providing data-driven feedback on the

AI can be used to analyze customer behavior, preferences, and feedback from various

stands for any activity exceeding the capacity of 4M (man,

AI can refine demand forecasting, making it more accurate and reliable. This makes
AI can optimize the production process to ensure a smooth flow. Predictive mainte-

AI can aid in the continuous improvement process. AI can identify deviations from
standard performance, highlighting areas for improvement. AI can also track the
production more closely aligned with actual demand, reducing overproduction.

machine, method, and material) against the actual require-

sources, enabling a deeper understanding of what the customer truly values.

ment [113]. This strain can come from unreasonable work

expectations, over-utilization of resources, lack of mainte-
nance, or even an unstable working environment [114]. Muri
can lead to several problems, such as employee burnout and
low morale, increased error rates due to stress and fatigue,
breakdown of machines due to overuse without proper main-
tenance, reduced life expectancy of equipment, and high
employee turnover rates [115].
AI can be used to optimize workloads and resources to
Table 3  How AI can aid in executing the five Lean principles

prevent Muri. For instance, ML models can analyze his-

torical production data to predict periods of high demand,
effectiveness of different strategies.
ing areas of waste and inefficiency.

allowing for proactive workload balancing. Regression and

LSTM models can be used for predicting periods of high
demand and RL algorithms for optimizing resource alloca-
tion. Also, AI can help automate repetitive or physically
demanding tasks, reducing human operator overburden.
Robotic process automation (RPA) can be utilized to auto-
Potential impact

mate repetitive tasks. Figure 11 visually illustrates the rela-

efficiently.

tionship between AI and the 3 M.

AI algorithms can automate and optimize various aspects
of the manufacturing process, reducing waste and enhancing
productivity. However, the exact choice of algorithms would
depend on the manufacturing setup’s specific circumstances
Perfection
Principle

and requirements, the data’s availability and quality, and the

Value

Flow

Pull

computational resources available.

The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5355

Fig. 10  Executing the five Lean core principles via AI

3.3 The 6S method and simplification. By separating unnecessary items and

waste from the manufacturing process [122], Sort elimi-
Takashi Osada introduced the 5S method in 1970 [116] nates non-value-added steps [123]. During this process,
to support the implementation of Lean principles. Com- items deemed unnecessary or misplaced receive a red tag
prising Sort, Set, Shine, Standardize, and Sustain, 5S is and are subsequently relocated, recycled, disposed of, or
a Lean tool that aims to eliminate unnecessary items that reassigned. This sorting activity optimizes floor space
do not contribute to production, rectifying an untidy and and eliminates broken, scrap, or excess raw materials
unhealthy work environment. A study across ten manu- [124]. Notably, research on automotive parts manufac-
facturing units revealed the positive impact of implement- turing companies demonstrated that the deployment of
ing 5S, showcasing continuous workplace improvement 6S positively influenced machine downtime and improved
and notable enhancements in employee relations and production efficiency, emphasizing its critical role in
motivation [117]. The positive impact of 5S extends to enhancing efficiency within this industry [125]. Addi-
enhancing manufacturing production quality [118] and tionally, Sort plays a crucial role in material management,
effectiveness [119]. Notably, the use of 5S + 1 (safety is eliminating waste in the form of excess or junk material
the sixth S) in organizing the workplace of a scientific [125]. By implementing 6S, an engineering firm could
instruments manufacturing company has yielded signifi- identify and remove unwanted equipment and materials
cant results in process improvement, continuous enhance- from the workplace, improving workflow efficiency and
ment, and waste reduction [120]. Through its systematic eliminating unnecessary movements [126]. Applying the
implementation, 6S is pivotal in promoting Lean practices 6S tool in health services reduced processing time, wait-
and driving positive outcomes across diverse industries ing hours, production cycles, and costs while enhanc-
and organizational contexts [121]. ing the quality and safety of medical services [127].
The Sort element of 6S enables enterprises to shorten
3.3.1 Sort in the 6S searching time, maintain a Lean and safe workspace, and
promptly detect equipment faults [128]. As a fundamental
The initial phase of implementing the 5S + 1 strategy approach, the 6S method enhances the production process
involves Sort, the foundation for improving workspace across diverse industries. Adequate implementation of 6S
management and enhancing work quality. Sort, derived is essential to increase process effectiveness by eliminat-
from Seiri, emphasizes the importance of order, selection, ing waste [129, 130].
 5356

Table 4  AI potential impact on the Muda reduction process in Lean Manufacturing
Muda Potential impact AI algorithms

Defects: errors that lead to poor-quality products or the need AI-powered quality control systems can identify and correct CNN can be used for image recognition in quality control
for rework errors in real-time, reducing the need for rework and elimi- systems. They can identify product defects faster and more
nating poor-quality products. accurately than human inspection, reducing error rates and
improving overall quality.
Overproduction: making more of a product than is required AI can analyze demand patterns to more accurately forecast Time series forecasting algorithms, like ARIMA or LSTM, can
can lead to unnecessary storage and maintenance costs needs, thus reducing overproduction. analyze historical demand data to predict future needs, pre-
venting overproduction by aligning production with demand.
Waiting: downtime between processes or waiting for previous Using AI to streamline scheduling, workflow optimization, RL algorithms can be used to optimize scheduling and resource
steps to be completed. and process automation can reduce downtime and wait allocation, minimizing waiting times between different stages
times. of the manufacturing process.
Non-utilized talent: not using employees’ skills, experience, or AI systems can be used for skills mapping and identifying NLP can be applied to analyzing employee feedback, job
knowledge to their full potential areas where employees’ skills can be best utilized, ensuring descriptions, and performance reviews to understand
the right people are doing the right jobs. employee skills and competencies better, ensuring they are
well-utilized and reducing skill waste.
Transportation: unnecessary movements of products or mate- AI-powered logistics tools can optimize routes and load capac- Route optimization algorithms (like Dijkstra’s or A*) can help
rials within the production process ities, reducing unnecessary transportation and motion. in logistics planning, reducing unnecessary transportation by
finding the most efficient routes.
Inventory: excess raw materials, work-in-progress (WIP), or AI can predict demand, allowing more efficient inventory Multi-layer perceptron (MLP), a type of neural network, can
finished goods that do not add value management reducing storage costs and potential waste from predict demand and manage inventory levels effectively,
expired or outdated stock. reducing the waste associated with excessive inventory.
Motion: unnecessary movement of people or equipment AI can help design more efficient workflows, reducing unnec- Computer vision (via CNN for instance) can analyze workflows
within the process essary movement within processes. and worker motions to identify unnecessary movements and
suggest more efficient workflows.
Excess processing: extra steps taken in production that do not AI can analyze processes to identify unnecessary steps, help- Unsupervised learning algorithms like clustering (K-means,
add value to the customer ing to streamline operations and reduce excess processing. hierarchical clustering) can help analyze process steps and
group similar activities, helping to identify and eliminate
redundant or unnecessary steps in the manufacturing process.
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5357

Fig. 11  Waste reduction via AI

3.3.2 Shine in the 6S systematic cleaning to sustain a healthy work environment,

thereby supporting the production of high-quality work
After completing the sorting process, maintaining order [133]. Clean equipment operates more efficiently, elimi-
within the workspace becomes crucial. Regular cleanup is a nating unnecessary costs arising from unnoticed problems.
fundamental aspect of a well-functioning production system. Manufacturers have recognized that maintaining cleanliness
In the context of 6S, the distribution of work is based on during daily maintenance reduces the likelihood of down-
Lean time and designated Lean areas [131]. Furthermore, time, ultimately leading to increased profit margins [134].
a Lean working environment facilitates the prompt identifi- Moreover, documented evidence indicates a reduction in
cation of faulty equipment caused by excessive vibrations, common accidents and risks within a Lean factory, attest-
leakages, misalignments, and other factors. Addressing these ing to the positive impact of a Lean workspace or shop floor
issues immediately is essential to prevent production losses [135].
or equipment failures. Thus, a Lean and organized work-
space plays a vital role in detecting and mitigating poten- 3.3.3 Set in the 6S
tial sources of failure or downtime. The concept of Shine,
derived from Seiso, emphasizes the importance of sweep- Set or Seiton in the 6S methodology emphasizes the impor-
ing, scrubbing, and ensuring cleanliness in the workspace. It tance of systemization, set in order, and organization. It plays
focuses on guiding operators or workers on the shop floor to a significant role in facility layout planning by determining
maintain a state of Lean for machines and the overall envi- the appropriate placement of various resources, including
ronment. Employees receive training on the adverse effects raw materials, machine tools, and semi-finished products.
of contamination and how to avoid them, fostering a risk-free The objective is to ensure that locating a necessary object
working environment [132]. The Shine principle advocates takes no more than 30 s [136]. Once the sorting process has
5358 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

eliminated unnecessary items, the implementation of set in organizational adaptation [146]. Efforts were made to inte-
order can commence. Set in order involves arranging objects grate the 6S concept and tools into selected standard texts
based on size and frequency of use. Additionally, labeling like ISO 9001, providing a valuable guideline [147].
is employed for ease of identification, and items are stored
or placed in locations according to their designated order. 3.3.5 Sustain in the 6S
Organizing tools based on usage frequency, set in order, min-
imizes movements and reduces search time. Labels, tapes, The Shitsuke element of 6S, also known as sustain, aims to
signs, and floor markings are utilized to facilitate the set ensure the ongoing maintenance and adherence to the new
in order process. This approach eliminates various wastes, and effective processes established as standards within the
including search, waste in motion, and excessive inventory organization. It emphasizes the importance of conducting
[137]. The adoption of 6S has resulted in numerous positive planned routine workplace inspections at specific intervals
changes. For instance, studies have shown a 60% reduction [148]. Implementing 6S helps eliminate obstacles that hinder
in employee movement and a 20% reduction in operational an efficient production process [148]. Sustain is seen as a
costs [138]. Furthermore, it has enhanced productivity, work critical step in maintaining the first four S’s and facilitating
and product quality, safety, health, and efficiency [139]. The the successful implementation of the sixth S, which pertains
implementation of set in order has also reduced waste in to safety [149]. The primary objective of sustain is to cul-
motion, such as excessive movements between production tivate a habit among employees to follow the best practices
halls and workstations [140]. consistently throughout the organization. Moreover, sustain
fosters improved human relationships within the company
3.3.4 Standardize in the 6S by promoting discipline and ensuring the continuous imple-
mentation of the 6S process [149]. The adoption of 6S has
Seiketsu, another element of 6S, encompasses standardi- motivated employees to create better working conditions
zation, tidiness, and sanitization. Its purpose is to enable by equipping them with tools such as value stream map-
seamless execution of repetitive tasks without hindrance or ping (VSM) and continuous improvement methodologies to
unnecessary time consumption [141]. The primary goal of reduce waste. When applied, the 6S method influences indi-
standardization is to establish the first three S’s—sort, set in vidual behaviors, establishing simple, precise, and efficient
order, and shine—as habitual practices, ensuring that work- rules that contribute to a disciplined and functional work
ers do not revert to previous disorganized methods [142]. environment [150]. However, implementing sustain can be
Seiketsu involves creating and monitoring standards for challenging as it involves changing traditional practices and
various events, such as inspections and safety plans. A safe maintaining new procedures that disrupt established routines
working environment is maintained by promoting standardi- [151]. Without the sustained application of 6S, the benefits
zation, and the potential disruptions caused by changes in and effectiveness of the methodology cannot be maintained
workers’ duties are minimized. This allows for shorter train- [152]. By upholding the 6S strategies in the workplace, the
ing times and smooth adaptation to new working environ- work environment is simplified, making it easier for every-
ment dynamics [143]. Moreover, standardization prevents one to adhere to the established rules and guidelines [153,
waste accumulation and fosters consistent procedures, con- 154].
tributing to a Lean working environment. Without standards,
achieving product quality improvement becomes impossible 3.3.6 Safety in the 6S
within an organization [143]. In the current landscape of
the overlapping Lean and Industry 4.0 (I4.0) environments, Safety is a paramount concern in the workplace [155],
standardization emerges as a critical parameter for ensuring focusing on preventive measures to safeguard workers from
workspace safety, reducing disruptions, and enhancing pro- hazardous conditions and provide them with a secure and
duction efficiency. 6S, as a business or industrial approach, stress-free environment. Studies emphasize safety’s signifi-
primarily focuses on improving the production process cant role in enhancing the overall work environment [156].
by elevating workplace standards [144]. Standardization Certain operations, such as welding, gas cutting, and casting
extends beyond maintenance procedures, encompassing processes, require additional preventive measures to mini-
aspects such as production planning, specifications, trans- mize the occurrence of incidents [157]. Implementing the
portation of goods, human resource management, customer 6S method is closely intertwined with various industrial,
service, and bookkeeping [144]. The successful implemen- manufacturing, and construction operations, contributing to
tation of 6S relies on employee commitment, sustainability safer work conditions [158, 159]. In the healthcare industry,
initiatives, and comprehensive training [145]. As compa- adopting 6S has led to waste reduction, decreased errors, and
nies adopt 6S practices, new production plant rules and increased productivity [160]. Moreover, applying 6S ensures
streamlined processes are established, enabling effective that operators use personal protective equipment (PPE) and
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5359

appropriate gear to prevent accidents [161]. The 5S + 1 In Lean Manufacturing, it is recognized that the people
method is a powerful catalyst for improving the quality of doing the work often understand it best. Therefore, they can
the work environment, with safety being a key aspect [162]. provide valuable insights into improving the work process
The Lean 6S methodology provides a systematic way [169]. By empowering all employees to suggest and make
to organize the workplace for efficiency and effectiveness. changes, organizations can tap into a vast pool of poten-
Table 5 shows ways AI could help accelerate the implemen- tial improvements and build a culture that embraces change
tation of 6S in the workplace. Implementing AI can sig- and strives for excellence [170]. Continuous improvement
nificantly assist in managing, optimizing, and sustaining an in Lean Manufacturing can lead to higher quality products,
organized and efficient workspace according to the Lean increased efficiency, lower costs, improved customer satis-
6S principles. Figure 12 shows a graphical summary of AI faction, a safer work environment, and a more prosperous
potential in this area. and sustainable business. AI can facilitate the implemen-
tation of Kaizen by providing a framework for systematic
3.4 Kaizen data collection and analysis, supporting decision-making,
and allowing for the visualization and communication of
Kaizen, also known as continuous improvement in Lean improvement efforts [171]. Table 6 shows how Kaizen meth-
Manufacturing, is a fundamental principle that aims to ods can be carried out via AI.
improve all functions continually and involve all employ- Deming’s Plan-Do-Check-Act (PDCA) approach is based
ees, from the CEO to the assembly line workers. It can on Kaizen. You can generate desired outcomes through sys-
include external customers and all sorts of end-users [163]. tematic and effective planning, implementation, evaluation,
It extends to procurement, shipping, management strate- and review processes. Figure 13 shows the relationship
gies, and others in addition to manufacturing [164]. In Lean between Kaizen implementation and AI.
Manufacturing, continuous improvement seeks to improve
every process in a company by focusing on enhancing the 3.5 Value stream mapping (VSM)
activities that generate the most value for the customer while
removing as many waste activities as possible [165]. This VSM is a Lean Manufacturing technique used to analyze,
process involves incremental improvements over time rather design, and manage the materials and information required
than large-scale changes [166]. By focusing on the more to bring a product to a customer or end-user [172]. It is
minor, manageable improvements, an organization can enact a visual tool that helps you see and understand the flow
substantial positive change over the long term [167]. It is a of material and information as a product goes through
cycle that involves identifying areas of potential improve- the value stream [173]. The primary goal of VSM is to
ment, implementing small changes, measuring the effects, identify every step of a process—both value-adding and
and then starting the cycle again [168]. Key to this process non-value-adding—from start to finish and eliminate
is the active participation of all members of an organization. waste [174]. Doing so, VSM allows businesses to reduce

Table 5  AI potential impact on 6S implementation

The S AI impact Algorithms

Sort AI can help automate the sorting process by identifying neces- Computer vision algorithms such as CNN can be used to
sary or unnecessary items for the workspace. recognize different tools or materials, categorize them, and
determine which are essential for work.
Set in order AI can assist in optimizing the arrangement of items for the Using RL algorithms, AI can learn through trial and error to find
most efficient workflow. the best arrangements that minimize movement and time.
Shine AI can help automate the shine process by identifying what Computer vision algorithms such as CNN can again be applied
areas require attention. to identify unclean or suboptimal conditions in the workspace,
flagging areas that need cleaning or maintenance.
Standardize AI can analyze data from the production process to develop Regression analysis or DT algorithms can be used to understand
standardized best practices. the relationship between different factors in the production
process and their impact on the final product quality.
Sustain AI can help monitor adherence to the established standards over Anomaly detection algorithms like Isolation Forest (IF) can be
time, identify deviations, and provide alerts when corrective used to monitor for abnormal behaviors that deviate from the
action is needed. standard.
Safety AI can identify potential safety risks in the workspace, such as Computer vision algorithms such as CNN can again be applied to
hazards like spills or cluttered pathways, or ensure workers recognize these safety risks and trigger alerts.
wear appropriate safety equipment.
5360 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 12  6S implementation via AI

Table 6  Kaizen methods via AI algorithms

Kaizen method AI potential and algorithm

Kaizen by reduction ML algorithms such as DT or regression models can identify factors contributing to waste or inefficiency and provide
insights on reducing them. For example, time series analysis can be used to predict machinery breakdowns to mini-
mize downtime and enhance maintenance schedules.
Kaizen by elimination ULA, such as clustering (e.g., K-means or hierarchical clustering) can help analyze process steps and group similar
activities, aiding in identifying and eliminating redundant or unnecessary steps.
Kaizen by improvement SLA, such as SVM or neural networks, can model complex relationships in the data to predict outcomes under dif-
ferent conditions, guiding improvement efforts. RL algorithms can also be used to iteratively improve processes by
learning from past performance.

lead time, eliminate non-essential activities, and improve elimination tools can be applied to alleviate it. Table 7
overall efficiency [175]. A typical VSM includes a visual shows how AI algorithms can help in executing the com-
representation of every step in a process, along with a mon elements of VSM.
variety of specific data related to each step, such as time These AI algorithms can identify inefficiencies and waste
taken, resources used, or the volume of products flowing and understand relationships between different variables and
through [176]. VSM is an essential tool in identifying bot- their impact on the overall process. In addition, AI can simu-
tlenecks within a production process. A bottleneck refers late different scenarios for the future state of the value stream
to any point in the process where the capacity is less than based on the data and the analysis results. RL algorithms
the demand placed on it, leading to inefficiencies, delays, can be used to iteratively learn from each simulation and
and waste [177]. Each process step is evaluated in terms of optimize the value stream layout for maximum efficiency
its capacity, demand, and throughput time. By comparing (prescriptive analysis). By harnessing AI in these ways, it
these parameters across all phases, VSM can identify the is possible to gain deeper insights into the value stream and
stage that cannot meet demand or has the longest through- identify opportunities for improvement that might not be
put time, typically where a bottleneck occurs [178]. Once immediately apparent to human observers. Figure 14 shows
a bottleneck is identified using VSM, waste reduction and how data analytics and AI can be integrated with VSM.
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5361

Fig. 13  Kaizen via AI

Table 7  Executing VSM elements via AI

Element AI potential and algorithm

Customer Using NLP to analyze customer behavior, feedback, and market trends to understand customer needs better and adjust the
value stream accordingly
Supplier Using NLP to analyze supplier performance data to identify potential issues and areas for improvement
Process steps Analyze the efficiency of each process step and identify waste or inefficiencies. RL or other optimization algorithms can sug-
gest ways to streamline these steps
Material flow Optimize the path products take from raw materials to finished goods. This can involve everything from reducing unneces-
sary movements (using path optimization algorithms like Dijkstra’s or A*) to optimizing inventory levels (using forecasting
methods like ARIMA or LSTM)
Information flow Streamline the flow of information, ensuring that necessary instructions and updates reach the right place at the right time.
This can involve using algorithms for process optimization or even ML models to predict essential information based on
past patterns
Data collection Automate the process of collecting data from various sources across the value stream. IoT sensors can gather real-time data
about machine performance, process times, and product quality, which can then be fed into an AI system for analysis
Process data Analyze data about each process step to identify patterns, correlations, or anomalies that might suggest potential problems or
areas for improvement. SLAs, such as DT, regression models, and clustering, can be utilized here
Lead time Analyze lead time and process time data to identify bottlenecks and sources of delay. Time series analysis can predict future
lead times based on historical data, and anomaly detection algorithms can identify unusually long lead times that may
indicate a problem
5362 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 14  Integration of AI and VSM

3.6 Total productive maintenance (TPM) three factors is determined by the six big losses of Lean
Manufacturing. Availability losses are caused by equipment
The 6S is TPM’s corner foundation, which was developed in failure and breakdown rates and set up and adjustment times.
the 1960s [179]. TPM is a systematic approach to maintain- Performance losses are caused by the duration of time for
ing and improving the integrity of production and quality idle and minor stops and during the times of reduced speed.
systems through the machines, equipment, processes, and And quality losses are caused by startup and production
employees that add business value to an organization [180]. rejections. These six big losses are the most common causes
TPM focuses on proactive and preventative maintenance to of waste in Lean Manufacturing and must be reduced in
maximize the operational efficiency of equipment [181]. It the manufacturing process to improve the OEE values [187,
blurs the distinction between the roles of production and 188]. Table 8 shows the eight pillars of TPM and how AI
maintenance by strongly emphasizing empowering operators can help in their implementation.
to help maintain their equipment [182]. The main objec- By implementing TPM, companies can experience
tives of TPM are to increase equipment effectiveness, create decreased machine downtime, improved product quality,
a comprehensive maintenance system for the equipment’s increased equipment productivity, better morale among
entire life cycle, and involve every employee, from top man- workers, and a safer work environment. Figure 15 shows
agement to the shop floor workers [183]. TPM is a critical the relationship between AI and maintenance strategies.
element of Lean Manufacturing as it helps to reduce waste
and increase efficiency. This results in a higher benchmark 3.7 Poka‑Yoke
value of OEE. This value depends on availability, perfor-
mance, and quality [184]. Availability depends on the actual Poka-Yoke, a term derived from Japanese, means “mistake-
production time out of the total available production time, proofing.” It is a mechanism or technique used in Lean
while version describes how fast the machine produced dur- Manufacturing to avoid (Yokeru) inadvertent errors (Poka)
ing the actual production time [185]. Quality refers to mak- [189]. The goal is to design a process so that it is impossible
ing accepted goods from the total output [186]. Each of these to make mistakes [190]. If an error cannot be prevented,
Table 8  TPM implementation via AI
Pillar Description AI impact

Focused improvement Kobetsu Kaizen is a systematic approach to identifying and eliminating AI algorithms, particularly ML models, can provide insights for predictive
losses/wastes maintenance (PdM) and prescriptive maintenance (RxM) strategies; notify-
ing operators when maintenance is required. Predictive models, such as
regression analysis, RF, or SVM, can be used for this purpose.
Autonomous maintenance Jishu Hozen are steps to equip the operators to maintain, clean, and inspect AI can identify areas for improvement by analyzing production data and
their equipment identifying inefficiencies or waste. SLAs such as neural networks or ULA—
clustering can be used here.
Planned maintenance Approach to prevent unplanned downtime AI can help by analyzing historical data to predict when maintenance will be
required as a preventive maintenance (PM) strategy. Time series analysis or
LSTM can be used for this purpose.
Training and education Training and Yojishu develop the necessary skills among employees to work AI can automate quality checks using computer vision and CNN to identify
autonomously and understand, maintain, and improve their processes defects or inconsistencies.
Early equipment management Actions to get new equipment to its optimal speed faster AI can analyze historical data from similar equipment to provide insights that
inform the design process. Regression models, DT, or RF algorithms can be
used.
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Quality maintenance Activities to set equipment conditions that preclude quality defects AI can be used in personalized learning systems that adapt to the individual’s
learning style and pace or to identify gaps in knowledge or skills. Algo-
rithms for recommendation systems, like collaborative filtering or content-
based filtering, can be employed here.
Office TPM Applying TPM techniques to administrative functions to improve efficiency Computer vision can identify safety hazards or monitor for compliance with
safety procedures. CNN can be used for image recognition tasks related to
safety and health.
Safety, health, environment (SHE) Ensuring a safe and healthy workplace AI can automate data analysis, identify inefficiencies, and suggest improve-
ments in administrative processes. Algorithms like neural networks for
complex tasks or regression or DT models for straightforward tasks can be
employed.
5363
5364 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 15  Maintenance strategies via AI

Poka-Yoke should help detect the error early to prevent it These algorithms can be trained to classify products or pro-
from becoming a defect [191]. Poka-Yoke is a fundamental cesses as “pass” or “fail” based on predefined standards.
aspect of Toyota’s approach to Lean Manufacturing, part of Computer vision algorithms such as CNN can be employed
their quest for perfection through continuous improvement. to inspect products or processes in real-time and detect any
This approach can be applied to any type of process in any anomalies or defects in the contexts where visual inspection
industry to increase its efficiency and reliability [192]. There is needed. Figure 16 shows the AI potential to improve the
are two main types of Poka-Yoke, the Control Poka-Yoke Poka-Yoke operations.
and the Detection Poka-Yoke. The Control Poka-Yoke aims
to prevent the occurrence of an error. They control the pro- 3.8 Heijunka
cess so the operator cannot produce a defect [193]. An exam-
ple might be a fixture that holds parts in a certain way for Heijunka is a Japanese term that means “leveling.” In the
assembly, ensuring it is physically impossible to assemble context of Lean Manufacturing, it refers to leveling the type
them incorrectly. The Detection Poka-Yoke aims to detect and quantity of production over a certain period [195]. This
an error as it occurs, thereby preventing it from becoming technique reduces fluctuations in demand and production,
a defect. This usually involves some form of inspection or enabling a smoother, more efficient production process and
checking mechanism [194]. For example, a sensor might supply chain [196]. The primary goal of Heijunka is to pro-
confirm whether a part has been installed correctly before duce goods at a constant rate so that variations in customer
allowing the process to continue. demand do not overburden production lines. This approach
AI can aid in implementing Control Poka-Yoke by provid- helps reduce waste and makes the manufacturing process
ing real-time monitoring and decision-making support. For more flexible and adaptable to the changing needs [197].
instance, AI systems can monitor machine performance and Heijunka operates on two dimensions: leveling by volume
stop the process if the machine operates outside predefined and leveling by type [198]. Leveling by volume involves
parameters. Suitable AI algorithms for Control Poka-Yoke ensuring that the total volume of items produced during a
might include anomaly detection algorithms such as IF, One- specific period remains constant, even if the mix of products
Class SVM, or Autoencoders. These algorithms can identify varies. This can reduce strain on resources and operations
unusual machine performance that might indicate a prob- by ensuring a steady workload [199]. The leveling by type,
lem. AI can enhance Detection Poka-Yoke by automatically on the other hand, involves varying the production of differ-
checking the product or process against predefined stand- ent kinds of products to respond to changes in demand. By
ards. If a product or process does not meet these standards, regularly changing which product is being manufactured,
the AI system can alert the operator or stop the process. manufacturers can meet varying customer demands without
ML algorithms suitable for the Detection Poka-Yoke might causing significant disruption or delay in production [200].
include classification algorithms such as DT, RF, or SVM. Implementing Heijunka can bring several benefits, such
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5365

Fig. 16  AI Poka-Yoke

as reduced inventory, improved lead time, better customer involves producing and delivering finished goods just in time
response, less overburden on the employees and equipment, to be sold, sub-assemblies just in time to be assembled into
and reduced resource waste [201]. Heijunka allows for a finished goods, and purchased materials just in time to be
smoother, more predictable production process that can bet- transformed into fabricated parts [204]. JIT is a crucial pillar
ter cater to customer needs while maintaining efficiency, a of TPS. It is based on a simple principle: to produce the right
core concept in Lean Manufacturing [202]. items at the right time in the right amounts. To effectively
AI can significantly contribute to implementing Heijunka. implement JIT, a company must accurately forecast demand
Accurate demand forecasting is critical to successful produc- [205]. The process requires close coordination with suppli-
tion leveling. AI can analyze historical sales data and other ers and an efficient production system capable of rapidly
relevant factors (like seasonality, market trends, and eco- transitioning between different product lines to respond to
nomic indicators) to predict future demand more accurately. changes in customer demand [206]. JIT must be coupled
Algorithms like ARIMA, SARIMA (seasonal ARIMA), or with very accurate demand forecasting and reduce supply
more complex ML models like LSTM can be used for time chain disruptions [207]. By ordering materials to arrive only
series forecasting. Once the demand is accurately forecasted, when needed for production, JIT could allow businesses to
AI can help devise a production schedule that meets this significantly cut the costs associated with storing, managing,
demand while evenly distributing the workload over time. and ensuring inventory, thus reducing inventory costs and
This involves complex decision-making considering multiple improving quality through smaller batch sizes and immedi-
constraints (capacity limitations, labor availability, etc.), which ate feedback so that any quality issues can be quickly iden-
can be handled efficiently using AI. RL algorithms (such as tified and corrected, reducing the risk of large batches of
Q-Learning or Deep Q-Learning) or optimization algorithms faulty products [208]. By producing goods to meet immedi-
such as GA or particle swarm optimization (PSO) can be used ate demand, JIT could allow businesses to avoid the ineffi-
for this purpose. Implementing Heijunka with AI can lead to a ciencies of overproduction and improve overall operational
more balanced and efficient production system with less waste, efficiency, thus increasing efficiency and flexibility through
lower costs, and improved responsiveness to customer demand. a quick response to changes in customer demand, allowing
Figure 17 shows the potential of AI in implementing Heijunka. companies to be more adaptable and responsive [209].
JIT requires accurate customer demand forecasting to
3.9 Just‑in‑time (JIT) know when and how much of a product to produce. AI
can analyze historical sales data and other factors (for
JIT is a production strategy that strives to improve a busi- instance, seasonality, market trends, etc.) to forecast
ness’s return on investment by reducing in-process inven- demand more accurately. Time series forecasting methods
tory and associated carrying costs [203]. It is a method that such as ARIMA or more complex models such as LSTM
5366 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 17  Heijunka via AI

can be used for this purpose. JIT aims to minimize inven- 3.9.1 Just‑in‑sequence (JIS)
tory levels. AI can predict when to order materials based
on anticipated demand, lead times, and desired safety JIS is a production strategy that aligns the manufacturing
stock levels. This can be achieved through Q-Learning sequence directly with the order sequence. It is essentially a
or Deep Q-Learning algorithms that can learn the opti- more precise version of JIT, where the production is timed to
mal time to order supplies to minimize holding costs and meet demand, and the production sequence is meticulously
avoid stockouts. JIT requires a flexible production sched- arranged [210]. JIS allows parts and components to reach
ule that can quickly respond to changes in demand. AI can the assembly line in the exact order and when to be used in
create optimized production schedules that reduce waste production [211]. This sequence corresponds directly with
and ensure resources are used efficiently. Optimization customer demand, ensuring that production is highly syn-
algorithms such as GA or PSO can be used for this pur- chronized to consumer needs. JIS is often used in industries
pose. For JIT to be successful, suppliers must be reliable where customization is high, and product complexity is sub-
and able to deliver materials quickly when needed. AI stantial, such as the automotive industry. For instance, cars
can analyze supplier performance data to select the best can be assembled in the exact sequence of customer orders,
suppliers and identify potential issues. DT algorithms, ensuring each vehicle on the assembly line is assembled with
clustering algorithms, or other classification algorithms the right configuration of options [212]. JIS offers the same
can be used for this purpose. benefits as JIT with the addition of reduced errors. When
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5367

Fig. 18  AI relationship with JIT and JIS

parts are delivered in the sequence they are supposed to be One of the challenges of JIS is that changes in demand
utilized, the risk of the wrong part being used is minimized can necessitate changes in the production and delivery
[213]. sequence. AI can provide real-time analysis of incoming data
5368 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 19  AI-based Kanban

to identify such changes and adjust the sequence accord- processing requirements. ML clustering algorithms such as
ingly. Real-time anomaly detection algorithms, such as IF or K-means and hierarchical clustering can be used to group
One-Class SVM, can be used to identify significant differ- similar articles based on factors like machine usage patterns,
ences in demand. Once that is done, AI can aid in determin- sequence of operations, part families, number of operators
ing the optimal sequence for producing and delivering parts needed, and skill level of operators.
based on factors like demand, lead times, and production
constraints. This involves complex decision-making that can 3.10 Kanban
be enhanced using AI algorithms (such as GA, PSO, or Deep
Q-Learning) to learn the optimal sequencing through trial Kanban is a scheduling system used in Lean Manufactur-
and error. Figure 18 shows the relationship between JIT and ing to optimize workflow and minimize ongoing WIP [218].
JIS with AI. The word “Kanban” in Japanese translates to “signboard” or
“billboard,” representing its visual nature. Kanban aims to
3.9.2 Cellular manufacturing identify potential bottlenecks in your process and fix them
so work can flow through it cost-effectively at an optimal
Cellular manufacturing is a process of manufacturing which speed [219]. Kanban uses visual signals to represent work
is a subsection of JIT/JIS. Cellular manufacturing aims to in different process stages, from request to completion. Kan-
move as quickly as possible to make a wide variety of similar ban limits the number of items in any stage of the process
products while making as little waste as possible [214]. In a and helps prevent overloading and identifying bottlenecks
cellular manufacturing layout, a manufacturing unit consists [220]. Rather than producing ahead of time according to a
of multiple cells organized in an assembly line. Each cell forecast, goods are only delivered when there is demand for
is dedicated to producing a single product type or similar them. This “pull” system helps reduce overproduction and
products. Equipment and workstations are arranged accord- waste [221]. For example, a card is attached to a batch of
ing to the sequence of operations, allowing for a smooth and work in a production environment. When the work is com-
efficient flow of materials and components through the pro- plete, the card is returned to the beginning of the process
duction process [215]. Implementing cellular manufacturing to signal that more work can be started. This ensures that
can improve material flow, reduce setup time, and reduce only the required amount of work is in the system at any
work-in-progress (WIP) [216]. However, this implementa- time, preventing overproduction and improving flow [222].
tion process can be complex and requires a thorough analysis Kanban helps align inventory levels with actual consump-
of processes and product families, as well as potentially sig- tion, creating more efficient processes, reducing waste, and
nificant changes to the layout of the production floor [217]. increasing flexibility [223]. Its simple implementation and
AI can assist in creating optimal cell designs by analyz- highly visual nature allow teams to understand and manage
ing historical production data to identify items with similar their work quickly. Kanban operates under limiting WIP to
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5369

Fig. 20  Jidoka principles

Fig. 21  Seven steps of SMED

balance demand and production capacity [224]. Too many not be enough inventory to meet demand, leading to delays
cards in the system can lead to overproduction, resulting and possible loss of sales [225]. Determining the correct
in excessive inventory and potential waste. Conversely, too number of Kanban cards and adjusting it based on chang-
few cards can lead to underproduction, where there might ing conditions is essential to managing a Kanban system.
5370 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 22  Relationship between AI and RCA​

This number may need to be adjusted as demand and other the manufacturing process and not just as an endpoint. By
conditions change. detecting and addressing problems immediately, companies
Determining the optimal number of Kanban cards to can enhance quality, improve efficiency, and reduce waste
minimize lead time and WIP while avoiding stockouts is [229].
a complex problem. AI optimization algorithms, such as AI can automate the process of monitoring machine per-
GA or PSO, can be used to find the optimal solution to this formance and stopping production if abnormal conditions
problem. Kanban relies on accurate demand forecasting to are detected due to internal or external factors such as a
determine how many Kanban cards (signals to produce or cyber-attack. Machine learning models can analyze sensor
move items) are needed. AI can analyze historical sales data data to identify anomalies that might indicate a problem.
and other factors to predict future demand. Time series fore- Anomaly detection algorithms such as IF, CNN, or Autoen-
casting methods like ARIMA or more complex ML models coders can be used for this purpose. Figure 20 shows differ-
like LSTM can be used for this purpose. Figure 19 shows ent principles of Jidoka.
the elements of an AI-based Kanban.
3.12 Single minute exchange of die (SMED)
3.11 Jidoka
SMED is a systematic approach to reducing the time to com-
Jidoka, also known as “autonomation,” is a vital principle of plete equipment changeovers. The “single minute” refers to
the TPS; it can be translated as “automation with a human the goal of reducing changeover time to the “single digit”
touch” [226]. Jidoka refers to stopping a production line or minutes (in other words, less than 10 min) [230]. SMED
process when a problem is detected. It empowers workers is crucial for reducing downtime and improving flexibil-
and automated systems to halt production when an anom- ity in manufacturing, allowing for smaller batch sizes and
aly, error, or defect is noticed, preventing defective products varied product runs [231]. SMED is especially valuable in
from moving on to the next stage of the process [227]. The industries requiring high mix, low volume production or cus-
principle behind Jidoka is that problems should be corrected tomization. SMED involves several techniques and stages
as soon as they occur, and the cause of the situation should [232]. The key lies in distinguishing between “internal”
be identified and addressed to prevent its recurrence [228]. setup operations, which can only be performed when the
This ensures that only products that have passed each stage process is stopped, and “external” setup operations, which
of the production process without issues are of high quality. can be done while the equipment runs. The goal is to convert
Jidoka highlights the importance of quality at every step of as many internal operations into external ones as possible,
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5371

enabling most setup tasks to be carried out without stopping roots are identified in terms of the machine, method, mate-
the machine [233]. Implementing SMED can dramatically rial, human resources, measurement, and mother nature (the
reduce changeover time, improving efficiency and reducing 6 Ms) [238].
waste in the manufacturing process [234]. AI can analyze textual data, such as maintenance logs or
AI can be used to optimize the sequence of changeover operator notes, to extract insights related to the root cause.
tasks, minimizing the total time required. RL or other opti- NLP techniques could be applied here, including topic mod-
mization algorithms (GA or PSO) can be used for this. In eling such as Latent Dirichlet Allocation (LDA) or sentiment
addition, by predicting when a machine will likely require analysis. Figure 22 shows the relationship between AI and
maintenance or a changeover, AI can help schedule these RCA.
operations to minimize disruption to the manufacturing pro-
cess. This can be achieved through predictive models like 3.14 Takt time
DT, RF, and SVM. Figure 21 shows seven levels of SMED.
Takt time is a concept in Lean Manufacturing represent-
3.13 Root cause analysis (RCA) ing the maximum time available to produce a product to
meet customer demand [239]. It aligns production speed
The “5 Why” is a systematic problem-solving technique with demand and is crucial in JIT manufacturing. The
used to find out the actual root cause by repeatedly asking term “Takt” is derived from the German word “Taktzeit,”
“why?” deeper and deeper [235]. RCA can be used in almost which translates to “cycle time.” It reflects the rhythm
any situation with a problem or defect. It is commonly used or beat of production that aligns with customer demand
in manufacturing, healthcare, and IT, but its principles can [240]. Understanding Takt time helps manufacturers iden-
be applied in virtually any context [236]. The main benefit tify the weakest link in production that limits the total
of RCA is that it helps to prevent the same issues from recur- throughput (bottleneck analysis) [241], balance produc-
ring by addressing them at the source. By correcting the tion, and improve efficiency [242]. Thus, Takt time helps
root cause, rather than merely addressing the symptoms or maintain a steady and uninterrupted movement of prod-
immediate issues, organizations can reduce the chances of ucts through a production system, reducing wait times,
the problem recurring, leading to more reliable and efficient inventory, and waste (continuous flow) [243].
operations [237]. Other methodologies of RCA include the AI can support implementing and optimizing continu-
fishbone diagram, where significant categories of problem ous flow, helping manufacturers reduce waste, improve

Fig. 23  DMAIC
5372 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

efficiency, and meet customer demand more effectively. root cause of the problem. Various statistical tools may be
AI can assist in implementing and optimizing Takt time used to understand the relationships between different fac-
through accurate demand forecasting. AI can analyze his- tors [246]. AI tools can significantly speed up the analysis
torical sales data and other factors like seasonality and phase by automatically sorting through massive datasets.
market trends, to predict future demand more accurately Complex algorithms can identify the root causes of issues
using models such as ARIMA, SRIMA, and LSTM. Once more rapidly and might uncover hidden variables that human
the Takt time is determined, AI can help create a produc- analysis might miss. Once the root cause has been identified,
tion schedule that matches this pace. It involves optimiz- potential solutions are developed, tested, and implemented.
ing the allocation of resources (machines, labor, materi- The goal is to rectify the problem or improve the process
als) to ensure that production keeps pace with the Takt based on the findings from the analyze phase [247]. ML
time. RL algorithms such as Deep Q-Learning can aid in models can run simulations on potential improvements, pre-
this task. AI can also analyze production data to identify dicting outcomes with higher accuracy. These predictions
bottlenecks. ML algorithms such as K-means clustering can be used to pinpoint the most effective solutions. The
or SVM can be used for bottleneck analysis. final phase involves maintaining the improved process by
continuously monitoring it and applying corrective measures
3.15 DMAIC as needed. Control charts and other tracking tools may be
used to ensure the process remains within acceptable lim-
DMAIC is an acronym that stands for Define, Measure, its [248]. AI can monitor various variables in real-time to
Analyze, Improve, and Control. It is a data-driven quality ensure that improvements are sustained. Should anomalies
strategy used to improve processes and is an integral part of occur, machine learning models can trigger alerts for imme-
the LSS methodology. The first step involves defining the diate action, preventing minor issues from becoming more
problem or project goals. During this phase, the team iden- significant problems. DMAIC is widely used for quality
tifies the issue, the process to be improved, and the objec- control, process improvement, and operational excellence.
tives. Stakeholders and project scope are also clearly defined Figure 23 depicts DMAIC levels.
[244]. AI can help in the initial scoping of the problem by
analyzing large datasets to identify patterns or anomalies
that may not be immediately obvious. NLP can mine cus- 4 Convolutional neural networks (CNN)
tomer reviews, internal reports, and other textual data to
clarify the issues that need addressing. In the “Measure” CNNs have found significant success in a broad spectrum of
phase, the current state of the process is evaluated, and base- fields due to their robust feature extraction and image rec-
line metrics are established. Data is collected to provide an ognition capability. In tasks akin to object detection, speech
objective basis for future comparison, offering insight into recognition, and NLP, the performance of CNNs closely
the process [245]. Advanced sensors and ML algorithms aligns with human capabilities. This has enabled the wide-
can capture and measure more granular and real-time data spread application of CNNs across numerous sectors [249].
than humans can manually accomplish. This real-time data Healthcare is one of the domains that greatly benefit from
collection improves the baseline metrics used for later the methodological advances of CNNs. Medical profession-
stages. The collected data is then analyzed to identify the als grapple with vast data volumes used for patient diagnosis,

Fig. 24  Basic CNN architecture

The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5373

monitoring, and treatment. CNNs, through their applications The “Zero Waste” and I4.0 initiatives have seen a rise in
such as medical image segmentation and electrocardiogram the application of CNNs in waste management processes.
processing, have proven to add more objectivity to diagno- Intelligent waste identification and recycling (IWIR) is one
sis, permit quicker and more accurate decisions, and allow domain where CNNs transform waste classification, detec-
for more personalized treatment protocols [250]. CNNs tion, and segmentation, which is pivotal for manufactur-
have also permeated the manufacturing and production sec- ing and production efficiency [251]. The prowess of CNNs
tors, creating a paradigm shift in operational efficiencies. extends into business and enterprise landscapes, including

Table 9  Dataset requirements of a CNN-based model

Requirement Impact

Volume The larger the dataset, the better the network can generalize its learned patterns to unseen data. Small datasets could result in
overfitting, where the model performs well on the training data but poorly on new, unseen data.
Quality Incorrect labels or poor-quality images can lead to a model learning incorrect associations.
Diversity Diversity in the dataset allows the model to generalize well and perform efficiently on unseen data. In classification tasks, diver-
sity can mean a range of images that show the object of interest in various positions, lighting conditions, or backgrounds.
Balance For classification tasks, a balanced dataset, where each class is equally represented, is beneficial. An imbalance, with some classes
having many more examples than others, can make CNN biased towards the overrepresented classes.

Table 10  Summary of selected performance measurements

Metric Definition

Accuracy This is the most straightforward metric. It is the ratio of the number of correct predictions to the total number of predictions. While it
can provide a general
/ idea of model performance, accuracy may not be the best metric in cases of class imbalance.
Accuracy = TP+TN TP+FN+TN+FP
Precision The ratio of the number of true positive predictions to the sum of true positive and false positive predictions. It is essentially a meas-
ure of a model’s
/ ability to correctly identify only relevant instances.
Precision = TP TP+FP
Recall Recall or sensitivity is the ratio of the number of true positive predictions to the sum of true positive and false negative predictions. It
measures the
/ model’s ability to identify all relevant instances.
Recall = TP TP+FN
F1 score The F1 score is the harmonic mean of precision and recall. It provides a balance between the two metrics and can be a better measure
than accuracy,
( especially
/ for imbalanced ) datasets.
F1 = 2 ∗ Precision∗Recall Precision+Recall

Fig. 25  CNN working mechanism

5374 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

marketing and inventory management. Their ability to iden- including convolutional, activation, pooling, and fully con-
tify and classify intricate patterns makes them instrumental nected layers [258]. Since the introduction of LeNet-5 in
in deciphering consumer behavior and preferences and opti- 1998, numerous innovations in CNN architectures have
mizing marketing strategies. For inventory management, the been proposed. These advancements can be classified
predictive capabilities of CNNs help to manage stock levels into parameter optimization, regularization, and structural
effectively, leading to operational efficiency and reduction reformulation. However, the performance improvement in
in waste [252]. In cybersecurity, CNNs are robust tools for CNNs primarily stems from the fundamental reformulation
identifying patterns related to cyber threats, malicious activ- and the design of new blocks [259]. CNN architectures are
ities, and system vulnerabilities. Their potential to process organized to address tasks like classification, segmentation,
large datasets in real-time allows for proactive detection and and object detection. Classification models use convolu-
mitigation of cyber threats, thus ensuring the security and tion and subsampling layers to encode input images into
integrity of enterprise operations [253]. CNNs are used to various steps, with the classifier being used to calculate
streamline various tasks. They are used to improve logistical class probability [260]. These models can also serve as
operations by optimizing route planning, predicting poten- feature extraction tools for segmentation and detection
tial bottlenecks, and enabling real-time tracking [254]. In tasks. Semantic segmentation models have seen several
the human resources (HR) domain, CNNs assist in talent introductions in recent years. A segmentation model usu-
acquisition by analyzing and categorizing vast amounts ally consists of encoding and decoding units, with fully
of candidate data to identify the best fit for a role. CNN is convolutional network (FCN) being the first segmentation
widely adopted in different operations and enterprise func- model proposed [261]. Object detection, on the other hand,
tions [255, 256]. aims to identify target types and their positions, answering
two questions: What is the object, and where is the object
4.1 Evolution of CNN in the image? Some of the most popular models for object
detection are Region-based CNN (R-CNN), You Only Look
CNNs are representative algorithms of DL; they are rapidly Once (YOLO), and Faster R-CNN [262]. Despite the power
becoming a critical component in the era of AI thanks to of CNN architectures in object classification, segmenta-
large-scale annotated data, powerful computing capability, tion, and detection tasks, they require large volumes of
and advanced architectures [257]. Primarily used for solv- data for training due to the numerous parameters and deep
ing various computer vision problems, CNNs are feedfor- layers involved. This requirement is often mitigated using
ward networks that extract features in images layer-by-layer methods such as transferring learning, data augmentation,
from input to output. They comprise multiple components, customizing layers, and optimizing hyperparameters. These

Table 11  The breakdown of cybersecurity threats in the TON_IoT dataset

Attack type Brief details

Distributed denial of service (DDoS) A DDoS attack attempts to disrupt the normal functioning of a network, service, or website by over-
whelming it with a flood of internet traffic. It is “distributed” because it originates from many different
sources, often botnets, making it difficult to stop [281, 282]
Ransomware This malicious software is designed to encrypt a victim’s files. The attacker then demands a ransom from
the victim to restore access to the data upon payment. Users are shown instructions on how to pay a fee
to get the decryption key [283, 284]
Backdoor Any method by which authorized and unauthorized users can get around standard security measures and
gain high-level user access (also known as root access) on a computer system, network, or software
application [285]
Injection Any attack that involves inserting harmful data into a system. A typical example is SQL Injection, where
malicious SQL code is inserted into an entry field for execution, often to manipulate the system’s data-
base [286]
Cross-site scripting (XSS) This occurs when a web application gathers malicious data from a user. The data is usually in the form of
a hyperlink containing harmful scripts and is most often launched from trusted sites, leading to scripts
being executed as coming from trusted sources [286]
Password cracking This refers to attempting to gain unauthorized access to a system or sensitive data by guessing passwords,
often using specialized software tools that generate high numbers of possible password combinations
[285, 286]
Scanning This identifies potential vulnerabilities in a computer, network, or communications system. It is often the
preliminary stage of an attack where the perpetrator seeks out weak points in the system’s security that
they can exploit [287]
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5375

Fig. 26  Percentage breakdown

of attacks

techniques facilitate the training of CNNs, allowing them several cases of the application of custom-built CNNs in
to deliver superior performance in various applications. different areas of the manufacturing enterprise.
Figure 24 shows the standard architecture of CNN that was Before diving into Sect. 5, it is worth mentioning that
initially used in image processing. data splitting is a technique commonly used in ML to divide
a dataset into distinct subsets for training, validation, and
4.2 Dataset requirements for CNN‑based models testing purposes. This is a crucial step in the model develop-
ment process, as it helps to ensure that the model is learning
Training CNNs typically requires extensive, diverse, and generalized patterns rather than memorizing the data, which
high-quality datasets. The necessity for such data stems from could lead to overfitting. Typically, a dataset is divided into
the complexity of these models and their depth, involving a training set used to train the model, a validation set for
multiple layers of neurons and a multitude of parameters tuning hyperparameters and model selection, and a test
that must be optimized during training [263, 264]. Table 9 set for assessing the model’s generalization performance.
lists these requirements and their potential impacts on the Typical ratios for splitting might be 70-15-15, 60-20-20, or
performance of a CNN. 80-10-10, among others, depending on the specific project
and dataset size. Some specialized techniques for data split-
4.3 Utilized metrics to evaluate the performance ting include k-fold cross-validation, stratified sampling, and
of CNN‑based models time-based splitting, especially for time series data. Data
splitting is applicable across various types of ML, including
Performance estimation aims to estimate the loss a pre- supervised, unsupervised, and reinforcement learning.
dictive model will incur on unseen data. This process is
a fundamental stage in any ML project [265]. Evaluat-
ing the performance of a CNN is crucial to assessing its 5 Case studies: applications of CNNs
effectiveness and accuracy in making predictions. There utilizing different datasets
are several metrics used to evaluate the performance of a
CNN model. Some of these metrics are calculated based 5.1 Detection of cybersecurity threats in industrial
on the confusion matrix, a valuable tool for understanding IoT devices
the version of a classification model. It presents the true
positives (TP), true negatives (TN), false positives (FP), The escalating integration of sensor technology within conven-
and false negatives (FN) in a tabular form. Table 10 lists tional manufacturing systems has escalated their susceptibility
some of these metrics used in Sect. 5.1–5.4 of this paper. to cyber threats [266]. These threats are increasing both in
Each of these metrics provides different insights, and frequency and potential harm. Beyond merely stealing intel-
their importance can vary depending on the specific task lectual property (IP), these attacks can extend to damaging
and the business context. Therefore, considering multiple manufacturing equipment, manipulating product designs and
metrics when evaluating a model’s performance is often blueprints, and altering manufacturing processes [267, 268].
necessary. Figure 25 shows a working mechanism of a In the first month of 2022 alone, the Center for Strategic and
CNN-based model. The following section will discuss International Studies (CSIS) reported over 11 cyber threats.
5376 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 27  Custom CNN architec-

ture used in detecting threats

These threats included penetrating state-operated servers

worldwide, such as the Belarusian Railway, to compromising
private companies, like during a Minecraft tournament [269].
Table 12  Average values of performance metrics This demonstrates the wide-reaching capabilities of cyber
Industrial IoT device Accuracy Precision Recall F1 score threats, affecting various segments of a nation’s economy.
Among these sectors, the manufacturing industry was the
Fridge sensor 96.65% 93.30% 93.27% 93.24%
most susceptible. Typically, these attacks target organizations
Garage door sensor 99.78% 99.45% 99.45% 99.45%
by exploiting their data traffic in cyberspace to disrupt, crip-
GPS tracker 99.17% 97.13% 97.09% 97.11%
ple, damage, or seize control of a server’s environment and/
Motion light sensor 94.51% 89.38% 89.28% 89.31%
or infrastructure [270]. To mitigate the impact of such threats,
Modbus device 99.94% 98.69% 98.69% 98.69%
several researchers have suggested various models to detect
Thermostat sensor 99.71% 98.14% 98.13% 98.14%
and minimize the damage they cause, including network intru-
Weather sensor 98.97% 97.36% 97.28% 97.31%
sion detection systems (NIDS) [271, 272].
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5377

5.1.1 Utilized dataset are scaled using SKlearn Robust Scaler to enhance outlier
robustness [290]. A split of 90/10 was used for training and
The TON_IoT dataset, a newly minted resource in the testing, respectively. The structure of the CNN model com-
domain of I4.0, first saw the light of day in 2020 [273–275]. prises dual 1D convolutional hidden layers operating over a
An academic and industry collaboration between the Engi- 1D sequence with a set of 32 filters (kernels), each having
neering and Information Technology School, Cyber Range a size of 2. These kernels retain values acquired during the
and IoT Labs of the University of New South Wales (UNSW) training phase. Note that batch normalization is employed
Canberra, and the Australian Defense Force Academy alongside each convolutional hidden layer to keep the mean
(ADFA) gave birth to this dataset [276–280]. This dataset output nearing 0 and the output standard deviation approach-
is explicitly designed for I4.0 networks, which encompass ing 1 by applying a transformation to normalize the input.
IoT and industrial IoT networks. The driving force behind These hidden layers are used for feature extraction. The
the development of such a dataset was to gauge the perfor- model uses rectified linear activation (ReLU) to enable the
mance and reliability of various cybersecurity applications learning of more complex functions, thus enriching the train-
powered by AI. The TON_IoT dataset includes seven IoT ing outcomes. Subsequently, a GlobalAveragePooling1D
devices: a fridge sensor, garage door sensor, GPS tracker, layer follows ReLU, ensuring a considerable amount of data
motion light sensor, modbus communication device, thermo- about the “less critical” outputs is maintained. To prevent
stat sensor, and weather sensor. Each device has undergone overfitting, dropout layers with a probability value of 0.2 and
the array of attacks outlined in Table 11. Figure 26 displays 0.3 are added, meaning a portion of the layer’s output will be
the proportional distribution of each type of attack instance dropped. The optimal number of LSTM cells is determined
across all seven IoT devices, drawing from a total pool of to be 100. The models utilize GlorotUniform or Xavier
407,932 records. Note that “Bening” refers to normal traffic Uniform as the default weight initializer. The models incor-
devoid of attacks. porate an attention mechanism to assist in the algorithm’s
learning process. Both models apply the Adam optimization
5.1.2 Methodology algorithm, maintaining a steady learning rate of 0.03. A con-
catenation layer follows the dropout layers, which arranges
The first action in any network intrusion detection process all previous outputs into a single vector-valued column along
is gathering data by simulating it or from diverse online a specific dimension. Figure 27 shows the architecture of the
sources, as exemplified in our usage of the TON_IoT dataset CNN model that was built for this task.
[288]. Following this, the dataset undergoes a preprocess-
ing phase before it is fed into the algorithms. Both majority 5.1.3 Results and discussion
and minority types are handled by combining down-sam-
pling and oversampling, respectively. Categorical features While not a traditional element of Lean Manufactur-
are encoded through a one-hot encoder [289], and features ing, cybersecurity has become increasingly crucial as

Fig. 28  Lean security for efficient manufacturing

5378 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

manufacturing processes become more digitally connected flow to customers. Strong cybersecurity can be seen as a fun-
and reliant on technology. Improving cybersecurity can damental part of Lean Manufacturing in the modern digital
align with and support the goals of Lean Manufacturing. age, as it supports many of the core goals and principles of
Table 12 shows the average accuracy, precision, recall, and Lean. It is an area that requires continuous attention and
F1 score values for each industrial IoT device. It can be seen resources, but the payoff is well worth it in terms of support-
from that table that our custom-built CNN architecture has ing lean operations and overall business success.
achieved high performance on all the sensors and devices Furthermore, leaders in the manufacturing sector
when benchmarked. oversee the security of various IT elements. This
Lean Manufacturing aims to eliminate waste, and cyber- includes user endpoints operating on Windows and
security incidents can cause significant waste. Cyber-attacks macOS and depreciated machinery utilizing outdated
can waste time and resources as employees deal with the Windows versions that have long ceased to be supported.
fallout. This can include lost production time, costs asso- Furthermore, the manufacturing space is saturated with
ciated with repairing compromised systems, and resources Operational Technology (OT). Big data offers real-
spent on damage control. Lean Manufacturing emphasizes time and clear insights into network-connected devices,
the creation of value for the customer. Cybersecurity can functions, communication patterns, and vulnerabilities.
contribute to this by ensuring the protection of customer It becomes an essential component of the comprehensive
data, thus enhancing the company’s reputation and increas- risk management strategy that an enterprise might
ing customer trust and satisfaction. Lean Manufacturing pursue. To secure a Lean Manufacturing enterprise
thrives on predictable, stable operations. Cybersecurity pro- effectively, ideally, all these considerations should
tects the reliability of the systems, processes, and technolo- converge into a NIDS via CNN (or any other ML-based
gies that drive these operations. This enables manufacturing method that has the potential to achieve superior results).
to proceed without interruption, ensuring a steady product Simplicity and automation are critical for maximizing

Table 13  AI4I 2020 Predictive Maintenance Dataset attributes description

Attribute Description

UID Each entry is assigned a unique number, ranging from 1 to 10,000, excluded from our analysis
Product ID Comprises of a prefixed letter (L, M, or H) that corresponds to low (50%), medium (30%), or high (20%) prod-
uct quality variants, followed by a unique serial number
Type Characterizes the product type as low (L), medium (M), or high (H), each representing 50%, 30%, and 20% of
all products, respectively
Air temperature It is generated via a randomized process and later normalized to exhibit a standard deviation of 2 K around 300
K value
Process temperature Formulated via a randomized process, normalized to indicate a standard deviation of 1 K, and subsequently
added to the air temperature plus 10 K
Rotational speed Calculated from a power of 2860 W, overlaid with normally distributed noise, and measured in revolutions per
minute (rpm)
Torque Torque measurements are centered around 40 Nm with a standard deviation (σ) of 10 Nm and do not feature
negative values
Time to tool wear Depending on the product quality variant (H/M/L), the process adds 5/3/2 min of tool wear. Also included is a
“machine failure” label, which signifies if a machine has failed at this specific data point due to any identified
failure modes
Tool wear failure (TWF) Occurs when the tool requires replacement or fails within a randomly selected tool wear time of 200–240 min
(occurring 120 times in our dataset). Tool replacement and tool failure events occur 74 times and 46 times,
respectively
Heat dissipation failure (HDF) If the difference between the air and process temperature falls below 8.6 K and the tool’s rotational speed is
under 1380 rpm, heat dissipation causes a process failure. This scenario occurs for 115 data points
Power failure (PWF) The product of torque and rotational speed (in rad/s) should be the power necessary for the process. A process
failure ensues if this power falls below 3500 W or surpasses 9000 W, occurring 95 times in the dataset
Overstrain failure (OSF) The process fails if the multiplication of tool wear and torque exceeds 11,000 min Nm for the L product variant
(12,000 for M, 13,000 for H) due to overstrain. This holds for 98 data points
Random failures (RNF) Independent of its process parameters, each process has a 0.1% probability of failing. This happens for 19 data
points in our dataset, which is higher than expected for 10,000 data points
Predicted machine failure If any of the above failure modes are true, the process fails, setting the machine failure label to 1. This scenario
is valid for 339 data points
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5379

Fig. 29  Custom CNN architec-

ture used in failure predictions

Table 14  CNN performance measurement 5.2 Early tool failure detection

Model Accuracy Precision Recall F-score
A thriving manufacturing setup generally hinges on
CNN-CNN 0.68 0.66 0.68 0.64 superior design standards and a well-structured strategy for
maintenance to prevent system breakdowns. The expense
of maintenance for an enterprise is immense, taking up
the effectiveness of scarce security resources, allowing a sizeable fraction of the total cost of production [291].
them to focus on more pressing concerns. Similar to The surge in maintenance costs over the last few years is
how automation revolutionized the manufacturing floor, attributed to technological progress. Thus, an empirically
it has significantly impacted security across endpoint, based upkeep strategy aids in mitigating equipment failures
cloud, and IoT. Automation removes the burden of and averting costly production process disruptions [292].
mundane tasks, improving the efficiency of your current Given this, considerable focus has been directed towards
security staff and allowing them to concentrate on more maintenance strategies in manufacturing entities. This
sophisticated security tasks. This enhances the quality focus is geared towards prolonging the operational life
and output of the NIDS, drastically reducing the average of tools and enhancing system availability [293], thereby
response time to seconds. Figure 28 shows how cyber mitigating reliability risks [294] and reducing the chances
resilience can enhance operations continuity. of interrupted production. Manufacturers should train,
Integrating DMAIC (Define, Measure, Analyze, Improve, plan, and optimize the usage and implementation of their
Control) into the NIDS can enhance cybersecurity in an maintenance strategies to maximize operational time and
intelligent factory with industrial IoT devices. The “Define” minimize downtime [295]. To monitor the functioning
phase initially sets clear objectives for what constitutes an and health status of equipment, sensors are deployed to
intrusion and identifies critical assets. The “Measure” phase gather data on aspects such as temperature, vibration, and
establishes key performance indicators and collects baseline other pertinent parameters. This data can then construct a
data through logs and sensors. The heart of the integration decision-making model [296]. The model scrutinizes the
lies in the “Analyze” and “Improve” stages. ML algorithms sensor-collected data and formulates a collection of decision
can scrutinize network traffic to identify anomalies, serv- guidelines. These guidelines can predict potential equipment
ing as a basis for root cause analysis if breaches occur. This breakdowns and suggest appropriate maintenance actions.
insight feeds into the “Improve” stage, where algorithms are Once the pivotal parameters associated with component
refined and system components possibly upgraded to meet failure have been identified, the model can offer an optimal
security needs better. Lastly, the “Control” phase focuses on maintenance action based on the anticipated mode of
maintaining the enhanced state. Real-time analytics moni- failure [297]. The decision-making model can also undergo
tor the network to flag anomalies, with automated responses real-time updates as new data is collected, enhancing
in place to act on detected intrusions. Performance metrics its predictions continually. This methodology can assist
are regularly reviewed, and the system is updated as needed. companies in minimizing downtime and boosting equipment
Through DMAIC, the NIDS becomes more robust, agile, and reliability by allowing them to detect and rectify issues
capable of evolving to meet new cybersecurity challenges. before they culminate in equipment breakdowns [298].
5380 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 30  Lean maintenance for efficient manufacturing

5.2.1 Utilized dataset this, we utilize one-hot encoding to convert categorical

inputs into binary form for more precise depiction. For
In light of the difficulty encountered in acquiring actual feature scaling, while ensuring resilience to outliers, the
maintenance datasets that are publicly accessible [299], we SKlearn Robust Scaler was employed. The CNN dense
relied on the “AI4I 2020 Predictive Maintenance Dataset,” layers used the Softmax activation function, while the
a synthetic dataset [300]. The dataset, made available by the hidden layers utilized the rectified linear activation (ReLU)
Engineering Department of the University of Applied Sci- function. In addition, it was trained for more than 14 epochs.
ences, located in Berlin, Germany, comprises 10,000 data Figure 29 shows the architecture of the CNN model that was
entries, each with 14 attributes shown in Table 13. In addi- built for this task.
tion, the dataset exhibits significant skewness, with only 339
instances identifying machine breakdown. While a 3.39% 5.2.3 Results and discussion
failure ratio (339 out of 10,000) typically echoes the patterns
seen in actual industrial control platforms, it could also be Table 14 shows the average values of the performance meas-
perceived as distressing within high-volume manufacturing urements of the built CNN model. According to the results
contexts. The data was divided using a 90/10 ratio for train- shown in the table, CNNs have performed poorly, thus high-
ing and validation purposes, respectively. lighting a significant concern when deploying AI systems
in manufacturing. CNNs usually need an extensive amount
5.2.2 Methodology of data to discern underlying trends effectively. If the data-
sets are not large enough, the CNNs may overfit or underfit
Since CNN requires numerical input, non-ordinal categorical the data, resulting in less-than-ideal performance. It is also
features could pose a challenge if represented as integers. critical to acknowledge that the efficiency of CNNs relies
Giving each category a numerical value could misguide on the particular problem, dataset, and model design. Even
the model, potentially resulting in unintended results or with the addition of an LSTM that is coupled by an attention
diminished efficiency, especially if the model presumes mechanism, the newly generated architecture has performed
an inherent sequence among categories. This could lead to less than the original proposed CNN model for this failure
estimations that land in between categories. To circumvent detection task.
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5381

Fig. 31  Steel surface defects

Early tool failure detection is an integral part of a Lean steel plates are vital in various sectors such as automotive,
Manufacturing system. Detecting tool failures early helps national defense, machinery, chemical, and light industries.
prevent the waste (Muda) of materials, energy, and effort The term “defect waste” refers to the production of any item
from running a faulty process. It also avoids generating that is not fit for its intended use, such as flawed materials,
defective products, which would otherwise need to be components, or products not made according to the required
reworked or discarded. This aligns with the lean principle specifications [302]. Producing steel is intricate, involving
of getting things right the first time rather than relying on several steps like heating, rolling, drying, and cutting. Dur-
inspections and corrections downstream. By identifying ing these phases, the steel surface may interact with many
potential issues before they lead to a breakdown, produc- machines before being prepared for dispatch. As a result,
tion can continue without unexpected stops and starts, thus several potential defects on the steel plate might arise that
enhancing the flow (Jidoka). Early detection of tool failures need identification. Flaws like cracks, scabs, curled edges,
allows for proactive maintenance or replacement of tools, cavities, and abrasions on the surface can appear due to raw
ensuring they are ready when needed and avoiding delays material and technological issues, negatively impacting the
that could disrupt the JIT process. Early tool failure detec- factory’s corrosion resistance, strength, and financial prof-
tion provides valuable data that can be analyzed to identify itability [303]. The growing steel demand has underscored
trends, root causes, and opportunities for process improve- the need for an efficient manufacturing system in the steel
ments. This aligns with the principle of continuous improve- processing industry that can produce products that match
ment, as it promotes learning and adaptation, leading to bet- consumers’ quality expectations. Manual detection is the
ter performance over time. Figure 30 shows the principle of prevalent method for detecting defects on steel surfaces, but
Lean maintenance. it increases labor costs, is subject to personal bias, and often
produces inaccurate and unreliable results [304]. Detecting
5.3 Metal surface inspection these defects on metal surfaces is a challenge due to their
random occurrence. This is due to the ambiguity between
The worldwide steel demand has been growing yearly, pre- normal and defect, the tiny defective area, and, in some
dominantly used in infrastructure developments. In 2021, cases, the drastic differences in appearances among different
the USA consumed 98 million metric tons [301]. Moreover, defects, thus making them hard to detect using manual labor
5382 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 32  Architecture of the custom-built CNN algorithm used in steel surface inspection

Table 15  Values of performance measurements steel above its recrystallization point, typically between 1000
and 1200 °F, simplifying its shaping and molding. Moreo-
Defect Accuracy Precision Recall F1 score
ver, this process allows the steel to be manipulated with-
Cr 99.63% 98.37% 99.45% 98.91% out causing any cracks. Hot-rolled steel is often employed
GG 99.27% 86.27% 97.78% 91.67% when exact shapes and measures are not necessary, such as
In 99.82% 100.00% 98.29% 99.14% structural steel parts in the construction sector or machin-
Pa 99.36% 98.26% 97.69% 97.97% ery and equipment manufacturing. Its popularity stems from
PS 99.08% 93.02% 99.17% 96.00% its strength, resilience, and cost-effectiveness. A distinctive
RD 98.99% 82.76% 80.00% 81.36% characteristic of hot-rolled steel is its scaly surface texture
RS 100.00% 100.00% 100.00% 100.00% and a bluish-gray oxide layer that emerges on the surface
Sc 99.18% 98.21% 94.02% 96.07% as it cools down [308]. A dataset featuring steel surface
SP 98.99% 96.67% 86.57% 91.34% imperfections, named NEU-CLS-64, has been made avail-
able by Northeastern University (NEU) [309]; the dataset
includes over 7000 grayscale images covering nine typical
[305]. Computer vision methods can supplement or replace defect types found in hot-rolled steel strips. The specified
manual detection in the defect classification process, so it defects are made up of rolling scale (RS), plaque (Pa), crack-
is crucial to explore the use of computer-based vision algo- ing (Cr), pitting surface (PS), inclusions (In), scratches (Sc),
rithms to automatically detect surface defects on the produc- grooves and gouges (GG), rolling dust (RD), and spots (Sp).
tion line [306]. Automated defect detection is a subset of An overview of these imperfections is shown in Fig. 31.
object detection that specifically identifies defects on the sur- Detecting these surface defects earlier can help eliminate
faces of various industrial products like steel, glass, printed their upstream root cause [310].
circuit boards (PCBs), wafers, wood, and more. Given the
significance of steel in numerous areas such as construction, 5.3.2 Methodology
vehicle parts, and infrastructure, introducing an automated
system for defect detection on steel surfaces could improve The dataset underwent a process known as image augmen-
the quality of steel products [307]. tation. This is a prevalent method in computer vision and
deep learning, where existing training images are manip-
5.3.1 Utilized dataset ulated using diverse transformations to generate slightly
altered copies of the original data. These manipulations can
Hot-rolled steel strips refer to thin steel sheets shaped using include a wide array of adjustments such as rotation, scal-
the hot-rolling method. This procedure involves heating the ing, cropping, flipping, the addition of noise, or alterations
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5383

in brightness or contrast, among others. Exposing the model detecting RD defects, making it the hardest to detect with
to a wide variety of possible inputs through these transfor- a high F1 score. Overall, the model performed very well in
mations can enhance the model’s ability to generalize to the detection process.
new, unseen data. Image augmentation is especially advan- The ability to detect defective products early in a manu-
tageous when the training data at hand is limited or shows facturing process is an integral facet of a Lean system. It
an imbalance, as it can effectively boost the quantity and intersects perfectly with a variety of Lean principles. Early
diversity of the training set without collecting more data. detection of defects is pivotal in minimizing the waste of
The main layers of the CNN used in this task include input, materials and resources. Rather than allowing faulty prod-
convolutional, activation, pooling, fully connected, dropout, ucts to traverse the entire production line, consuming time,
and output. The input layer is the initial layer that accepts an energy, and materials at each stage, any defects can be
image as a matrix of pixels. Depending on the image’s color swiftly addressed. This process conserves resources and
scheme (e.g., colored or grayscale), the number of chan- curtails expenses. Lean systems aspire to incorporate qual-
nels might differ. Convolutional layers apply a set of filters ity into every step of the process. When faults are detected
(kernels) to the input image or the output of previous lay- early, the approach inherently becomes self-regulating.
ers. Convolutional layers help to detect local features in the This means issues are caught and rectified as part of the
image, such as edges, shapes, and textures, which are then regular workflow rather than treated as an anomaly or an
used for the classification task. The non-linear activation afterthought. Lean thinking places considerable emphasis
function, such as ReLU, is applied after every convolutional on creating value for the customer. In this mindset, any-
layer. These functions add the necessary non-linearity to the thing that does not offer value to the customer is deemed a
model, which allows it to learn from complex data. Pool- waste. Naturally, defective products fall into this category,
ing layers are used to down-sample the feature maps while as they are unsaleable and even if they are sold, they lead
preserving the most important information. This helps to to customer dissatisfaction. Thus, by detecting and rectify-
reduce the computational complexity and to provide some ing defects early, Lean systems ensure that every product
degree of translation invariance. There are different types of that reaches the customer’s hands is valuable to them. The
pooling operations, with max pooling and average pooling ability to detect defective products equips the system with a
being the most common. After several iterations of apply- consistent source of feedback about the production process.
ing convolutional and pooling layers, the output is flattened This information is instrumental in identifying bottlenecks
(converted from a 2D feature map to a 1D vector) and fed or inefficiencies in the design, thereby serving as a founda-
into one or more fully connected layers (also known as dense tion for continuous improvement efforts. Finally, the early
layers). These layers perform high-level reasoning and make detection of defective products in a manufacturing process
the final decisions based on the features extracted by the significantly bolsters its lean nature. It eliminates waste,
preceding layers. The last fully connected layer has the same ensures quality, delivers value to customers, and promotes a
number of neurons as there are classes in the classification culture of continuous improvement.
task. Dropout layers can be inserted between layers to reduce
overfitting. During training, dropout layers randomly set a 5.4 Skin disease detection
fraction of input neurons to 0 at each update, which helps
prevent the model from relying too much on any feature. Skin cancer is the most prevalent form of cancer worldwide,
The final layer of the network, typically utilizing a Soft- affecting one in five individuals by age 70. Many entities are
Max activation function, produces the probabilities for each diligently working to alter this grim statistic. The journey
class. The class with the highest chance is selected as the begins with excising a questionable skin sample for labora-
output prediction. Figure 32 shows the architecture of the tory analysis, known as a skin biopsy. Such a biopsy can
custom-built computer vision CNN algorithm for steel sur- confirm the presence and type of skin cancer [311]. Skin
face inspection. cancer is alarmingly the most common cancer diagnosis in
the USA, with over 5 million new cases recorded yearly,
5.3.3 Results and discussion a figure that continues to climb. The impact of skin can-
cer extends far beyond health, contributing to a significant
Table 15 shows the performance benchmarks for the cus- economic burden, with treatment costs exceeding $8 bil-
tom-built CNN algorithm used in detecting steel surface lion annually in the USA alone [312]. The rise in cancer
defects. According to the results in the table, the RS defect incidents is linked to factors like ethnicity, age, gender,
was detected 100% of the time, and the model performed smoking, environmental shifts, harmful sun or radiation
best when detecting this defect. This could be because it exposure, viruses, alcohol, poor diet, inadequate exercise,
appears as a dark, flaky residue on the surface of the steel. and past medical history [313]. Recent data indicate that
Furthermore, the model had the lowest performance when melanoma, a deadly skin cancer type often mistaken for
5384 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 33  Breakdown of the seven classes before and after augmentation

Fig. 34  Architecture of the custom-built CNN image algorithm used for skin disease detection

harmless skin lesions, is responsible for an estimated 6750 Table 16  Values of performance measurements
deaths in the USA each year [314]. Encouragingly, most
patients can be cured if diagnosed and treated early enough Class Accuracy Precision Recall F1 score
before the disease progresses to a fatal stage [315]. However, nv 98.00% 98.94% 87.73% 93.00%
timely diagnosis with accurate screening methods is cru- mel 98.93% 93.93% 99.02% 96.41%
cial to prevent skin cancer from metastasizing to other body bkl 99.17% 94.71% 99.69% 97.14%
parts, diminishing the success of treatments and heighten- bcc 99.86% 99.03% 100.00% 99.51%
ing the risk of mortality [316, 317]. A failure to diagnose akiec 99.98% 99.84% 100.00% 99.92%
early has been linked to less than 14% survival rates, while vasc 99.99% 99.92% 100.00% 99.96%
early detection can boost survival rates to nearly 97% [318]. df 99.99% 99.92% 100.00% 99.96%
The diagnostic process generally involves a dermatologist
examining suspicious lesions visually, then an examina-
tion under polarized light magnification, dermoscopy, and poses three main challenges: reliance on the dermatolo-
finally, a biopsy [319]. This diagnostic approach, however, gist’s expertise (with evidence suggesting that even the most
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5385

skilled professionals only accurately diagnose skin cancer All surviving images were initially taken using both polar-
in about 80% of cases), the lengthy time span between the ized and non-polarized dermatoscopy. To safeguard patient
commencement of the examination and final result commu- privacy, the dataset creators stripped all images of metadata,
nication to the patient, and the ongoing decline in public thereby preventing unintended patient identification. Nota-
healthcare funding and a global shortage of qualified derma- bly, the Medical University of Vienna and the University of
tologists [320]. The primary objective of implementing Lean Queensland’s ethics review committees approved the data
principles in healthcare is to enhance value by minimizing collection and publication processes. Most images in this
the daily challenges encountered by patients and healthcare collection prominently feature a malignant or benign tumor
staff [321, 322]. Thus, instead of just concentrating on cost- situated at the image’s core. These tumors, or neoplasms, are
cutting, Lean Healthcare places a premium on maintaining unusual tissue growths and a characteristic attribute of many
superior quality, safety, satisfaction, and staff morale [323, types of cancer, instigated by excessive cellular division
324]. It revolves around establishing processes that enhance or a failure of cells to die at the appropriate time. Benign
the patient’s experience and eradicating or refining those tumors, typically slow-growing, retain distinct boundaries,
that do not add value, such as long wait times for services, remaining confined to their initial location without invading
healthcare providers, or test results [325, 326]. neighboring organs, earning them the “noninvasive” label
[328]. Ordinarily, benign tumors do not cause complications
5.4.1 Utilized dataset and, once surgically removed, seldom reappear. Yet, some
benign tumors can evolve into malignant tumors, clusters of
The HAM10000 dataset, boasting 10,000 images, is a valu- cells displaying uncontrollable and invasive growth. Capa-
able resource compiled over two decades by the Department ble of spreading to other organs (such as the liver, lungs,
of Dermatology at the Medical University of Vienna, Aus- brain, or bones) via the bloodstream or lymphatic system,
tria, and Cliff Rosendahl’s skin cancer clinic in Queensland, early detection of these malignant tumors allows for surgical
Australia [327]. This treasure trove of high-quality images removal and potential subsequent treatments like chemo-
includes resolutions of 128 × 128 and 256 × 256, with some therapy, radiotherapy, or immunotherapy [329, 330]. This
of the collected visuals hailing from the pre-digital camera dataset incorporates images featuring seven categories of
era. The dataset consists of images extracted from both Pow- pigmented skin lesions. Diagnostic procedures will likely
erPoint presentations and Excel databases and those digi- classify over 95% of detected skin anomalies into one of
tized from dia-positives. All images underwent a meticulous these seven categories [331]. Clinicians face the task of
quality check to ensure no loss of information during the distinguishing between these lesions and diagnosing them
extraction process. Notably, the image count in this dataset accurately within standard categories. Except for vascular
does not equal the number of unique lesions as the collection lesions, the skin lesions depicted in this dataset are melanin-
includes varied images of the same lesion captured at differ- pigmented. Melanin, a natural skin pigment, is produced
ent angles, magnification levels, or using other camera sys- by specific skin cells known as melanocytes. Genetic varia-
tems. Images that were insufficiently magnified, out of focus, tions among individuals dictate the levels of different types
or otherwise substandard were culled from the collection. of melanin in their bodies, influencing their skin, hair, and

Fig. 35  Value of Lean Healthcare and AI

5386 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

eye colors. Conversely, vascular lesions are pigmented by Augmentation techniques involve alterations to spatial
hemoglobin, a protein found in red blood cells that facilitates properties, including horizontal and vertical flipping,
oxygen and carbon dioxide transportation to and from body rotating, adjusting image brightness, shifting images hori-
tissues. Notably, this dataset does not include non-pigmented zontally and vertically, and varying the magnification of
lesions, which offer greater diversity and diagnostic catego- existing images. CNN models, with their hidden neurons,
rization options [332]. Finally, the classes in the dataset are often depend on the size and diversity of the training data-
nv (melanocytic nevi), mel (melanoma), bcc (basal cell car- set for superior performance on complex tasks [336]. Data
cinoma), df (dermatofibroma), akiec (actinic keratoses and augmentation can aid in replicating real-world scenarios
intraepithelial carcinoma), bkl (benign keratosis), and vasc since dermoscopic images are taken at various angles and
(vascular skin lesions). settings, sometimes inverted, using different camera speci-
fications [337]. Finally, an 80/20 split was used for the
5.4.2 Methodology model’s training, validation, and testing, respectively. Fig-
ure 34 shows the architecture of the custom-built computer
The HAM10000 dataset exhibits an imbalance that poses vision CNN algorithm for skin disease detection.
challenges in training CNN models for intricate tasks
[333]. This imbalance is particularly evident in medi- 5.4.3 Results and discussion
cal applications where specific incidents are infrequent,
leading to a skewed dataset. This imbalance can result in Table 16 shows the results obtained for the performance
biased predictions and impact model performance [334]. measurement values for detecting each of the seven classes
Computer vision algorithms often provide enhanced detec- in this dataset. This custom-built CNN algorithm did an
tion performance with larger image quantities. However, excellent job in detecting bcc, akiec, vasc, and df skin
acquiring many images, particularly medical images, can diseases with F1 score values nearing 100%. While at the
be challenging due to privacy regulations associated with same time, it did a slightly lower performance detecting
medical practices worldwide. Consequently, image insuf- nv, mel, and bkl skin diseases.
ficiency is a frequent problem and can cause overfitting The Lean methodology is about reducing waste and
during model training. Data augmentation techniques are improving efficiency, and this aligns well with the benefits
employed to tackle these challenges, increasing the sample that AI and CNNs bring to healthcare. Early cancer detection
size and equalizing the number of images per class. Fig- can significantly reduce the treatment costs and potential
ure 33 shows the breakdown of each of the seven classes complications that arise from late-stage diagnoses. This
before and after augmentation. As such, data augmenta- minimizes the waste of resources, including time and money.
tion addresses data insufficiency and overfitting [335]. CNNs can analyze and learn from vast amounts of data,

Fig. 36  Forecasting process flow

The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5387

Fig. 37  Data role in forecasting

sometimes surpassing the accuracy of human clinicians. This continuous improvement. Figure 35 shows the value of Lean
ensures that the treatment plan developed is based on the most Healthcare and AI.
accurate diagnosis possible, maximizing value for the patient,
which is a crucial aspect of Lean thinking. Furthermore, 5.5 Multinominal time series forecasting
the implementation of CNNs also aids in standardizing
the diagnostic process. Standardization, a fundamental Predictive analysis, or forecasting, is a process that
Lean principle, leads to predictable results and reduces involves predicting future outcomes based on historical
the likelihood of errors. CNNs follow the same process data analysis and insights gleaned from knowledge and
in analyzing images, ensuring a consistent approach. By experience. In today’s rapidly evolving world, forecasting
automating the diagnostic process, CNNs can free up doctors’ has become an invaluable tool in preempting blind guesses.
time to focus on more complex patient care tasks. Patients It plays a significant role across various sectors such as
can have quicker access to diagnostics, reducing anxiety and marketing, sales, industrial operations, economics, and
waiting times. Finally, using CNNs for cancer detection can finance, enabling the formulation of strategic plans
streamline healthcare systems by reducing waste, improving for future events [338]. However, these forecasts must
quality, standardizing the diagnostic process, and enabling accurately reflect reality, as misguided outcomes could

Fig. 38  Architecture of the CNN model used in forecasting

5388 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 39  Actual vs. predicted

closing prices for n = 10, 30,
and 60 days

Table 17  MSE, RMSE, and MAE values for n = 10, 30, and 60 days
severely hamper decision-makers who largely rely on these
predictions [339]. Consequently, managers, operators,
N days MSE RMSE MAE policymakers, and statisticians express immense interest
10 8.732875e + 06 2955.143782 2253.266995 in forecasting. Despite its ancient roots tracing back to the
30 8.544010e + 06 2923.013793 2405.412035 dawn of civilization when people endeavored to predict
60 1.371060e + 07 3702.783044 2871.178257 weather patterns, forecasting only emerged as a recognized
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5389

field in the 1940s. Its importance gained traction alongside performance. On the contrary, businesses that choose to be
global developments. Rapid expansion in forecasting was listed on the NYSE are often seen as more secure and well-
observed during the 1960s and 1970s, especially with established. It is worth noting that the Dow includes equi-
computers aiding in computational tasks [340]. With ties from the Nasdaq and NYSE. In contrast to the Nasdaq
the advent of AI and other advanced technologies, the Composite and the Dow, the S&P 500 is generally viewed
application of forecasting has extended to various areas, as a more comprehensive reflection of market performance
including business, economics, management, politics, across all sectors. Relative to the S&P 500, the Russell 2000
medicine, and more. Forecasting is a multi-faceted process has a higher correlation to the economic performance of the
in the manufacturing industry that combines historical USA and is commonly employed to gauge the performance
knowledge, real-time data analysis, and proficiency to of small and medium-sized enterprises (SMEs). None of the
prepare for future decision-making or problem-solving. It five mentioned indices—Nasdaq, NYSE, Dow Jones Indus-
assists enterprises in managing their inventory levels and trial Average, S&P 500, and Russell 2000 —is specifically
optimizing production while keeping costs to a minimum dedicated to the manufacturing sector. They all comprise a
[341]. Forecasting in this sector can leverage a wide range broad range of industries and companies. However, for the
of data, including inventory, production, labor, facilities, analysis presented in this paper, the Dow Jones Industrial
demand, equipment, material, cash, and energy [342]. In the Average was selected since a good number of its constituent
current market scenario, manufacturing firms are operating companies, such as 3 M, Caterpillar, and Boeing, are major
amidst escalating uncertainties and economic volatility. players in the manufacturing sector. Figure 37 shows the role
Projecting future sales, anticipating volume fluctuations, of data in forecasting.
understanding product life cycles, and accommodating
product variety are essential in adjusting manufacturing 5.5.2 Methodology
and production structures to meet customer demands [343].
Moreover, with intensifying competition among various The analyzed dataset is stochastic, containing 1985 recorded
manufacturing industries and technological advancements, instances spread across 84 features, with the target output
manufacturing firms must embrace strict pricing policies being the closing price of the stock index. A sequence of
and innovate their machinery to remain competitive [344]. preprocessing steps was carried out to streamline the data.
Integrating global markets through the Internet has also Firstly, any missing values were addressed by filling them
necessitated increased production transparency. Emerging with data from either the preceding or subsequent row
technologies have underscored the need to enhance product according to their immediate temporal context. Despite this,
quality, diversify offerings, and optimize product life any rows persistently containing missing values were entirely
cycles [345]. Therefore, forecasting provides a pathway to removed. The crux of the data processing was the creation of
aid in meeting these needs and beyond through dynamic “n” lag features for all 82 input variables that spread along
planning, intelligent factory analysis, and reduced waste the time series feature. Here, “n” represents the number of
in manufacturing and inventory. All of which are critical days that the model will forecast into the future. In time series
for an efficient and effective production model in the analysis, a lagged feature refers to a feature whose value at
manufacturing industry. Figure 36 shows the process flow a particular timestamp corresponds to the value of the origi-
of forecasting in enterprise operations. nal feature at an earlier timestamp. Following the lag feature
creation, the data normalized to fit within the range [0, 1].
5.5.1 Utilized dataset This step was critical in ensuring numerical stability dur-
ing the subsequent neural network training phase. The entire
Due to the lack of access to a related manufacturing data- dataset was divided into a training set (comprising 80% of
set, an alternative dataset was utilized. This dataset con- the data) and a test set (consisting of the remaining 20%).
tains several daily features of five stock markets: S&P 500, The model then underwent training for 100 epochs, adopting
Nasdaq, Dow Jones Industrial Average, Russell 2000, and the “One Cycle Policy” learning rate schedule. This learn-
NYSE Composite from 2010 to 2017 [346]. The Nasdaq has ing rate modulation strategy is known for yielding improved
a reputation for being a hub for technological and innovative training outcomes. Following the training, the model was
enterprises, including digital and biotech firms that stand at harnessed to generate predictions on the training and test sets.
the forefront of their respective industries. Consequently, Subsequently, these predictions were re-scaled to the original
equities listed on the Nasdaq are typically associated with price scale using the stored minimum and maximum values
growth potential, though they may exhibit more volatility from the normalization step. The resulting model’s intricate
and are frequently used as indicators of technology sector architecture is depicted in Fig. 38.
5390 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 40  Forecasting in Lean Manufacturing

5.5.3 Results and discussion Therefore, the one with the smallest MAE is typically pre-
ferred among several models. However, one limitation of
The final piece of the analysis was the graphical representa- MAE is its inability to convey the relative scale of the
tion of the actual and predicted closing prices for both the errors, making it challenging to differentiate between large
training and test datasets. This visualization provided a clear, and small discrepancies. MSE can be utilized to address
comparative insight into the performance of the model (see this issue. MSE is calculated as the mean of the squared
Fig. 39). Several error metrics were computed to evaluate differences between the predicted and actual values. It
the model’s performance, including the mean squared error serves as a measure of the quality of a prediction model.
(MSE), mean absolute error (MAE), and root mean squared MSE considers variance (the deviation among predicted
error (RMSE). The values of these metrics are shown in values) and bias (the gap between the forecast and the
Table 17. actual value). A lower MSE, close to zero, ideally signifies
MAE gives an estimate of the average magnitude of a better predictive model. RSME is the square root of MSE
errors in the predictions. A zero MAE implies that the and similar to MSE in terms of interpretation; the only
predictions are perfectly accurate. A model with a smaller difference is that the RMSE number is in the same unit
MAE is generally considered superior, with a value of zero as the forecasted value. The increase in MSE, RMSE, and
signifying that the projections are entirely devoid of errors. MAE with increasing lag values suggests that the model’s
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5391

Table 18  Areas where Lean principles can be incorporated into the inspection process
Area Process

Standardized procedures Develop and implement standardized inspection procedures and checklists to ensure consistency and
thoroughness across all inspections. Standardization helps minimize errors, reduce rework, and
ensure that all critical aspects of the structure are evaluated
Efficient resource allocation Optimize utilizing resources, such as personnel, equipment, and time, to perform inspections. This can
involve scheduling inspections to minimize downtime and travel and ensuring the appropriate tools
and equipment are available for each inspection task
Preventive maintenance Prioritize preventive maintenance activities based on inspection findings to address potential issues
before they become critical. By identifying and addressing problems early, repair costs can be mini-
mized, and the overall efficiency of maintenance operations can be improved
Training and skill development Provide training and skill development opportunities for inspection personnel to ensure they are well-
versed in the latest inspection techniques and technologies. This can lead to more accurate and effi-
cient inspections, reducing the need for rework and minimizing the risk of overlooking critical issues
Continuous improvement Regularly review and refine inspection procedures, incorporating feedback from inspection teams and
lessons learned from past inspections. This commitment to continuous improvement helps ensure the
inspection process remains adequate and efficient
Performance measurement and feedback Establish performance metrics for the inspection process and use these metrics to evaluate the effec-
tiveness and efficiency of inspections. Provide feedback to inspection teams based on these perfor-
mance measurements, encouraging improvements and adjustments as needed
Collaboration and communication Encourage open communication and collaboration between inspection teams, engineers, and mainte-
nance personnel. This fosters a shared understanding of the bridge’s condition and the importance of
timely maintenance and repairs, helping to ensure that inspection findings are acted upon promptly
and efficiently

Fig. 41  Negative (normal) and

positive crack (faults) images
[347, 348]

performance is declining as the number of lagged values what is needed when needed, reducing overproduction, one
(or the time horizon of the forecast) increases. Reasons of the most critical types of waste in Lean Manufacturing.
could be due to the small size of the dataset, as DL models Overproduction results in excess inventory that incurs
require more data for training, and another variable is the storage costs and can become obsolete, tying up capital
stochastic nature of the dataset itself. better utilized elsewhere. Accurate forecasting minimizes
With continuous forecasting and subsequent adjustments underproduction, preventing missed sales opportunities
based on real-time data, the manufacturing process keeps and ensuring customer satisfaction. Reliable forecasting
improving. Each forecast provides an opportunity to smoothens the production process, reducing bottlenecks
learn and refine future predictions. Thus, providing a and downtime and thus enhancing the overall flow of
more accurate outlook is an essential component of Lean operations. When production is aligned with actual
Manufacturing as it underpins many of its core principles. demand, resources such as manpower, machinery, and
Precise forecasting allows manufacturers to produce just materials are used more efficiently, creating a smoother,
5392 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 42  CNN model

more consistent flow of value to the customer. In a 5.6.1 Utilized dataset

Lean system, products should be produced based on
customer demand (pull) rather than projected sales The quality and quantity of the dataset significantly influ-
(push). Accurate forecasting helps manufacturers shift ence the performance of CNN. To ensure optimal results,
towards a pull system, which reduces inventory levels, the network requires an extensive collection of images. The
minimizes storage costs, and results in a more flexible and dataset titled Concrete Crack Images for Classification [347,
responsive manufacturing process. Accurate forecasting 348] contains concrete images having cracks. The dataset is
allows employees to plan their work better, reducing stress divided into negative (normal) and positive crack (faults)
caused by last-minute changes, overworking, or idling. It images for image classification. Each class has 20,000
provides employees the stability and predictability needed images with a total of 40,000 images with 227 × 227 pixels.
to perform their best work, respecting their time and well- The high-resolution images have variance in terms of surface
being. Figure 40 shows the role of forecasting in Lean finish and illumination conditions. An 80/20 split was used
Manufacturing. for training and testing, respectively. Figure 41 shows an
example of the images found in the dataset.
5.6 Concrete crack detection
5.6.2 Methodology
Concrete structure inspections are crucial for several rea-
sons, including ensuring public safety, maintaining struc- A custom-built CNN model tailored (see Fig. 42 for an
tural integrity, identifying repair and maintenance needs, approximate illustration) for classifying colored concrete
prolonging the structures lifespan, and facilitating efficient images into “images with crack” and “images without crack”
resource allocation. Inspections are essential for maintaining integrates several components. The model often commences
safe, functional, and reliable infrastructure. Regular inspec- with an input layer, designed to receive the concrete image
tions help identify and address potential issues before they data. Given the colored nature of the images, the input
become critical, ensuring the safety of all users, prolong- dimensions typically account for the height, width, and three
ing the structure’s lifespan, and enabling efficient resource color channels (red, green, and blue). Following this, several
allocation for maintenance and repairs. The Lean concept, convolutional layers are introduced. These layers employ
which focuses on reducing waste and improving efficiency, filters (or kernels) to slide over the input image, detecting
can be applied to inspecting concrete in several ways. By features by computing dot products and generating feature
incorporating Lean principles into the inspection of con- maps. The convolutional layers are frequently paired with
crete, authorities can optimize the inspection process, reduce activation functions like the Rectified Linear Unit (ReLU) to
waste, and improve overall efficiency. This can lead to better introduce non-linearity, enabling the model to learn intricate
structural management, more effective maintenance activi- patterns. Pooling or subsampling layers intersperse between
ties, and safer and more reliable infrastructure. Table 18 convolutional layers, predominantly to reduce dimensionality,
shows some areas where Lean principles can be applied to focusing on dominant features and enhancing computational
the bridge inspection process. efficiency. Max pooling is a favored technique, where the
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5393

Fig. 43  Loss and accuracy in both training and validation for CNN model

Table 19  Values of performance measurements which, in this case, typically consists of two neurons cor-
Model Accuracy Precision Recall F1 score
responding to the two classes: “images with crack” and
“images without crack.” A softmax activation function is
Custom CNN 99.0375% 99.0505% 99.0375% 99.0374% utilized here to output the probabilities for each class. The
class with the higher probability determines the final clas-
sification. Training the model necessitates a loss function,
maximum value from a group of importance in the feature like categorical cross-entropy for this two-class problem,
map is chosen, effectively condensing the data. and an optimizer like Adam or SGD (Stochastic Gradi-
Deep CNN models usually integrate several convolu- ent Descent) to adjust weights based on the loss gradient.
tional and pooling layers in sequence, with each succes- During the training phase, the model iteratively refines its
sive layer aiming to recognize more complex features. weights by comparing its predictions to the true labels,
Post these layers, the network integrates one or more aiming to minimize the loss and improve accuracy. The
fully connected (dense) layers, which interpret the recog- model’s effectiveness is gauged using a validation data-
nized features and make decisions based on them. Before set, ensuring it generalizes well to unseen data. Figure 43
reaching the final classification, dropout layers might be shows loss and accuracy per epoch in both training and
interspersed within the fully connected layers to prevent validation for the CNN model. One epoch is when an
overfitting by randomly deactivating a fraction of neurons entire dataset is passed forward and backward through the
during training. Concluding the model is the output layer, neural network only once.
5394 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

5.6.3 Results and discussion surroundings, the impact of lighting, and the minimal dis-
tinction between areas with and without cracks. The evolu-
Table 19 shows the performance measurements results tion of DL provides a solution by autonomously discerning
obtained when using the CNN model to detect cracks in the characteristics of areas with and without cracks in the
concrete surfaces. It can be seen that a slightly larger than captured images. This technological progression in DL facil-
99% performance was recorded, which could be attributed to itates quicker and more accurate identification of structural
various reasons, such as the characteristics of the dataset and damages compared to traditional human-led inspections or
performance saturation. The dataset’s features, such as size, earlier digital detection methodologies. With the unavoid-
complexity, and noise levels, can influence the performance able scenario of aging infrastructure across the nation, the
of various models. lack of federal funding to keep up with the demand to reha-
Identifying cracks in concrete structures can be intricate bilitate older people and maintain the new, and the increased
because of factors like varying light exposures, intricate struggle to hire and retain a quality workforce ever since

Fig. 44  Number of discussion threads for each brand [353]

Table 20  Description of categories found in the dataset

Field Description

tweet_id A distinct, anonymized identifier for each tweet, which is referenced by response_tweet_id and in_response_to_
tweet_id
author_id A separate, anonymized user ID is assigned, with @ symbols in the dataset replaced by the corresponding
anonymized user IDs
inbound Whether the tweet is “inbound” to a company doing customer support on Twitter. This feature is useful when re-
organizing data for training conversational models
created_at Date and time when the tweet was sent
text The content of the tweet, where sensitive information like phone numbers and email addresses are substituted with
placeholders such as __email__
response_tweet_id A list of tweet IDs that are replies to the current tweet, separated by commas
in_response_to_tweet_id The tweet ID to which the current tweet is replying, if applicable
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5395

Table 21  Sample of the UPS tweet_id author_id inbound text

tweets before preprocessing
119,329 UPSHelp FALSE @105,859 UPS has a service called My Choice that will allow
your packages to be delivered to UPS Store. ^ML [Link]
co/​BPZKn​dqww1[Link]
119,330 105,859 TRUE @105,860 I wish Amazon had an option of where I can just
get it shipped to the ups store so I can avoid a lot of the
struggle
119,331 105,860 TRUE They reschedule my shit for tomorrow [Link]

COVID-19 hit the country the spring of 2020, now more are crucial for proactive management and sustainability in
than ever, an AI-based concrete inspection is a viable construction projects.
strategy not just to reduce waste in the traditional bridge
inspection process but also to keep the public safe which 5.7 Sentiment analysis
is the ultimate goal. In the end, future research in concrete
fault detection using image classification algorithms can be Sentiment analysis is a computational method used to deter-
directed towards several areas to improve these techniques’ mine the emotional tone behind a series of words. This tech-
overall efficiency, accuracy, and applicability. Integrating nique is commonly applied to text to understand the senti-
Lean principles with concrete crack detection via CNNs ments, opinions, or attitudes expressed therein. It is often
offers significant advancements in Structural Health Man- used in business and social media monitoring as it allows
agement (SHM) by enhancing efficiency, accuracy, and companies to track public opinion, conduct market research,
overall value. Employing CNNs for crack detection not only and monitor brand and product sentiment in real-time [349].
speeds up the inspection process but also enhances reliabil- The process typically involves categorizing the opinions
ity, thereby reducing the time waste—a core Lean objective. expressed in a piece of text as positive (+ 1), negative (−1),
Furthermore, the accuracy provided by CNNs minimizes or neutral (0). Advanced sentiment analysis models might
human error, ensuring repairs are conducted precisely where also capture more complex emotions and intensities. These
needed, which aligns with Lean’s focus on reducing rework models utilize various methods, including ML models such
and avoiding unnecessary material use. Additionally, the as CNNs, statistical techniques, and NLP tools, to interpret
integration facilitates data-driven decision-making, allowing and classify the emotional states conveyed by textual data
for predictive maintenance and resource optimization, which [350].

Table 22  Utilized functions

Function Goal

pandas This library is fundamental in handling structured data. It provides data structures and functions for manipulating and analyzing
structured data, such as text-based datasets. In addition, it allows easy reading, writing, and manipulation of tabular data.
numpy While numpy is primarily used for numerical computations, it works well with pandas and other libraries to support an efficient
Python data manipulation interface. It is also used to convert data to numpy arrays, often the input type needed for machine
learning libraries.
matplotlib This is a plotting library used for data visualization, which is a critical aspect of preprocessing steps as it helps understand the
distribution and characteristics of the data.
json JSON data is quite common in web data, and the json library allows Python to read JSON files into a format that can be manipu-
lated as Python objects. This is useful in preprocessing, as many text datasets might come in JSON format.
regex Text data often needs cleaning and formatting. The re-library allows for advanced string manipulation using regular expressions.
This can be used to remove unwanted characters, extract specific information, and perform many other string preprocessing
tasks.
nltk The Natural Language Toolkit library is used for many natural language processing tasks. For example, it includes capabilities for
tokenizing text (breaking the text up into words or other meaningful components), identifying parts of speech, stemming and
lemmatization (reducing words to their root form), and much more.
emoji This library helps in handling and manipulating emojis in text data. For example, in text preprocessing, it is often helpful to con-
vert emojis to their meaning in words or remove them, which can be efficiently done using this library.
contractions This library handles contractions in English text, crucial in text preprocessing steps as it expands contracted words to their
complete form. For instance, converting “don’t” to “do not.” This can help maintain consistency in the text data and improve the
performance of downstream tasks like text classification or sentiment analysis.
5396 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 45  CNN model used in sentiment analysis

In management, understanding the sentiment among natural language’s richness in conveying human experiences,
employees can help in assessing overall job satisfaction, the datasets fueling innovation often do not reflect contem-
organizational climate, and the effectiveness of internal porary language usage. This Twitter dataset provides a sub-
communications. By analyzing feedback from employee stantial volume of everyday English conversations (primar-
surveys, emails, or online communication platforms, man- ily) between customers and support agents on the platform
agement can identify areas of concern, improve employee and offers three key benefits over other conversational text
engagement, and enhance workplace culture. Moreover, datasets. The first key is that it resembles a purpose-driven
sentiment analysis applied to customer feedback, social dataset where customers reach out to support with the intent
media interactions, and market trends allows managers to to resolve specific issues, resulting in a narrower range of
make informed decisions about brand positioning, cus- discussion topics compared to unrestricted conversational
tomer service strategies, and product development. This datasets like the Reddit Corpus. The second key benefit is
proactive approach to management helps in maintaining that it has authenticity in the broader demographic represen-
a positive public image and building strong relationships tation of the users than those in the Ubuntu Dialog Corpus
with both customers and employees [351]. In the manufac- and exhibits more recent and natural typed text usage than
turing sector, sentiment analysis offers substantial benefits the Cornell Movie Dialog Corpus. The final key is concise
by optimizing product design and quality control. Manu- due to Twitter’s character constraints that encourage support
facturers can use sentiment analysis to scrutinize reviews agents to provide more genuine, non-scripted responses and
and feedback on their products from various channels, prompt problem descriptions and solutions. Additionally, the
including online forums, social media, and customer sur- platform’s message length restrictions are ideal for recur-
veys. This analysis helps in identifying common issues rent neural networks. Figure 44 shows a breakdown of the
or desired features, guiding the development of products number of discussion threads for these brands.
that better meet consumer expectations. Furthermore, The dataset is formatted as a CSV file, with each row rep-
sentiment analysis can play a crucial role in monitoring resenting a tweet. Various columns provide additional infor-
the supply chain and vendor relationships by analyzing mation. All included conversations have a minimum of one
communications and feedback, thus identifying potential customer inquiry and one company response. The inbound
disruptions or dissatisfaction. This capability enables man- field can be used to determine which user IDs are associated
ufacturers to address issues proactively, maintain smooth with company accounts. Table 20 shows a detailed descrip-
operations, and ensure the stability of their supply chains tion of the fields found in the dataset. Table 21 shows a
[352]. sample of some of the initial formats of the UPS tweets that
were taken from the dataset.
5.7.1 Utilized dataset
5.7.2 Methodology
The Twitter customer support [353] dataset built with
PointScrape [354] is a comprehensive, contemporary col- Preprocessing of a text dataset is a crucial step in any sen-
lection of three million tweets and responses from more timent analysis task. It transforms raw text data into an
than 20 global brands to foster progress in natural language understandable format for sentiment analysis models and
understanding and conversational models while examining algorithms. Since text data can be unstructured, noisy, and
current customer support practices and outcomes. Despite ambiguous, preprocessing helps to clean and organize the
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5397

data so that algorithms can easily interpret it [355]. The extract_conversation function was used to pull out individ-
process starts with filtering the DataFrame only to include ual back-and-forth dialogs between different users or par-
inbound tweets directed to UPSHelp. This is done by check- ticipants. The function was applied to each user’s specific
ing if the text of each tweet begins with “@UPSHelp.” This tweets to extract all conversations. The extracted conversa-
is a part of the data preprocessing phase when working with tions (prompts) are then fed (input) into the CNN model (see
pandas DataFrame in Python. This step is crucial for focus- Fig. 45 for its architecture) for sentiment analysis.
ing on specific subsets of data based on certain conditions or The model starts with an input layer that takes in vec-
criteria. Filtering is essential because it allows you to focus torized representations of text data, possibly embeddings
on particular parts of your dataset that are relevant to the that transform words into a format suitable for processing
analysis or model you are working on. It helps remove unim- by neural networks. The initial convolutional layers in the
portant or irrelevant data, reduce the computational cost, model, while typically associated with image processing, are
and increase your results’ accuracy. The text data cleaning employed here to capture local dependencies and contextual
continues for further processing or analysis through a series cues within the text. Each convolutional layer uses multiple
of steps to transform the raw text into a more digestible form, filters to scan through the input text data, producing feature
making it easier for ML models to learn from. The clean- maps that highlight different aspects of the text’s sentiment.
ing process involved removing URLs, Twitter usernames, Following the convolutional layers, max pooling layers are
memorable characters, and stop words, converting all text utilized to reduce the dimensionality of the feature maps.
to lowercase, performing lemmatization and stemming, and This reduction helps in distilling the most salient features
handling emojis by replacing them with their textual descrip- that are essential for sentiment analysis, thus improving the
tion using the emoji library. The pd.to_datetime() function efficiency and effectiveness of the model. Pooling layers
in pandas converts the “created_at” field to datetime format. summarize the outputs of neighboring groups of neurons in
Then, the text under “author_id” field was grouped, and the the feature maps into a single neuron, which retains only the
tweets within each group were sorted based on their data. strongest signal from each region (like the most dominant
This is to ensure that the tweets are in chronological order sentiment cues). A concatenate layer follows, merging out-
for each user and to select the first tweet from each group puts from different stages of the convolution and pooling lay-
as the starting point of each conversation. Multiple libraries ers. This concatenated output combines both high-level and
were utilized in Python for various tasks related to preproc- detailed features, providing a comprehensive feature set that
essing text datasets. Table 22 shows a summary of Python enhances the model’s ability to interpret complex sentiment
functions that were utilized in this process and what was expressions in text. The network also includes dense layers
their function in the preprocessing steps. towards the end, which integrate all features extracted from
The expand_contractions function was used in text pre- the previous layers into a format suitable for making final
processing to convert contracted words into their complete predictions. These fully connected layers are pivotal for the
forms. A contraction is a shortened version of a word or classification task, synthesizing all the extracted information
multiple words where an apostrophe replaces the missing into a coherent prediction about the sentiment conveyed in
letter(s). Examples include “don’t” for “do not,” “it’s” for the text. Dropout layers interspersed among the dense layers
“it is,” or “I’m” for “I am.” Expanding contractions can be help prevent overfitting by randomly dropping units (neu-
important for several reasons, including standardization, rons) during the training phase, which forces the network to
disambiguation, and improving text understanding. Text develop more robust features that are not reliant on specific
data often comes from a variety of sources and in various aspects of the training data. Finally, the output layer con-
formats. By expanding contractions, you can ensure your cludes the model’s architecture. In sentiment analysis, this
text data is more standardized, which can help improve layer would typically comprise units corresponding to differ-
the performance of your subsequent analysis or machine ent sentiment classes, such as positive, negative, and neutral.
learning model. Some contractions can have more than one The network outputs a probability distribution across these
meaning depending on the context (for example, “it’s” could classes, indicating the likelihood of each sentiment being
mean “it is” or “it has”). Expanding these contractions can the correct interpretation of the input text. Thus, while the
help disambiguate their meaning. Some sentiment analysis architecture bears similarities to those used in image pro-
tasks or models may perform better with total words rather cessing, its components and structure are adeptly tailored
than contractions. For instance, specific sentiment analy- to handle the nuances of sentiment analysis in textual data,
sis tools might better understand that “is not” is a negation emphasizing the extraction and interpretation of sentiment-
while “isn’t” could be overlooked. The expand_contractions related features throughout the network. In addition, this add
function helps automate this process of turning contrac- another dimension to the versatility of CNNs as it has been
tions into full words, improving the quality and consistency employed in data analytics, image classification, forecasting,
of your text data for further processing or analysis. The and text analytics.
 5398

Table 23  Results of sentiment analysis

Input Customer part Agent part Customer Agent
sentiment senti-
ment

“Customer: severe weather dont say backed cant deliver severe weather dont say backed cant deliver today dont im sorry hear please dm tracking look mc −1 0
today dont treat us like stupid treat us like stupid
Agent: im sorry hear please dm tracking look mc”
“Customer: want package lazy ass delivery driver left want package lazy ass delivery driver left another need assistance please provide tracking number ms −1 0
another person apartment clue person apartment clue
Agent: need assistance please provide tracking number
ms”
“Customer: hey fun talking support told package trailer hey fun talking support told package trailer city idea time experiencing seasonal delays best way keep 0 1
city idea whe delivered arrived arent supposed know whe delivered arrived arent supposed know delivery updated package track package updates shows deliv-
delivery date isnt point whole company date isnt point whole company ery scan package delivered ms
Agent: time experiencing seasonal delays best way
keep updated package track package updates shows
delivery scan package delivered ms”
“Customer: hello id love know delayed package next “hello id love know delayed package next day air ser- “sure look please dm tracking number km 1 0
day air service finally delivered today help thanks vice finally delivered today help thanks im happy received package please let us know ques-
Agent: sure look please dm tracking number km delivered earlier today thanks much” tions concerns future team help mm”
Customer: delivered earlier today thanks much
Agent: im happy received package please let us know
questions concerns future team help mm”
“Customer: billing center unable log morning pay corp “billing center unable log morning pay corp account “dont think talked technical support assistance kg 0 −1
account havent called yet website terribly slow getting attached would suggest contacting technical support asap help
Agent: dont think talked technical support assistance kg error multiple machines across geographic regions” get resolved kg”
Customer: havent called yet website terribly slow
getting attached error multiple machines across geo-
graphic regions
Agent: would suggest contacting technical support asap
help get resolved kg”
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5399

5.7.3 Results and discussion

their package’s status. This approach not only offers practical assistance but also conveys

Agents frequently exhibited neutral sentiments, focusing on procedural responses without

customer’s frustration about the delivery issue and asks for tracking information to look
significant emotional engagement. In the first interaction, the agent empathizes with the
Negative sentiment from agents is less common but notable in the fifth interaction. Here,

acknowledges the service delays and advises the customer on how to stay updated with

interactions also demonstrate a neutral sentiment, where the agents stick to protocol by
direct assistance. This could be interpreted as a lack of empathy or immediate support,

requesting further details to assist the customers, thus focusing solely on resolving the
technical support, which might seem like shirking responsibility rather than providing

into the matter further, maintaining a professional demeanor. The second and fourth
Categorizing sentiments in texts during sentiment analysis

the agent’s response is perceived as dismissive, with a recommendation to contact

A distinctly positive sentiment is observed in the third interaction, where the agent
is a critical task that determines the accuracy and effective-
ness of the sentiment analysis model. The following three
categories were defined in this case:

a positive, helpful, and proactive attitude towards customer service.

– Positive: expresses favorable opinions or emotions
– Negative: reflects unfavorable opinions or emotions
– Neutral: indicates absence of strong emotions or opinions
contributing to a negative sentiment classification.

The determination of whether a text is positive, nega-

issues without additional emotional expressions.

tive, or neutral in sentiment analysis largely depends on the
method and technology used (CNN in this case). Table 23
shows a sample of the results obtained when conducting
sentiment analysis on the UPSHelp subset.
The sentiment analysis results from customer-agent inter-
actions reveal distinct emotional dynamics that can be cat-
egorized under two primary groups: customer sentiment and
Agent sentiment analysis

agent sentiment. Each of these groups encompasses various

sub-categories reflecting the range of emotional responses
recorded during these exchanges. Table 24 shows the final
understanding of the five interaction samples shown in
Table 23.
Sentiment analysis integrates with Lean Manufacturing
by enhancing its core objective: the systematic elimination
gratitude and relief following the resolution of a delivery delay. This interaction shows

of waste and continuous improvement in all facets of pro-

frustration during their interactions. For instance, in the first interaction, the customer

explanation provided was inadequate and dismissive. Similarly, in the second interac-
is noticeably upset due to a service delay attributed to severe weather, feeling that the

simply reports technical issues with accessing the billing center, describing the situa-
negligence and criticizes the overall delivery service, highlighting a clear dissatisfac-

tional distress or satisfaction. In the third interaction, the customer sarcastically com-
tion, the sentiment further deteriorates as the customer accuses the delivery driver of

Neutral Neutral sentiments are expressed when customers discuss issues without explicit emo-

ments on the inefficiency of the service, showing resignation to the situation without
direct anger. The fifth interaction also falls under this category, where the customer
Positive A positive sentiment emerges in the fourth interaction, where the customer expresses

customer contentment and appreciation for the service’s efforts to rectify the issue,

duction. Lean Manufacturing emphasizes the philosophy of

Negative The negative sentiment category captures customers expressing dissatisfaction and

continuous improvement, or Kaizen, which involves every-

one in the organization from the top management to the shop
floor employees. Sentiment analysis can be used to gauge
the effectiveness and reception of continuous improvement
initiatives by analyzing employee feedback, comments, and
discussions. By understanding the sentiment of the work-
tion factually and without expressing frustration or anger.

force regarding changes and new practices, management can

adjust strategies to better align with employee expectations
and comfort, thus fostering a more receptive and motivated
workplace environment. Lean principles stress the impor-
marking a positive end to their experience.
tion with how their situation was handled.

tance of respect for people, and sentiment analysis can play

a crucial role in this area by providing insights into employee
morale and engagement. Regular analysis of communica-
tion channels, such as internal emails, feedback surveys, and
Table 24  Understanding sentiment analysis

meeting notes, allows managers to detect signs of dissatis-

Category Customer sentiment analysis

faction or resistance among employees early. This proac-

tive approach ensures that the workforce remains motivated
and engaged, which is crucial for sustaining the high levels
of operational efficiency required by Lean methodologies.
Lean Manufacturing focuses on value creation through the
eyes of the customer. Sentiment analysis applied to customer
feedback, product reviews, and social media can reveal what
customers truly value in a product or service. These insights
can be directly used to enhance VSM processes by identify-
ing and focusing on aspects that matter most to customers,
5400 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 46  The four components of AI-aided manufacturing

thereby ensuring that all efforts are directed towards value- monitoring, condition tracking, and proactive maintenance
added activities. In Lean Manufacturing, identifying and strategies.
eliminating waste in processes is crucial. Sentiment analysis
can assist in pinpointing problematic areas or processes by 6.1 Opportunities
analyzing sentiments expressed in project debriefs, prob-
lem-solving session outputs, and quality circle discussions. ML and DL have evolved into the bedrock of data-centric
Understanding employee sentiments regarding processes methodologies for constructing AI-aided manufactur-
or tools can help identify bottlenecks or non-value-added ing systems. The primary facets in creating such systems,
activities that may not be readily apparent through tradi- which are revolutionized by newer ML and DL algorithms
tional quantitative methods. daily, are feature extraction and innovative decision-making
models. Through ML and DL, AI-powered manufacturing,
specifically intelligent production, has gained unparalleled
6 Opportunities, challenges, risks, and cost instruments for thorough and efficient feature engineering.
The key opportunities AI can offer in intelligent automation
The cornerstone elements of any AI-aided manufacturing within manufacturing are manifold and impactful.
system encompass the manufacturing component, the data Adopting AI-guided techniques like heuristic optimiza-
acquisition component, the real-time monitoring component, tion models for ideal feature selection has found applications
and the problem-solving component (see Fig. 46). It is in intelligent automated manufacturing. Strategies rooted in
worth noting that the manufacturing component takes computational intelligence, such as the genetic algorithm
in raw materials and delivers the final product. On the and particle swarm optimization, can pinpoint the optimal
other hand, the data acquisition component supports subset of features that enhance the performance of the manu-
intelligent manufacturing strategies across the different facturing system. A pivotal innovation ML introduced to AI-
phases, powering the live monitoring and problem-solving aided manufacturing is comprehensive and automated fea-
components. Constant tracking plays a vital role in fostering ture engineering. While feature selection aims to streamline
improvements in quality. It is helpful in process health the quantity of the extracted features by choosing a subset
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5401

that bolsters the efficacy of decision-making models, feature generalization and overfitting; hence, data preprocessing or
engineering aspires to construct complex models from raw manual feature engineering may be needed before employ-
data for better extraction of significant patterns and insights. ing complex models for such datasets [360]. One potential
This is something that traditional data processing and fea- solution is to apply data fusion or augmentation methods to
ture extraction methods find challenging [356]. For instance, the small data collected in the early stages of experiments
through high-level data abstraction, DL generates signifi- to estimate the data range and generate synthetic data. How-
cant meaning from data. Feature engineering can incorporate ever, this approach may not always be suitable, depending
automatic feature selection. DL-based strategies like autoen- on the process’s nature and variability. Lack of access to big
coders or deep belief networks (DBN) can be employed for data may necessitate performing manual feature engineer-
high-level data abstraction while performing dimensional- ing and pre-selection before applying any DL models [361].
ity reduction and automated feature selection, which finds Collecting data in a manufacturing environment can
numerous applications in AI-aided manufacturing. It is also pose significant challenges. In many scenarios, the place-
proven that CNNs can perform sophisticated feature engi- ment of sensors is not straightforward, or the data collected
neering on image-based data and time series. Before the may be severely contaminated by machine vibration and
advancement of ML, the model’s performance in real-time ambient noise. Additionally, monitoring processes often
data monitoring heavily relied on the kind of signal process- require data acquisition from multiple sensors, which is
ing and the application of methods such as Fourier or wave- not straightforward due to potential desynchronization and
let transforms. While data preprocessing remains critical in integration difficulties when handling diverse data sources.
attaining an accurate monitoring model, ML and DL have This further complicates the use of extensive multi-source
placed increased importance on feature engineering in creat- data. Another significant challenge is data quality. Con-
ing a trustworthy monitoring model [357]. Feature selection trary to what many manufacturers believe, owning a large
should be an integrated part of intelligent decision-making quantity of data does not necessarily equate to having a
model development. Utilizing data-driven approaches for valuable dataset for analysis. Often, data quality outweighs
automatic feature engineering can help extract hidden pat- the amount, as noisy or irregularly sampled measure-
terns from the data and may eliminate the need for selecting ments have limited usefulness. Moreover, anomaly and
complex sensory systems or decision-making models. fault detection datasets are frequently imbalanced, with
The developments in ML and DL have facilitated the most observations representing normal machine/process
development of complex models for improved feature engi- conditions. This imbalance can create a bias towards the
neering and pattern identification. Additionally, these tech- dominant data group during classification. Methods such
niques enable the integrating of different models to gener- as up-sampling might not always be feasible, while down-
ate innovative hybrid models. For instance, SOMs can be sampling could lead to other issues related to data size.
employed for unsupervised learning and data grouping, One approach could be considering the initial cost and
thereby facilitating the recognition of machine states. Fur- modifying the loss function before executing the classi-
thermore, SOMs can be harnessed for feature processing fication task. Classification of imbalanced data can also
and data reduction, feeding another ML model for super- be reinterpreted as an anomaly detection problem using
vised learning. For example, the successful application of unsupervised learning methods and defining an anomaly/
fusing CNNs and recurrent neural networks (RNNs) allows damage threshold [362].
the newly generated model to comprehend the dataset’s tem- The choice of a suitable model hinges on multiple
poral and spatial features [358]. factors, which include the purpose of monitoring, dataset
size and dimensionality, the extent of feature engineering
6.2 Challenges carried out on the data, computation time, and hardware
constraints. For example, traditional supervised DL models
The volume of available data directly influences the selec- are utilized in regression-based prediction tasks. In some
tion of the appropriate model and its structure design. A cases, both supervised and unsupervised models could be
significant portion of research literature relies on small to employed depending on the availability of labeled data.
medium-sized datasets. The shift towards acquiring big data Methods like SVM and CNNs have been extensively used
necessitates a focused approach towards model selection and for classification tasks. Meanwhile, models like k-means,
precise hyperparameter tuning. It is a general guideline that fuzzy C-means, SOMs, or hierarchical clustering can be
the sample size should be at least ten times larger than the employed for pattern recognition when labeled data is
number of parameters in a DL model [359]. Aside from data unavailable [363]. DL models like autoencoders could be
size, high-dimensional data may require specific attention directly utilized for data reduction and feature engineering
for feature engineering. In particular, dealing with high- in a hybrid model combined with a supervised regression
dimensional small datasets can pose challenges regarding or classification model. RNNs have proven to be a powerful
5402 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

Fig. 47  Data-driven challenges in AI-aided manufacturing

tool for time series data, where sequence is critical. account for sensor malfunctions, and there is a need to
CNNs have also gained popularity for the classification distinguish between system faults and sensor failures
and regression of image-based data [364]. CNNs have [365]. For instance, contact sensors such as thermometers
shown promise for intelligent automated manufacturing or vibration-based transducers could be damaged by
in recent years. Yet, many investigations still focus on factors such as cooling fluid, lubricant, dust, machine
developing better intelligent systems using DL models. motor, or vibration. Therefore, it is vital to acknowledge
Comparative studies between different ML and DL models sensor failure and its impact on the monitoring model’s
in intelligent manufacturing are needed. However, one performance [366]. Despite the uncertainties associated
should keep in mind that the superiority of DL over ML with models derived from lab-scale investigations, they can
models depends on the size and quality of the data. DL serve as a baseline to develop and train models from real-
models are more challenging to tune, and there is a lack world collected data. Considering the exponential growth
of systematic approaches for designing the architecture and in data collection and the challenges tied to machine and
hyperparameter tuning of DL models in the literature. process health uncertainties, incremental learning should
One of the primary concerns in this category pertains to be implemented to expand existing intelligent models
laboratory-based research’s ability to evaluate the efficacy [367]. Transfer learning can effectively address the lack
of sensing technology or data-driven models to represent the of labeled prognosis data in industrial setups. Moreover,
unpredictable nature of processes in real-world scenarios. machine-to-machine interaction in manufacturing units
The model developed using single-source data may not can cause data variation, impacting the developed models’
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5403

integrity. Cloud computing can significantly bridge the gap 6.3 Risks

between advanced monitoring and prediction algorithms
from the research lab to the industry due to its enhanced The complete adoption of AI will likely yield far greater
computational efficiency and data storage capability. A fog benefits than attempting to stimulate progress without it.
computing layer can be incorporated into process monitoring Incorporating AI can enhance the chances of success as it
systems to address potential issues such as network delays, expedites the rate of experimentation, learning, and data-
reliability, and security attacks in cloud-based systems. The informed improvements across all the disciplines of the
introduction of a fog computing layer has been shown to manufacturing enterprise [66, 369]. One significant risk
reduce the response latency of monitoring systems [368]. in AI-aided manufacturing is the intense reliance on high-
Future cloud and fog computing research can further address quality data. A core hurdle for any organization in this sector
the uncertainty challenges related to developing online is to tackle data fragmentation by cultivating a data-centric
monitoring systems. The industrial IoT enables online culture. Ensuring that the correct data is fed into the problem
monitoring using multi-source data, which should be further is crucial. Solid data forms the foundation of AI modeling in
emphasized in AI-aided manufacturing systems. the manufacturing process. If a project lacks adequate data,
Mar r ying Lean met hodologies wit h AI can the outcome will be suboptimal. However, the absence of
streamline operations and facilitate a more dynamic flawless data initially should not deter manufacturing units
work environment. However, the success of this from initiating the process. It is crucial to have a system
integration hinges on employee involvement. Current ensuring that all data is fed into the AI mechanism promptly,
Lean management models rely on harnessing workforce enabling the most recent dataset analysis for the most accu-
expertise to drive enterprise value. When combining AI rate insights and solutions. Advances in datasets have driven
with Lean, staff input remains crucial for shaping new tech significant AI breakthroughs. Collecting, classifying, and
frameworks and job roles, leading to quicker and more labeling datasets for training algorithms is challenging, espe-
reliable results. AI should complement, not replace, human cially when dealing with comprehensive datasets that mirror
skills to maximize sustainable value and competitive edge. real-world scenarios. Regular audits of the manufacturing
The preliminary step ensures a shared understanding of processes are necessary to maintain the ongoing integrity
AI’s role across different teams. Figure 47 summarizes of the data being inputted, ensuring it is accurate and not
the data-driven challenges associated with AI-aided compromised. For instance, each time there is an upgrade in
manufacturing. the manufacturing process, it is vital to double-check that the
data being fed into the AI system is not disrupted. Product

Fig. 48  Waste reduction associated with AI-aided Lean Manufacturing

5404 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

upgrades always carry some risk; a thorough quality assur- the most up-to-date individuals to comprehend. Depend-
ance process is crucial to mitigate this risk. Otherwise, faulty ing on the neural network’s depth, the level of explanation
data might disrupt your algorithms. If data is sourced from and transparency can be quite limited. For instance, an AI
multiple points, it is vital to cross-verify the data from one system might monitor over a thousand operational data
source against another before applying any machine learn- signals to identify the most efficient production processes
ing. AI is bound to become a fundamental aspect of every from a pool of millions. Manual checking would be pro-
manufacturing strategy sooner or later, and solid data man- hibitively labor-intensive. This “black box” problem of AI is
agement processes will ensure its success. not new and has become even more pertinent with modern,
As AI systems can process tremendous volumes of potent DL solutions and sophisticated models. Meanwhile,
operational data from diverse sources in manufacturing, the these models can outperform humans in complex tasks like
issues of transparency and trust gain prominence. Some AI defect detection, automated quality control, or translation of
approaches function like a black box, offering little insight machine codes. However, the more complex the model, the
into the behind-the-scenes decision-making process. This less explainable it tends to be. This challenge needs to be
can raise concerns within and outside the organization, par- addressed in the context of AI-aided manufacturing.
ticularly when decisions on critical optimization factors like Even with data availability, manufacturing processes can
resource allocation, production schedules, quality control, encounter issues related to inherent biases in the datasets. If
and more are made solely by machines. Teams that employ biased data is fed to the AI algorithms, the system is expected
AI to manage substantial manufacturing operations need to to perpetuate those biases. For instance, in manufacturing,
demonstrate both performance and a degree of transparency datasets focused on certain types of materials, products, or
to build trust within the organization, ensuring responsi- machines can unintentionally narrow the AI’s understanding,
ble use of technology. Many AI systems in manufacturing creating a biased perspective. This could potentially hinder
employ neural networks—complex interlinked node sys- the exploration and optimization of other production lines or
tems—that may not easily reveal their rationale for spe- equipment different from the current focus. It is crucial to
cific decisions. Inputs and outputs may be visible, but the ensure that both the data and the algorithms analyzing it align
underlying decision-making process remains hidden. With with the organization’s principles, goals, and values. Current
their intricate and nuanced structures, these neural network- AI systems might struggle with the nuanced ethical and value-
based systems are often charged with coordinating multiple based judgments of their suggested outcomes. For example,
operations across different manufacturing platforms and there could be instances where systems fed with biased data
processes. They dynamically allocate resources, adjust pro- resulted in works that disproportionately disadvantaged spe-
duction schedules, provide insights, and act autonomously cific product categories or processes, unintentionally pro-
round-the-clock, making their decisions difficult for even moting an unequal distribution of resources or attention. If

Fig. 49  Role of generative AI

The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5405

Fig. 50  A damaged body panel

and a damaged T-shirt gener-
ated by AI

the input data is flawed or biased, the output will reflect the important to continually verify the output of these intelligent
same—the classic case of “garbage in, garbage out.” And machines, striving to understand how the algorithms and data
since a machine generates the outcome, there is a tendency to arrived at their results.
assume it is accurate. At times, this bias in AI could result in a
self-perpetuating cycle. If the AI system continually receives 6.4 Cost
skewed data (as that is what the algorithm has been trained
on), it will continue to generate similarly biased results. For Before incorporating AI technology into a Lean Manufac-
instance, in a manufacturing scenario, an AI system might turing framework, examining its compatibility with current
overly focus on optimizing one specific production line, systems and machinery is crucial. This involves assessing
thereby inadvertently overlooking the potential of others. In data formats, communication standards, and hardware
other words, biases can often be blind spots. Organizations configurations. Ensure that the new AI systems can effort-
may lack the expertise to adequately assess these data sets and lessly meld with existing infrastructure and implement any
filter out biased data or assumptions, potentially perpetuating required modifications for smooth integration. Therefore, a
these problems. This can further disadvantage certain aspects comprehensive financial assessment is crucial before inte-
of the manufacturing process. To prevent such scenarios, it is grating AI systems into Lean Manufacturing settings. This

Fig. 51  The AI impact on the manufacturing enterprise

5406 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

should account for initial acquisition and setup expenses improving AI model training and performance across manu-
and recurring maintenance and operations costs. Addition- facturing domains. Moreover, generative AI’s ability to auto-
ally, the evaluation must weigh these expenditures against mate product designs, predict and simulate cyber threats, and
potential advantages like increased operational efficiency, ensure seamless interoperability across platforms underscores
waste reduction, and enhanced product quality. Figure 48 its role in advancing global competitiveness and innovation
shows the waste reduction associated with coast when in manufacturing. Generative AI can come in different forms
implementing AI integrated with Lean Manufacturing. such as large language models (LLM) or large visual models
(LVM) or a combination of them.
Generative AI refers to a subset of artificial intelligence
7 Current and future trends in AI technologies that can generate new content, ranging from text,
images, and music to code and beyond, based on the patterns
At the dawn of the fourth industrial revolution, generative and data they have learned from. Generative AI produces out-
artificial intelligence (AI) emerged as a force capable of puts that mimic the training data in form and content. Unlike
transforming the manufacturing landscape. The integra- traditional AI tools that are primarily analytical, generative
tion of generative AI within manufacturing systems holds AI can create novel content. For instance, it can write original
immense potential to revolutionize design processes, optimize articles, compose music, or generate realistic images from
production, and enhance cybersecurity measures. By lever- textual descriptions. Generative AI models learn from vast
aging advanced algorithms, such as Generative Adversarial amounts of existing content. This training allows the models
Networks (GANs), Variational Autoencoders (VAEs), and to understand and replicate complex patterns and styles inher-
Transformers, generative AI facilitates the creation of real- ent in the data. The quality and diversity of the training data
istic manufacturing data and simulations. These capabilities significantly influence the outputs’ accuracy and authentic-
enable the generation of “golden datasets”—robust, diverse ity. Generative AI finds applications in diverse fields such as
data collections that mimic real-world conditions, thereby entertainment, where it can generate music and video game

Fig. 52  Optimizing Lean Manufacturing via AI integration to reduce waste

The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5407

environments; in the arts, for creating digital paintings and according to individual customer specifications without
drawings; in business, for generating marketing content and sacrificing speed or increasing costs significantly. AI can
reports; and in technology, for coding and software develop- generate and monitor complex quality control systems that
ment. One of the most famous examples of generative AI is go beyond human capabilities to detect defects or inconsist-
OpenAI ChatGPT platform that can generate images and texts encies (see Fig. 46). By continually learning from new data,
using prompts given to it by humans [370]. Figure 49 shows a these systems can adapt to changes in production materi-
summary of the role of generative AI in different areas. als or methods, ensuring high-quality outputs consistently,
Integrating digital technologies signifies a milestone which is crucial for maintaining Lean standards. This inte-
for achieving global competitiveness and leadership in the gration enables smarter, more responsive manufacturing
rapidly evolving manufacturing industry landscape. Within processes that can significantly enhance productivity and
this context, generative AI emerges as a transformative reduce waste.
force capable of redefining the manufacturing enterprise.
The ability of generative AI to create realistic manufacturing
data and simulations of various manufacturing scenarios, 8 Conclusions and final thoughts
including rare or unusual conditions, presents unparalleled
opportunities for manufacturing systems. Generative AI can The profound transformation witnessed in the manufactur-
address the scarcity of effective datasets for AI model train- ing sector can be mainly attributed to the advent of revo-
ing across manufacturing systems through its capability to lutionary concepts like the IoT, cloud-based technologies,
create high-quality, diverse datasets. It can generate data that mobile internet, and AI. AI has smoothed the pathway from
closely mimics real-world manufacturing conditions, includ- automated manufacturing to intelligent automated manufac-
ing data from machines, sensors, and devices, thus facilitat- turing. As industries grapple with an ever-expanding trove
ing better model training and sharing. This approach enables of data, the urgency to manage, store, and process this vast
the creation of “golden datasets” that are comprehensive information landscape underscores the criticality of data-
and robust, ensuring consistent AI model performance. centric manufacturing within the realm of intelligent produc-
For example, GANs and VAEs could be used to automate tion. Figure 51 shows a summary of AI impact across the
product designs that meet specific functional requirements manufacturing enterprise.
while minimizing material usage and production costs. This The advent of data-driven methodologies has transformed
approach could hold the potential to accelerate the design the manufacturing landscape by incorporating ML and DL
phase while also introducing a new level of creativity and techniques. These techniques have been instrumental in cre-
efficiency in product development. Generative AI can model ating intelligent systems that can effectively monitor process
unique manufacturing data features and effectively represent health and condition. In the context of AI-aided manufac-
different manufacturing layouts and modalities. Generative turing, these techniques have vastly improved automated
AI can produce realistic images and videos of manufacturing feature selection and data management for large and high-
processes under various conditions, including instances of dimensional datasets. They have made data fusion optimiza-
defects, anomalies, and inefficiencies (see Fig. 50). Training tion feasible and facilitated the creation of intelligent hybrid
AI models on this data can improve their ability to detect models capable of descriptive, diagnostic, predictive, and
waste and quality issues in real-time visually. prescriptive as well as classification and regression ana-
Integrating generative AI with Lean Manufacturing offers lytics—crucial aspects of Lean Manufacturing. However,
a unique opportunity to enhance the efficiency and innova- despite these significant advances, challenges persist, espe-
tion of manufacturing processes. Generative AI can accel- cially concerning the size and quality of data captured. There
erate the product design process by automatically generat- is a need for further exploration and comparison of DL con-
ing design alternatives based on specified constraints and cepts and their opportunities and limitations with traditional
requirements. This technology can use generative design ML models. Comparative studies among different basic and
algorithms to explore possible configurations, optimizing complex DHL models are essential. Small data sets pose
for materials used, weight reduction, durability, and manu- challenges that need to be addressed with data fusion tech-
facturing ease. This not only speeds up the design phase but niques. It is vital to bridge the gap between laboratory-scale
also ensures that the designs are optimized for Lean Manu- results and real-world conditions. Investigating process
facturing principles from the start. Lean Manufacturing aims uncertainty and implementing cloud and fog computing can
to meet customer demands efficiently without overproduc- provide a more accurate reflection of AI-aided Lean Manu-
tion. Generative AI facilitates mass customization, which facturing scenarios.
allows for the production of customized products with the Incorporating AI vision systems into Lean Manufacturing
efficiency of mass production. By using AI to adjust manu- environments enhances the detection of waste and inefficien-
facturing processes quickly, companies can produce goods cies through ongoing surveillance. The real-time nature of this
5408 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

monitoring allows production supervisors to intervene swiftly, References

minimizing further waste. Lowering scrap and rework rates is
pivotal in Lean Manufacturing due to their impact on produc- 1. Verma S, Sharma R, Deb S, Maitra D (2021) Artificial intelli-
tion expenses and throughput. These AI systems spot flaws gence in marketing: systematic review and future research direc-
tion. Int J Inf Manag Data Insights 1:100002. [Link]
early in the manufacturing sequence and automate the segre- 1016/j.​jjimei.​2020.​100002
gation of faulty goods from quality products. This automation 2. Davenport T, Kalakota R (2019) The potential for artificial intel-
serves to diminish scrap volumes and labor expenditures. AI is ligence in healthcare. Future Healthc J 6:94–98. [Link]
set to revolutionize the landscape of Lean Manufacturing. AI 10.​7861/​futur​ehosp.6-​2-​94
3. Javaid M, Haleem A, Singh RP, Suman R (2022) Artificial intel-
facilitates enhanced data flow among humans, systems, and ligence applications for Industry 4.0: a literature-based study.
devices, optimizing products and processes. As firms refine J Ind Integr Manag 07:83–111. [Link]
human-technology interactions through automation, AI stands 86222​13000​40
poised to elevate operational efficiency. Self-learning algo- 4. Iyer LS (2021) AI enabled applications towards intelligent
transportation. Transp Eng 5:100083. [Link]
rithms can significantly advance factors like product quality, treng.​2021.​100083
speed to market, and cost efficiency. 5. Ahmed S, Alshater MM, Ammari AE, Hammami H (2022)
Eliminating waste and improving efficiencies is as much a Artificial intelligence and machine learning in finance: a bib-
principle of Lean Manufacturing as AI. AI-aided Lean Man- liometric review. Res Int Bus Finance 61:101646. [Link]
org/​10.​1016/j.​r ibaf.​2022.​101646
ufacturing must focus on acquiring large quality datasets to 6. Moore S, Bulmer S, Elms J (2022) The social significance of
harness data-driven methods’ powerfully. Furthermore, the AI in retail on customer experience and shopping practices. J
pivotal role of feature engineering should be recognized, Retail Consum Serv 64:102755. [Link]
emphasizing an integrated approach that combines feature onser.​2021.​102755
7. Kopalle PK, Gangwar M, Kaplan A et al (2022) Examining
selection and expert decision-making. This will better reveal artificial intelligence (AI) technologies in marketing via a
the hidden patterns in the data, leading to more efficient and global lens: current trends and future research opportunities.
intelligent monitoring in an AI-aided Lean Manufacturing Int J Res Mark 39:522–540. [Link]
environment. 2021.​11.​002
8. Soni N, Sharma EK, Singh N, Kapoor A (2020) Artificial intel-
For AI systems to truly enhance Lean Manufacturing, ligence in business: from research and innovation to market
staff engagement is vital. Therefore, equipping employees deployment. Procedia Comput Sci 167:2200–2210. [Link]
with the skills to understand and utilize the data these sys- org/​10.​1016/j.​procs.​2020.​03.​272
tems produce is crucial. Highlighting these technologies’ 9. Takeuchi H, Yamamoto S (2019) AI service system develop-
ment using enterprise architecture modeling. Procedia Comput
advantages can facilitate employee engagement and ensure Sci 159:923–932. [Link]
effective use of the AI systems. Figure 52 shows a quick 10. Ransbotham S, Kiron D, Gerbert P, Reeves M (2017) Reshap-
overview of how Lean Manufacturing can be integrated into ing business with artificial intelligence. MIT Sloan Manag Rev
AI. 11. Singh R, Gill SS (2023) Edge AI: a survey. Internet Things
Cyber-Phys Systems 3:71–92. [Link]
2023.​02.​004
12. Schmitt M (2023) Automated machine learning: AI-driven
Author contribution Mohammad Shahin took care of conceptualiza- decision making in business analytics. Intell Syst Appl
tion, methodology, investigation, original draft, methodology, and final 18:200188. [Link]
revisions. Ali Hosseinzadeh took care of the dataset, methodology, 13. Batelle H (2021) The data & analytics flywheel: what is it and
and the review. Mazdak Maghanaki designed the graphical illustra- how do you create your own? In: AtScale. [Link]
tions, investigation, and helped with the initial LR. F. Frank Chen le.​com/​blog/​intro​ducing-​data-​analy​tics-​flywh​eel/. Accessed 14
contributed in resources. Finally, all authors read and approved the Jul 2023
final manuscript. 14. Gill SS, Xu M, Ottaviani C et al (2022) AI for next genera-
tion computing: emerging trends and future directions. Internet
Funding The reported research work received partial financial sup- Things 19:100514. [Link]
port from Office of Naval Research MEEP Program (Award Num- 15. Rahman MS, Bag S, Hossain MA et al (2023) The new wave
ber: N00014-19-1-2728) and Department of Energy/NNSA (Award of AI-powered luxury brands online shopping experience: the
Number: DE-NA0004003), as well as from the Lutcher Brown Dis- role of digital multisensory cues and customers’ engagement. J
tinguished Chair Professorship fund of the University of Texas at San Retail Consum Serv 72:103273. [Link]
Antonio. onser.​2023.​103273
16. Küfner T, Uhlemann TH-J, Ziegler B (2018) Lean data in man-
Data availability The data that support the findings of this study are ufacturing systems: using artificial intelligence for decentral-
available from the corresponding author, AH, upon reasonable request. ized data reduction and information extraction. Procedia CIRP
72:219–224. [Link]
Declarations 17. Czvetkó T, Kummer A, Ruppert T, Abonyi J (2022) Data-
driven business process management-based development of
Competing interests The authors declare no competing interests. industry 4.0 solutions. CIRP J Manuf Sci Technol 36:117–132.
[Link]
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5409

18. Patel L (2020) Lean AI: how can startups successfully scale Technovation 124:102763. [Link]
customer acquisition and revenue growth with a Lean team? on.​2023.​102763
O’Reilly Media, Inc. 35. Parteka A, Kordalska A (2023) Artificial intelligence and pro-
19. Twinprai N, Boonrod A, Boonrod A et al (2022) Artificial ductivity: global evidence from AI patent and bibliometric data.
intelligence (AI) vs. human in hip fracture detection. Heliyon Technovation 125:102764. [Link]
8:e11266. [Link] on.​2023.​102764
20. Cheng X, Zhang X, Cohen J, Mou J (2022) Human vs. AI: 36. Forghani R (2023) A practical guide for AI algorithm selection
understanding the impact of anthropomorphism on consumer for the radiology department. Semin Roentgenol 58:208–213.
response to chatbots from the perspective of trust and relation- [Link]
ship norms. Inf Process Manag 59:102940. [Link] 37. Agarwal R, Bjarnadottir M, Rhue L et al (2023) Addressing algo-
1016/j.​ipm.​2022.​102940 rithmic bias and the perpetuation of health inequities: an AI bias
21. Tekic Z, Füller J (2023) Managing innovation in the era of aware framework. Health Policy Technol 12:100702. [Link]
AI. Technol Soc 73:102254. [Link] org/​10.​1016/j.​hlpt.​2022.​100702
2023.​102254 38. Pyrkov A, Aliper A, Bezrukov D et al (2023) Quantum comput-
22. Lin Y-C, Chen C-T, Sang C-Y, Huang S-H (2022) Multiagent- ing for near-term applications in generative chemistry and drug
based deep reinforcement learning for risk-shifting portfolio discovery. Drug Discov Today 28:103675. [Link]
management. Appl Soft Comput 123:108894. [Link] 1016/j.​drudis.​2023.​103675
10.​1016/j.​asoc.​2022.​108894 39. Nouioua T, Belbachir AH (2023) The quantum computer for
23. Bagheri AB, Rouzi MD, Koohbanani NA et al (2023) Potential accelerating image processing and strengthening the security of
applications of artificial intelligence (AI) and machine learning information systems. Chin J Phys 81:104–124. [Link]
(ML) on diagnosis, treatment, outcome prediction to address 10.​1016/j.​cjph.​2022.​11.​006
health care disparities of chronic limb-threatening ischemia 40. US Bureau of Labor Statistics McKinsey Global Institute analy-
(CLTI). Semin Vasc Surg. [Link] oi.​o rg/​1 0.​1 053/j.​s emva​ sis. [Link] Accessed 26 Jul 2023
scsurg.​2023.​06.​003 41. Shahin M, Chen FF, Hosseinzadeh A et al (2023) Waste reduc-
24. Elbasheer M, Longo F, Nicoletti L et al (2022) Applications of tion via image classification algorithms: beyond the human eye
ML/AI for decision-intensive tasks in production planning and with an AI-based vision. Int J Prod Res 0:1–19. [Link]
control. Procedia Comput Sci 200:1903–1912. [Link] 10.​1080/​00207​543.​2023.​22256​52
10.​1016/j.​procs.​2022.​01.​391 42. Dillmann R (1988) Machine learning strategies for knowledge
25. Thomas T, Sharma P, Gupta DK (2023) Sustainable application acquisition in autonomous robot systems. IFAC Proc Vol 21:5–
of ML/AI methods in DCW understanding & screening. Mater 15. [Link]
Today Proc. [Link] 43. Holzinger A, Saranti A, Angerschmid A et al (2023) Toward
26. Petrick LM, Shomron N (2022) AI/ML-driven advances in human-level concept learning: pattern benchmarking for AI algo-
untargeted metabolomics and exposomics for biomedical appli- rithms. Patterns 100788. [Link]
cations. Cell Rep Phys Sci 3:100978. [Link] 100788.
xcrp.​2022.​100978 44. Robyns S, Helsen S, Weckx S et al (2023) An intelligent data
27. Schreckenberg F, Moroff NU (2021) Developing a maturity- capturing framework to improve condition monitoring and
based workflow for the implementation of ML-applications anomaly detection for industrial machines. Procedia Comput
using the example of a demand forecast. Procedia Manuf Sci 217:709–719. [Link]
54:31–38. [Link] 45. Zhang C, Zhou G, Li J et al (2023) A multi-access edge com-
28. Fahle S, Prinz C, Kuhlenkötter B (2020) Systematic review on puting enabled framework for the construction of a knowledge-
machine learning (ML) methods for manufacturing processes sharing intelligent machine tool swarm in industry 4.0. J Manuf
– identifying artificial intelligence (AI) methods for field appli- Syst 66:56–70. [Link]
cation. Procedia CIRP 93:413–418. [Link] 46. Rashidifar R, Chen FF, Bouzary H, Shahin M (2023) A math-
procir.​2020.​04.​109 ematical model for cloud-based scheduling using heavy traffic
29. Ajayan J, Nirmal D, Tayal S et al (2021) Nanosheet field effect limit theorem in queuing process. In: Kim K-Y, Monplaisir L,
transistors-a next generation device to keep Moore’s law alive: Rickli J (eds) Flexible automation and Intelligent Manufactur-
an intensive study. Microelectron J 114:105141. [Link] ing: the human-data-technology Nexus. Springer International
10.​1016/j.​mejo.​2021.​105141 Publishing, Cham, pp 197–206
30. Smith AR, Lugo-Fagundo E, Fishman EK et al (2022) More 47. Sundberg L, Holmström J (2023) Democratizing artificial intel-
from Moore’s law: the journey to toy story and implications ligence: how no-code AI can leverage machine learning opera-
for radiology. J Am Coll Radiol 19:592–593. [Link] tions. Bus Horiz. [Link]
10.​1016/j.​jacr.​2022.​01.​009 48. Zhang H, Song M, Wang Y (2023) Does AI-infused operations
31. Civit-Masot J, Luna-Perejón F, Corral JMR et al (2021) A capability enhance or impede the relationship between informa-
study on the use of Edge TPUs for eye fundus image segmen- tion technology capability and firm performance? Technol Fore-
tation. Eng Appl Artif Intell 104:104384. [Link] cast Soc Change 191:122517. [Link]
1016/j.​engap​pai.​2021.​104384 2023.​122517
32. Liu H, Tang W (2023) Quantum computing for power systems: 49. Rashidifar R, Chen FF, Shahin M et al (2024) Reinforcement
tutorial, review, challenges, and prospects. Electr Power Syst learning-based model for optimization of cloud manufacturing-
Res 223:109530. [Link] based multi objective resource scheduling: a review. In: Silva
33. Servou E, Behrendt F, Horst M (2023) Data, AI and governance FJG, Pereira AB, Campilho RDSG (eds) Flexible automation
in MaaS – leading to sustainable mobility? Transp Res Interdiscip and Intelligent Manufacturing: establishing bridges for more sus-
Perspect 19:100806. [Link] tainable Manufacturing systems. Springer Nature Switzerland,
34. Plantec Q, Deval M-A, Hooge S, Weil B (2023) Big data as an Cham, pp 133–140
exploration trigger or problem-solving patch: design and inte- 50. Ruschioni G, Malchiodi D, Zanaboni AM, Bonizzoni L (2023)
gration of AI-embedded systems in the automotive industry. Supervised learning algorithms as a tool for archaeology: clas-
sification of ceramic samples described by chemical element
5410 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

concentrations. J Archaeol Sci Rep 49:103995. [Link] 67. Ngo VM, Nguyen HH, Van Nguyen P (2023) Does reinforce-
10.​1016/j.​jasrep.​2023.​103995 ment learning outperform deep learning and traditional portfolio
51. Fosić I, Žagar D, Grgić K, Križanović V (2023) Anomaly detec- optimization models in frontier and developed financial markets?
tion in NetFlow network traffic using supervised machine learn- Res Int Bus Finance 65:101936. [Link]
ing algorithms. J Ind Inf Integr 33:100466. [Link] 2023.​101936
1016/j.​jii.​2023.​100466 68. Zhou C, Gao Y, Chen EJ et al (2023) Deep learning technologies
52. Zoppi T, Ceccarelli A, Puccetti T, Bondavalli A (2023) Which for shield tunneling: challenges and opportunities. Autom Constr
algorithm can detect unknown attacks? Comparison of super- 154:104982. [Link]
vised, unsupervised and meta-learning algorithms for intrusion 69. Zhou Y, Huang Z, Jiang Q et al (2023) Deep learning in preclini-
detection. Comput Secur 127:103107. [Link] cal antibody drug discovery and development. Methods. [Link]
cose.​2023.​103107 doi.​org/​10.​1016/j.​ymeth.​2023.​07.​003
53. Zhou Z, Liu Y, Hu T, Wang C (2023) Two unsupervised learning 70. Chang Y, Li Z, Saju G et al (2023) Deep learning-based rigid
algorithms for detecting abnormal inactivity within a household motion correction for magnetic resonance imaging: a sur-
based on smart meter data. Expert Syst Appl 230:120565. [Link] vey. Meta-Radiol 100001. [Link] oi.o​ rg/1​ 0.1​ 016/j.m
​ etrad.2​ 023.​
doi.​org/​10.​1016/j.​eswa.​2023.​120565 100001.
54. Wan TH, Tsang CW, Hui K, Chung E (2023) Anomaly detection 71. Santorsola M, Lescai F (2023) The promise of explainable deep
of train wheels utilizing short-time Fourier transform and unsu- learning for omics data analysis: adding new discovery tools to
pervised learning algorithms. Eng Appl Artif Intell 122:106037. AI. New Biotechnol 77:1–11. [Link] oi.o​ rg/1​ 0.1​ 016/j.n​ bt.2​ 023.​
[Link] 06.​002
55. Choi Y, Yoon G, Kim J (2022) Unsupervised learning algorithm 72. Yassir A, Jai Andaloussi S, Ouchetto O et al (2023) Acoustic fish
for signal validation in emergency situations at nuclear power species identification using deep learning and machine learning
plants. Nucl Eng Technol 54:1230–1244. [Link] algorithms: a systematic review. Fish Res 266:106790. [Link]
1016/j.​net.​2021.​10.​006 doi.​org/​10.​1016/j.​fishr​es.​2023.​106790
56. Zamri N, Pairan MA, Azman WNAW et al (2022) A comparison 73. Shahin M, Chen FF, Bouzary H et al (2022) A novel fully con-
of unsupervised and supervised machine learning algorithms to volutional neural network approach for detection and classifica-
predict water pollutions. Procedia Comput Sci 204:172–179. tion of attacks on industrial IoT devices in smart manufacturing
[Link] systems. Int J Adv Manuf Technol 1–13. [Link] oi.o​ rg/1​ 0.1​ 007/​
57. Quintero-Gull C, Aguilar J (2022) LAMDA-HSCC: a semi- s00170-​022-​10259-3
supervised learning algorithm based on the multivariate data 74. Mehrish A, Majumder N, Bharadwaj R et al (2023) A review
analysis. Expert Syst Appl 202:117479. [Link] oi.o​ rg/1​ 0.1​ 016/j.​ of deep learning techniques for speech processing. Inf Fusion
eswa.​2022.​117479 99:101869. [Link]
58. Liu M, Zhou M, Zhang T, Xiong N (2020) Semi-supervised 75. Shahin M, Chen FF, Hosseinzadeh A et al (2022) A deep hybrid
learning quantization algorithm with deep features for motor learning model for detection of cyber attacks in industrial IoT
imagery EEG recognition in smart healthcare application. Appl devices. Int J Adv Manuf Technol 123:1973–1983. [Link]
Soft Comput 89:106071. [Link] org/​10.​1007/​s00170-​022-​10329-6
106071 76. Shahin M, Hosseinzadeh A, Chen FF et al (2024) Deploying
59. Huang D, Jiang J, Zhou F, Ouyang S (2021) A distributed algo- optical character recognition to improve material handling and
rithm for graph semi-supervised learning. Pattern Recognit Lett processing. In: Silva FJG, Ferreira LP, Sá JC (eds) Flexible auto-
151:48–54. [Link] mation and Intelligent Manufacturing: establishing bridges for
60. Shahin M, Chen FF, Hosseinzadeh A (2023) Machine-based more sustainable Manufacturing systems. Springer Nature Swit-
identification system via optical character recognition. Flex Serv zerland, Cham, pp 510–517
Manuf J. [Link] 77. Shahin M, Chen FF, Hosseinzadeh A et al (2024) Deploying
61. Sen D, Aghazadeh A, Mousavi A et al (2019) Data-driven semi- computer-based vision to enhance safety in industrial environ-
supervised and supervised learning algorithms for health moni- ment. In: Silva FJG, Ferreira LP, Sá JC (eds) Flexible automation
toring of pipes. Mech Syst Signal Process 131:524–537. [Link] and Intelligent Manufacturing: establishing bridges for more sus-
doi.​org/​10.​1016/j.​ymssp.​2019.​06.​003 tainable Manufacturing systems. Springer Nature Switzerland,
62. Klar M, Glatt M, Aurich JC (2023) Performance comparison Cham, pp 503–509
of reinforcement learning and metaheuristics for factory layout 78. Bouzary H, Chen FF, Shahin M (2021) Using machine learn-
planning. CIRP J Manuf Sci Technol 45:10–25. [Link] ing for service candidate sets retrieval in service composition of
10.​1016/j.​cirpj.​2023.​05.​008 cloud-based manufacturing. Int J Adv Manuf Technol 115:941–
63. Shakya AK, Pillai G, Chakrabarty S (2023) Reinforcement learn- 948. [Link]
ing algorithms: a brief survey. Expert Syst Appl 231:120495. 79. Bouzary H, Chen FF, Shahin M (2021) Natural language process-
[Link] ing for comprehensive service composition in cloud manufac-
64. Dai Z, Xie P, Huang Y et al (2023) Optimization method of turing systems. Procedia Manuf 55:343–349. [Link]
power grid material warehousing and allocation based on multi- 1016/j.​promfg.​2021.​10.​048
level storage system and reinforcement learning. Comput Electr 80. Sharma R, Mithas S, Kankanhalli A (2014) Transforming deci-
Eng 109:108771. [Link] sion-making processes: a research agenda for understanding the
108771 impact of business analytics on organisations. Eur J Inf Syst
65. Liu P, Zhou J, Lv J (2023) Exploring the first-move balance point 23:433–441. [Link]
of Go-Moku based on reinforcement learning and Monte Carlo 81. Shahin M, Chen FF, Maghanaki M, Hosseinzadeh A, Zand N,
tree search. Knowl-Based Syst 261:110207. [Link] Khodadadi Koodiani H (2024) Improving the concrete crack
1016/j.​knosys.​2022.​110207 detection process via a hybrid visual transformer algorithm.
66. Khodadadi Koodiani H, Erfanian N, Majlesi A et al (2023) Cali- Sensors 24(10):3247
brating equations to predict the compressive strength of FRP- 82. Xie K, Wu Y, Xiao J, Hu Q (2016) Value co-creation between
Confined columns using optimized neural network model. Struc- firms and customers: the role of big data-based cooperative
tures 56:105060. [Link]
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5411

assets. Inf Manage 53:1034–1048. [Link] 99. Radecka K (2022) Improving business maintenance processes
2016.​06.​003 using Lean Manufacturing tools-case study. Zeszyty Naukowe.
83. Wu X, Zhu X, Wu G-Q, Ding W (2014) Data mining with big Organizacja i Zarządzanie/Politechnika Śląska, 156
data. IEEE Trans Knowl Data Eng 26:97–107. [Link] oi.o​ rg/1​ 0.​ 100. Argiyantari B, Simatupang TM, Basri MH (2022) Transportation
1109/​TKDE.​2013.​109 performance improvement through lean thinking implementa-
84. Bertsimas D, Kallus N (2020) From predictive to prescriptive tion. Int J Lean Six Sigma 13:622–647. [Link]
analytics. Manag Sci 66:1025–1044. [Link] IJLSS-​06-​2020-​0075
mnsc.​2018.​3253 101. Cekerevac Z, Dvorak Z, Prigoda L (2022) Lean manufacturing
85. Li H, Sheu PC-Y (2021) A scalable association rule learning vs Covid-19. Mest J 10:1–11. [Link]
heuristic for large datasets. J Big Data 8:86. [Link] 10.​01.​01
1186/​s40537-​021-​00473-3 102. Gupta S, Jain SK (2013) A literature review of lean manufac-
86. Naouali S, Ben Salem S, Chtourou Z (2020) Clustering categori- turing. Int J Manag Sci Eng Manag 8:241–249. [Link]
cal data: a survey. Int J Inf Technol Decis Mak 19:49–96. [Link] 10.​1080/​17509​653.​2013.​825074
doi.​org/​10.​1142/​S0219​62201​93000​64 103. Medonos M (2021) Leanness level of manufacturing compa-
87. Goldani MH, Safabakhsh R, Momtazi S (2021) Convolutional nies - a survey on lean manufacturing implementation. Acta
neural network with margin loss for fake news detection. Inf Acad Karviniensia 21:54–65. [Link] oi.​o rg/​1 0.​2 5142/​a ak.​
Process Manag 58:102418. [Link] 2021.​012
102418 104. Venugopal V, Saleeshya PG (2019) Manufacturing system sus-
88. Bottmer L, Croux C, Wilms I (2022) Sparse regression for large tainability through lean and agile initiatives. Int J Sustain Eng
data sets with outliers. Eur J Oper Res 297:782–794. [Link] 12:159–173. [Link]
org/​10.​1016/j.​ejor.​2021.​05.​049 105. Leksic I, Stefanic N, Veza I (2020) The impact of using differ-
89. Bouzary H, Chen FF, Shahin M (2019) Optimal composition of ent lean manufacturing tools on waste reduction. Adv Prod Eng
tasks in cloud manufacturing platform: a novel hybrid GWO-GA Manag 15:81–92. [Link]
approach. Procedia Manuf 34:961–968 106. Kumar N, Shahzeb Hasan S, Srivastava K et al (2022) Lean man-
90. (2023) A novel approach to voice of customer extraction using ufacturing techniques and its implementation: a review. Mater
GPT-3.5 Turbo: linking advanced NLP and Lean Six Sigma 4.0. Today Proc 64:1188–1192. [Link] oi.o​ rg/1​ 0.1​ 016/j.m
​ atpr.2​ 022.​
[Link] Accessed 28 Aug 2023 03.​481
91. Radecka K (2022) Improving Bu Siness maintenance processes 107. Vaněček D, Pech M, Rost M (2018) Innovation and lean produc-
using lean manufacturing tools – case study. Sci Pap Silesian tion. Acta Univ Agric Silvic Mendel Brun 66:595–603. [Link]
Univ Technol Organ Manag Zesz Nauk Politech Slaskiej Ser doi.​org/​10.​11118/​actau​n2018​66020​595
Organ Zarzadzanie 395–405. [Link] 108. Suhardi B, Anisa N, Laksono PW (2019) Minimizing waste using
3466.​2022.​156.​27 lean manufacturing and ECRS principle in Indonesian furniture
92. Nguyen D-M, Duong T-K, Enterprises (2022) Characteristics and industry. Cogent Eng 6:1567019. [Link]
lean outcome: an empirical evidence from Vietnam manufactur- 916.​2019.​15670​19
ing enterprises. Manag Syst Prod Eng 30(2):98–108. [Link] 109. Vinodh SKB, Devarapu S, Siddhamshetty S G (2017) Applica-
org/​10.​2478/​mspe-​2022-​0013 Edición 2 (June 2022) tion of lean approach for reducing weld defects in a valve com-
93. Ulewicz R, Czerwińska K, Siwiec D, Pacana A (2022) Analysis ponent: a case study. Int J Lean Six Sigma 8:181–209. [Link]
of the systemic approach to the concept of lean manufacturing doi.​org/​10.​1108/​IJLSS-​01-​2016-​0001
- results of empirical research: Analiza systemowego podejścia 110. Katayama H (2017) Legend and future horizon of lean concept
do koncepcji lean manufacturing - wyniki badań empiryc- and technology. Procedia Manuf 11:1093–1101. [Link]
znych. 25:375–395. [Link] 10.​1016/j.​promfg.​2017.​07.​227
94. García Alcaraz JL, Morales García AS, Díaz Reza JR et al (2022) 111. Alsaffar I, Ketan H (2018) Integration of lean six Sigma and
Machinery lean manufacturing tools for Improved sustainabil- Ergonomics: a proposed model combining Mura Waste and a
ity: the Mexican Maquiladora industry experience. Mathematics RULA Tool to Examine Assembly workstations. IOP Conf Ser
10:1468. [Link] Mater Sci Eng 433:012061. [Link]
95. Sancha C, Wiengartena F, Longonia A, Pagell M (2019) The 433/1/​012061
moderating role of temporary work on the performance of lean 112. Sakao T, Hara T, Fukushima R (2020) Using product/service-
manufacturing systems. Int J Prod Res 58:4285–4305. [Link] system family design for efficient customization with lean prin-
doi.​org/​10.​1080/​00207​543.​2019.​16514​58 ciples: model, method, and tool. Sustainability 12:5779. [Link]
96. Tripathi V, Chattopadhyaya S, Mukhopadhyay AK, Saraswat S, doi.​org/​10.​3390/​su121​45779
Sharma S, Li C, Rajkumar S (2022) Development of a data‐ 113. Radin Umar RZ, Tiong JY, Ahmad N, Dahalan J (2023) Develop-
driven decision‐making system using lean and smart manufac- ment of framework integrating ergonomics in lean’s muda, Muri,
turing concept in industry 4.0: a case study. Math Probl Eng and Mura concepts. Prod Plan Control 0:1–9. [Link]
2022(1):3012215 1080/​09537​287.​2023.​21896​40
97. Adreeva T, Shvidenko D, Popova L (2021) Development of a lean 114. Cirjaliu B, Draghici A (2016) Ergonomic issues in lean manufac-
quality management system: an integrated management system | turing. Procedia - Soc Behav Sci 221:105–110. [Link]
SpringerLink. IHNS 2010. Lecture notes in Business Information 10.​1016/j.​sbspro.​2016.​05.​095
Processing, vol 46. Springer, Berlin, Heidelberg. [Link] 115. Hopp WJ (2018) Positive lean: merging the science of efficiency
10.​1007/​978-3-​642-​12494-5_​13 with the psychology of work. Int J Prod Res 56:398–413. [Link]
98. Nascimento DLDM, Goncalvez Quelhas OL, Gusmão Caiado doi.​org/​10.​1080/​00207​543.​2017.​13873​01
RG, Tortorella GL, Garza-Reyes JA, Rocha-Lona L (2020) 116. Abdulmalek F, Rajgopal J, Needy K (2015) A classification
A lean six sigma framework for continuous and incremental scheme for the process industry to guide the implementation of
improvement in the oil and gas sector. Int J Lean Six Sigma lean. Eng Manag J 18:15–25. [Link]
11(3):577–595. [Link] 2006.​11431​690
117. Sangode PB (2018) Impact of 5s methodology on the efficiency
of the workplace. Study Manuf Firms
5412 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

118. Randhawa JS, Ahuja IS (2017) 5S–a quality improvement tool 135. Singh J, Singh H, Singh P (2020) The impact of 5S practices on
for sustainable performance: literature review and directions. Int the performance of manufacturing industry: an empirical inves-
J Qual Reliab Manag 34(3):334–361 tigation. Singh, Jagdeep and Singh, Harwinder and Singh, Par-
119. Latif MA, Vang J (2021) Top management commitment and lean deep, The Impact of 5S Practices on the Performance of Manu-
team members’ prosocial voice behaviour. Int J Lean Six Sigma facturing IndustryNovember 8, The IUP Journal of Operations
12(6):1289–1309. [Link] Management, Vol. 20, No. 2, May 2021, pp. 35–49, Available at
120. Gupta S, Jain SK (2015) an application of 5S concept to organ- SSRN: [Link] srn.c​ om/a​ bstra​ ct=3​ 95878​ 6. Accessed 1 Aug 2022
ize the workplace at a scientific instruments manufacturing 136. Lopes R, Freitas F, Sousa I (2015) Application of lean manufac-
company. Int J Lean Six Sigma 6:73–88. [Link] turing tools in the food and beverage industries. J Technol Manag
1108/​IJLSS-​08-​2013-​0047 Innov 10:120–130. [Link]
121. Khan SA, Kaviani MA, Galli BJ, Ishtiaq P (2019) Application 03000​13
of continuous improvement techniques to improve organization 137. Patel VC, Thakkar HA, Case (2014) Study: 5s implementation
performance: a case study. Int J Lean Six Sigma 10:542–565. in ceramics manufacturing company | PDF | Lean Manufactur-
[Link] ing | Employment. In: Scribd. Bonfring Int J Ind Eng Manag
122. Kawalec M, Pacana A, Rewilak J (2018) Development of the Sci 4(3), August Accessed 1 Aug 2022
5S method implementation on the grounds of collected expe- 138. Teplická K, Hurná S, Seňová A (2021) Design of workplace
rience. Zesz Nauk Organ Zarządzanie Politech Śląska z 119. layout using the 5s method in the area of quality management
[Link] system. Quality-Access to Success 22(183)
123. Swarnkar B, Verma D (2017) Implementation of ‘5S’in a small 139. Chi H (2011) 5S implementation in Wan Cheng industry manu-
scale industry: a case study. Int J Eng Res Appl 07:44–48. facturing factory in Taiwan . The Graduate School University
[Link] of Wisconsin-Stout May, Accessed 1 Aug 2022
124. loo Tan YL, Tan CL (2018) Reducing scrap at a multinational 140. Furman J, Małysa T (2021) The use of lean manufacturing
toy manufacturer in Malaysia - Tan – 2018 - Global Business (LM) tools in the field of production organization in the metal-
and Organizational Excellence - Wiley Online Library.,[Link] lurgical industry. Metalurgija 60(3–4):431–433
onlin​e libr​a ry.​w iley.​c om/​d oi/​f ull/​1 0.​1 002. Accessed 1 Aug 141. Fernandes JPR, Godina R, Matias JCO (2019) Evaluating the
2022 impact of 5S implementation on occupational safety in an auto-
125. Singh J, Vikas R, Sharma R (2014) Implementation of 5S prac- motive industrial unit | SpringerLink. Springer Proc Math Stat
tices: a review. Uncertain Supply Chain Manag 2:155–162. 281(April):139–148, Accessed 1 Aug 2022
[Link] 142. Agrahari R, Dangle PA, Chandratre KV Implementation Of 5S
126. Sati S, Adam A (2019) Evaluating the effectiveness of 5S imple- methodology in the small scale industry a case study | Seman-
mentation in the industrial sector tic Scholar. Int J Sci Technol Res 4(04), APRIL 2015,ISSN
127. de Barros LB, Bassi LDC, Caldas LP, Sarantopoulos A, Zeferino 2277–8616. Accessed 1 Aug 2022
EBB, Minatogawa V, Gasparino RC (2021) Lean healthcare tools 143. Zuliana N, Radityo BA, Tualeka AR (2019) The influence work
for processes evaluation: an integrative review. Int J Environ Res periods and knowledge of 5S to the 5S implementation at logis-
Public Health 18(14):7389. [Link] tic section in heavy equipment fabrication company GRESIK.
7389 Malays J Med Health Sci 15:19–22
128. Randhawa JS, Ahuja IS (2017) 5S implementation methodolo- 144. Mehta V, Dave P (2020) Impact of 5S and lean manufacturing
gies: literature review and directions. Int J Prod Qual Manag techniques in various organisations to enhance the productiv-
20:48. [Link] ity. 2:421–436. [Link]
129. Kgotso MA (2020) lean 7s methodology framework to improve 145. Gamboa P, Singgih ML (2021) Lean manufacturing improve-
efficiency of organisational performance at a packaging and ment using ECRS and TRIZ methods: literature review. Pro-
fulfilment SME. Proceedings of the International Conference ceedings of the Second Asia Pacific International Conference
on Industrial Engineering and Operations Management Dubai, on Industrial Engineering and Operations Management, Sura-
UAE, March 10–12 karta, Indonesia, September 14–16, 2021
130. Soliman M (2018) Healthcare is ripe for lean. Ind Manag 146. Romana FA (2021) Lean management implementation in small
60(1):2018. Available at SSRN [Link] and medium sized companies–a success case study in a manu-
93. Accessed 16 May 2023 facturing process. J Intercult Manag 13(1):88–121
131. Patel M, Kiran MB (2022) The review on various strategies 147. Chiarini A (2011) Integrating lean thinking into ISO 9001: a
adopted for implementing and sustaining 5S in a manufacturing first guideline. Int J Lean Six Sigma 2(2):96–117. [Link]
industries. Proceedings of the International Conference on Indus- org/​10.​1108/​20401​46111​11350​00
trial Engineering and Operations Management Istanbul, Turkey, 148. Ulrych W, The 5S method and its influence on employee work
March 7–10 requirement practices which can hamper lean service intro-
132. Wolniak R (2020) Analysis of the 5S method functioning duction | Journal of Positive Management. Journal of Positive
in a production company. Zeszyty Naukowe. Organizacja i Management [online]. 1 (2020) October T. 10, nr 2, s. 30–41.
Zarządzanie/Politechnika Śląska 146:523–531. [Link] [accessed 1.8.2022]. [Link]
10.​29119/​1641-​3466.​2020.​146.​37 149. Yogesh Y, Kulkarni S (2016) Study and analysis of the fac-
133. Abu F, Gholami H, Mat Saman MZ et al (2019) The implemen- tors affecting sustainability of 5S. Study Anal Factors Affect
tation of lean manufacturing in the furniture industry: a review Sustain 5S 4(6):178–181
and analysis on the motives, barriers, challenges, and the applica- 150. Dimitrescu A, Babiş C, Niculae E et al Impact on quality of
tions. J Clean Prod 234:660–680. [Link] production using 5S method. Journal of Research and Innova-
ro.​2019.​06.​279 tion for Sustainable Society (JRISS) 1, issue 1, 2019 ISSN:
134. Pheng LS Towards TQM – integrating Japanese 5-S principles 2668 – 0416. Thoth Publishing House 81 [Link]
with ISO 9001:2000 requirements | Emerald Insight. The TQM 33727/​JRISS.​2019.1.​11:​81-​86
Magazine 13(5):334–341. [Link] oi.o​ rg/1​ 0.1​ 108/E ​ UM000​ 0000​ 151. Patel VC, Thakkar H, March review on implementation of
005859. Accessed 1 Aug 2022 5S in various organization (2014). Int J Eng Res Appl ,ISSN:
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5413

2248–9622, Vol. 4, Issue 3(Version 1), pp.774–779. Accessed for environmental improvements. Procedia CIRP 81:1219–1224.
1 Aug 2022 [Link]
152. Omogbai O, Salonitis K (2017) The implementation of 5S lean 169. Knechtges P, Decker MC (2014) Application of Kaizen Method-
tool using system dynamics approach. Procedia CIRP 60:380– ology to Foster Departmental Engagement in Quality Improve-
385. [Link] ment. J Am Coll Radiol 11:1126–1130. [Link] oi.o​ rg/1​ 0.1​ 016/j.​
153. Subramanian M (2018) Kumar senthil implementing lean jacr.​2014.​08.​027
manufacturing principle in an automobile valve manufactur- 170. Tetteh HA (2012) Kaizen: a process improvement model for the
ing industry with simulation analysis – a case study. J Balkan Business of Health Care and perioperative nursing professionals.
Tribological Assoc 24(3):600–607 AORN J 95:104–108. [Link] oi.o​ rg/1​ 0.1​ 016/j.a​ orn.2​ 011.1​ 1.0​ 01
154. Shahin M, Chen FF, Hosseinzadeh A et al (2024) Robotics multi- 171. Helmold M (2020) Lean management and Artificial Intelligence
modal recognition system via computer-based vision. Int J Adv (AI). Manag Prof 131–137
Manuf Technol. [Link] 172. Batwara A, Sharma V, Makkar M, Giallanza A (2023) Towards
155. Shahin M, Chen FF, Hosseinzadeh A et al (2023) Enhanced smart sustainable development through value stream mapping – a
safety implementation in 5S + 1 via object detection algo- systematic literature review. Heliyon 9:e15852. [Link]
rithms. Int J Adv Manuf Technol 1–21. [Link] 10.​1016/j.​heliy​on.​2023.​e15852
s00170-​023-​10970-9 173. Nascimento DL, de Quelhas M, Moyano-Fuentes OLGJ et al
156. International Journal of Modern Engineering Research (IJMER) (2022) Circular value stream mapping 40: proposed general
(2022) Vol.3, no. 3, pp. 1654–1660, 2013 SafetyCulture, Avail- model and application to a digital 3D printing recycling factory.
able: [Link] Accessed on Feb 16, Accessed Sustain Prod Consum 34:600–612. [Link]
1 Aug 2022 2022.​10.​012
157. Sukdeo N (2017) The application of 6S methodology as a lean 174. Teriete T, Böhm M, Sai BK et al (2022) Event-based frame-
improvement tool in an ink manufacturing company. In: 2017 work for digitalization of value stream mapping. Procedia CIRP
IEEE International Conference on Industrial Engineering and 107:481–486. [Link]
Engineering Management (IEEM). pp 1666–1671 175. Ferreira W, de Armellini P, Santa-Eulalia F, de Thomasset-
158. Purohit SR, Shantha V (2015) Implementation of 5S methodol- Laperrière LA V (2022) Extending the lean value stream map-
ogy in a manufacturing industry. Int J Sci Eng Res 6(8):225–231 ping to the context of industry 4.0: an agent-based technology
159. Tang JR, Sethuprakhash V, Kamal MFM, Mei JH (2021) Imple- approach. J Manuf Syst 63:1–14. [Link]
mentation of 5S, practice in university electrical and electronic 2022.​02.​002
laboratories. Int Scientifica Electron J 55(1):171–185. [Link] oi.​ 176. Hernandez Marquina MV, Zwolinski P, Mangione F (2021)
org/​10.​32744/​ISSN Application of Value Stream Mapping tool to improve circular
160. Burroni L, Bianciardi C, Romagnolo C et al (2021) Lean systems. Clean Eng Technol 5:100270. [Link] oi.o​ rg/1​ 0.1​ 016/j.​
approach to improving performance and efficiency in a nuclear clet.​2021.​100270
medicine department | SpringerLink. Clin Transl Imaging 9:129– 177. Sultan FA, Routroy S, Thakur M (2021) A simulation-based per-
139. [Link] formance investigation of downstream operations in the Indian
161. Cabrera JL, Corpus OA, Maradiegue F, Álvarez JC (2020) Surimi Supply Chain using environmental value stream mapping.
Improving quality by implementing lean manufacturing, SPC, J Clean Prod 286:125389. [Link] oi.o​ rg/1​ 0.1​ 016/j.j​ clepr​ o.2​ 020.​
and HACCP in the food industry: a case study. South Afr J Ind 125389
Eng 31:194–207. [Link] 178. Renfro CP, Smith K, Wilson CA et al (2022) Use of value stream
162. Sommer AC, Blumenthal EZ (2019) Implementation of lean and mapping to improve the medication synchronization process. J
six Sigma principles in ophthalmology for improving quality of Am Pharm Assoc 62:740–749e3. [Link]
care and patient flow. Surv Ophthalmol 64:720–728. [Link] 2021.​12.​007
org/​10.​1016/j.​survo​phthal.​2019.​03.​007 179. Au-Yong CP, Azmi NF, Myeda NE (2022) Promoting employee
163. Sundararajan N, Terkar R (2022) Improving productivity in participation in operation and maintenance of green office build-
fastener manufacturing through the application of lean-kaizen ing by adopting the total productive maintenance (TPM) concept.
principles. Mater Today Proc 62:1169–1178. [Link] J Clean Prod 352:131608. [Link] oi.o​ rg/1​ 0.1​ 016/j.j​ clepr​ o.2​ 022.​
1016/j.​matpr.​2022.​04.​350 131608
164. Ochieng W, Ndungu F, Kamau L (2022) M119 impact of Kai- 180. Tortorella GL, Fogliatto FS, Cauchick-Miguel PA et al (2021)
zen implementation on TAT in the Nairobi hospital laboratory Integration of industry 4.0 technologies into total productive
(Kenya). Clin Chim Acta 530:S255. [Link] oi.o​ rg/1​ 0.1​ 016/j.c​ ca.​ maintenance practices. Int J Prod Econ 240:108224. [Link]
2022.​04.​217 org/​10.​1016/j.​ijpe.​2021.​108224
165. Washizu A, Nakano S (2022) Exploring the characteristics of 181. Pascal V, Toufik A, Manuel A et al (2019) Improvement indica-
smart agricultural development in Japan: analysis using a smart tors for total productive maintenance policy. Control Eng Pract
agricultural kaizen level technology map. Comput Electron Agric 82:86–96. [Link]
198:107001. [Link] 182. Eti MC, Ogaji SOT, Probert SD (2004) Implementing total pro-
166. Johanson T, Hunter G, Johnston L, Robb M (2022) 61.4 action, ductive maintenance in Nigerian manufacturing industries. Appl
not perfection: a COVID-19 Response Story about using a lean Energy 79:385–401. [Link]
Kaizen event to Develop Wraparound Care for Children and 007
Youth with ASD and Complex needs. J Am Acad Child Adolesc 183. McKone KE, Schroeder RG, Cua KO (2001) The impact of
Psychiatry 61:S83. [Link] total productive maintenance practices on manufacturing per-
167. Maarof MG, Mahmud F (2016) A review of contributing factors formance. J Oper Manag 19:39–58. [Link]
and challenges in implementing Kaizen in Small and Medium S0272-​6963(00)​00030-9
enterprises. Procedia Econ Finance 35:522–531. [Link] 184. Antônio Mendonça P, da Piedade Francisco R, de Souza Rabelo
10.​1016/​S2212-​5671(16)​00065-4 D (2022) OEE approach applied to additive manufacturing sys-
168. Bellgran M, Kurdve M, Hanna R (2019) Cost driven Green Kai- tems in distributed manufacturing networks. Comput Ind Eng
zen in pharmaceutical production – creating positive engagement 171:108359. [Link]
5414 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

185. Oechsner R, Pfeffer M, Pfitzner L et al (2002) From overall 202. Dhiravidamani P, Ramkumar AS, Ponnambalam SG, Subra-
equipment efficiency (OEE) to overall Fab effectiveness (OFE). manian N (2018) Implementation of lean manufacturing and
Mater Sci Semicond Process 5:333–339. [Link] oi.o​ rg/1​ 0.1​ 016/​ lean audit system in an auto parts manufacturing industry – an
S1369-​8001(03)​00011-8 industrial case study. Int J Comput Integr Manuf 31:579–594.
186. Basak S, Baumers M, Holweg M et al (2022) Reducing produc- [Link]
tion losses in additive manufacturing using overall equipment 203. Li X, Zhao S, Shen Y et al (2023) Integrated parameter optimi-
effectiveness. Addit Manuf 56:102904. [Link] oi.o​ rg/1​ 0.1​ 016/j.​ zation approach: just-in-time (JIT) operational control strategy
addma.​2022.​102904 for TBM tunnelling. Tunn Undergr Space Technol 135:105040.
187. Antosz K, Pasko L, Gola A (2020) The Use of Artificial Intelli- [Link]
gence methods to assess the effectiveness of Lean Maintenance 204. Li X, Tang L, Ling J et al (2023) Digital-twin-enabled JIT
Concept Implementation in Manufacturing Enterprises. Appl design of rock tunnel: methodology and application. Tunn
Sci 10:7922. [Link] Undergr Space Technol 140:105307. [Link]
188. Hosseinzadeh A, Frank Chen F, Shahin M, Bouzary H (2023) tust.​2023.​105307
A predictive maintenance approach in manufacturing systems 205. Chung W, Talluri S, Kovács G (2018) Investigating the effects
via AI-based early failure detection. Manuf Lett 35:1179–1186. of lead-time uncertainties and safety stocks on logistical perfor-
[Link] mance in a border-crossing JIT supply chain. Comput Ind Eng
189. Saurin TA, Ribeiro JLD, Vidor G (2012) A framework for 118:440–450. [Link]
assessing poka-yoke devices. J Manuf Syst 31:358–366. 206. Kesen SE, Baykoç ÖF (2007) Simulation of automated guided
[Link] vehicle (AGV) systems based on just-in-time (JIT) philosophy in
190. Al-Araidah O, Jaradat MAK, Batayneh W (2010) Using a fuzzy a job-shop environment. Simul Model Pract Theory 15:272–284.
poka-yoke based controller to restrain emissions in naturally [Link]
ventilated environments. Expert Syst Appl 37:4787–4795. 207. Bookbinder JH, Ülkü MA (2021) Freight Transport and Logistics
[Link] in JIT systems. In: Vickerman R (ed) International Encyclopedia
191. Chen X, Kurdve M, Johansson B, Despeisse M (2023) Ena- of Transportation. Elsevier, Oxford, pp 107–112
bling the twin transitions: Digital technologies support envi- 208. Kim SC, Shin KS (2019) Negotiation model for optimal Replen-
ronmental sustainability through lean principles. Sustain Prod ishment Planning considering defects under the VMI and JIT
Consum 38:13–27. [Link] Environment. Asian J Shipp Logist 35:147–153. [Link]
192. Antonelli D, Stadnicka D (2016) Classification and efficiency 10.​1016/j.​ajsl.​2019.​09.​003
estimation of mistake proofing solutions by fuzzy inference. 209. Abdul Nazar KP, Madhusudanan Pillai V (2018) Mixed-model
IFAC-Pap 49:1134–1139. [Link] sequencing problem under capacity and machine idle time con-
07.​651 straints in JIT production systems. Comput Ind Eng 118:226–
193. Aboumatar HJ, Winner L, Davis R et al (2010) Applying lean 236. [Link]
Sigma solutions to Mistake-Proof the Chemotherapy Prepara- 210. Werner S, Kellner M, Schenk E, Weigert G (2003) Just-in-
tion process. Jt Comm J Qual Patient Saf 36:79–AP4. [Link] sequence material supply—a simulation based solution in elec-
doi.​org/​10.​1016/​S1553-​7250(10)​36014-4 tronics production. Robot Comput-Integr Manuf 19:107–111.
194. Grout JR, Toussaint JS (2010) Mistake-proofing healthcare: [Link]
why stopping processes may be a good start. Bus Horiz 211. Wagner SM, Silveira-Camargos V (2011) Decision model for the
53:149–156. [Link] application of just-in-sequence. Int J Prod Res 49:5713–5736.
195. Hüttmeir A, de Treville S, van Ackere A et al (2009) Trading [Link]
off between heijunka and just-in-sequence. Int J Prod Econ 212. Kubasáková I, Kubáňová J (2020) The usage of Simulation just
118:501–507. [Link] in Sequence Road Transport for Company. Transp Res Procedia
196. Korytkowski P, Wisniewski T, Rymaszewski S (2013) Multi- 44:255–263. [Link]
variate simulation analysis of production leveling (heijunka) - a 213. Heinecke G, Lamparter S, Lepratti R, Kunz A (2013) Advanced
case study. IFAC Proc Vol 46:1554–1559. [Link] Supply Chain Information for Rule-based sequence adaptions on
3182/​20130​619-3-​RU-​3018.​00285 a mixed-Model Assembly line with unreliable Just-In-Sequence
197. Schumacher S, Hall R, Bildstein A, Bauernhansl T (2022) Lean deliveries. IFAC Proc Vol 46:1902–1907. [Link]
production systems 4.0: systematic literature review and field 3182/​20130​619-3-​RU-​3018.​00095
study on the digital transformation of lean methods and tools. 214. Motahari R, Alavifar Z, Zareh Andaryan A et al (2023) A multi-
Int J Prod Res 0:1–23. [Link] objective linear programming model for scheduling part families
21595​62 and designing a group layout in cellular manufacturing systems.
198. Tomašević I, Stojanović D, Slović D et al (2021) Lean in High- Comput Oper Res 151:106090. [Link]
Mix/Low-Volume industry: a systematic literature review. Prod 2022.​106090
Plan Control 32:1004–1019. [Link] 215. Saraçoğlu İ, Süer GA, Gannon P (2021) Minimizing makespan
2020.​17820​94 and flowtime in a parallel multi-stage cellular manufacturing
199. Slomp J, Bokhorst JAC, Germs R (2009) A lean production company. Robot Comput-Integr Manuf 72:102182. [Link]
control system for high-variety/low-volume environments: a org/​10.​1016/j.​rcim.​2021.​102182
case study implementation. Prod Plan Control 20:586–595. 216. Wu L, Zhao Y, Feng Y et al (2021) Minimizing makespan of sto-
[Link] chastic customer orders in cellular manufacturing systems with
200. Kolberg D, Knobloch J, Zühlke D (2017) Towards a lean auto- parallel machines. Comput Oper Res 125:105101. [Link] oi.o​ rg/​
mation interface for workstations. Int J Prod Res 55:2845– 10.​1016/j.​cor.​2020.​105101
2856. [Link] 217. Rafiee M, Kayvanfar V, Mohammadi A, Werner F (2022) A
201. Obeidat MS, Al-Aomar R, Pei ZJ (2014) Lean Manufacturing robust optimization approach for a cellular manufacturing sys-
implementation in the Sewing Industry. J Enterp Transform tem considering skill-leveled operators and multi-functional
4:151–171. [Link] machines. Appl Math Model 107:379–397. [Link]
1016/j.​apm.​2022.​02.​028
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5415

218. Weflen E, MacKenzie CA, Rivero IV (2022) An influence dia- Intell Ind AI4I Artif Intell Ind AI4I 2020 Third Int Conf AI4I
gram approach to automating lead time estimation in Agile Kan- 6–9. [Link]
ban project management. Expert Syst Appl 187:115866. [Link] 235. DeZoort FT, Pollard TJ (2023) An evaluation of root cause anal-
doi.​org/​10.​1016/j.​eswa.​2021.​115866 ysis use by internal auditors. J Acc Public Policy 42:107081.
219. Lei H, Ganjeizadeh F, Jayachandran PK, Ozcan P (2017) A sta- [Link]
tistical analysis of the effects of Scrum and Kanban on software 236. Ito A, Hagström M, Bokrantz J et al (2022) Improved root cause
development projects. Robot Comput-Integr Manuf 43:59–67. analysis supporting resilient production systems. J Manuf Syst
[Link] 64:468–478. [Link]
220. Jou Lin C, Frank Chen F, Min Chen Y (2013) Knowledge kanban 237. Yue W, Chai J, Wan X et al (2023) Root cause analysis for pro-
system for virtual research and development. Robot Comput- cess industry using causal knowledge map under large group
Integr Manuf 29:119–134. [Link] environment. Adv Eng Inf 57:102057. [Link]
04.​020 aei.​2023.​102057
221. Lage Junior M, Godinho Filho M (2010) Variations of the kan- 238. Linville JW, Schumann D, Aston C et al (2016) Using a six sigma
ban system: literature review and classification. Int J Prod Econ fishbone analysis approach to evaluate the effect of extreme
125:13–21. [Link] weather events on Salmonella positives in young chicken slaugh-
222. Xanthopoulos AS, Koulouriotis DE, Botsaris PN (2015) Single- ter establishments. J Food Prot 79:2048–2057. [Link] oi.o​ rg/1​ 0.​
stage Kanban system with deterioration failures and condition- 4315/​0362-​028X.​JFP-​16-​173
based preventive maintenance. Reliab Eng Syst Saf 142:111– 239. Beachum D (2005) Lean manufacturing beefs up marginspull
122. [Link] systems, takt time, and one-piece flow benefit the operation of a
223. Yang L, Zhang X (2009) Design and application of Kanban con- powder coating system. Met Finish 103:20–25. [Link]
trol system in a multi-stage, mixed-model assembly line. Syst 10.​1016/​S0026-​0576(05)​80010-8
Eng - Theory Pract 29:64–72. [Link] 240. Zhang W, Hou L, Jiao RJ (2021) Dynamic takt time decisions
8651(10)​60070-1 for paced assembly lines balancing and sequencing consider-
224. Shahabudeen P, Sivakumar GD (2008) Algorithm for the design ing highly mixed-model production: an improved artificial bee
of single-stage adaptive Kanban system. Comput Ind Eng colony optimization approach. Comput Ind Eng 161:107616.
54:800–820. [Link] [Link]
225. Nakashima K, Gupta SM (2012) A study on the risk management 241. Tang J, Liu Y, Lin K, Li L (2023) Process bottlenecks identifi-
of multi Kanban system in a closed loop supply chain. Int J Prod cation and its root cause analysis using fusion-based clustering
Econ 139:65–68. [Link] and knowledge graph. Adv Eng Inf 55:101862. [Link]
226. Ching JM, Williams BL, Idemoto LM, Blackmore CC (2014) 10.​1016/j.​aei.​2022.​101862
Using lean automation with a human touch to improve medica- 242. Cheng S-F, Nicholson BE, Epelman MA et al (2013) A dynamic
tion safety: a step closer to the perfect dose. Jt Comm J Qual programming approach to achieving an optimal end-state along
Patient Saf 40:341-AP3. [Link] oi.o​ rg/1​ 0.1​ 016/S ​ 1553-7​ 250(14)​ a serial production line. IIE Trans 45:1278–1292. [Link] oi.o​ rg/​
40045-X 10.​1080/​07408​17X.​2013.​770183
227. Dey BK, Sarkar B, Seok H (2021) Cost-effective smart autono- 243. Subramaniyan M, Skoogh A, Bokrantz J et al (2021) Artificial
mation policy for a hybrid manufacturing-remanufacturing. intelligence for throughput bottleneck analysis – state-of-the-art
Comput Ind Eng 162:107758. [Link] oi.o​ rg/1​ 0.1​ 016/j.c​ ie.2​ 021.​ and future directions. J Manuf Syst 60:734–751. [Link]
107758 10.​1016/j.​jmsy.​2021.​07.​021
228. Sartal A, Martinez-Senra AI, Cruz-Machado V (2018) Are all 244. Smętkowska M, Mrugalska B (2018) Using six Sigma DMAIC
lean principles equally eco-friendly? A panel data study. J Clean to improve the quality of the production process: a case study.
Prod 177:362–370. [Link] oi.o​ rg/1​ 0.1​ 016/j.j​ clepr​ o.2​ 017.1​ 2.1​ 90 Procedia - Soc Behav Sci 238:590–596. [Link] oi.o​ rg/1​ 0.1​ 016/j.​
229. Garza-Reyes JA, Kumar V, Chaikittisilp S, Tan KH (2018) The sbspro.​2018.​04.​039
effect of lean methods and tools on the environmental perfor- 245. Monday LM (2022) Define, measure, analyze, improve, control
mance of manufacturing organisations. Int J Prod Econ 200:170– (DMAIC) methodology as a roadmap in quality improvement.
180. [Link] Glob J Qual Saf Healthc 5:44–46. [Link]
230. Toki GFI, Ahmed T, Hossain ME et al (2023) Single minute JQSH-​22-​X2
exchange die (SMED): a sustainable and well-timed approach for 246. Nowotarski P, Szymanski P, Rzepecka P (2019) DMAIC method
Bangladeshi garments industry. Clean Eng Technol 12:100592. of quality improvement of ground works processes: case study.
[Link] IOP Conf Ser Earth Environ Sci 221:012002. [Link]
231. Almomani MA, Aladeemy M, Abdelhadi A, Mumani A (2013) 1088/​1755-​1315/​221/1/​012002
A proposed approach for setup time reduction through integrat- 247. Jamil N, Gholami H, Saman MZM et al (2020) DMAIC-based
ing conventional SMED method with multiple criteria decision- approach to sustainable value stream mapping: towards a sus-
making techniques. Comput Ind Eng 66:461–469. [Link] oi.o​ rg/​ tainable manufacturing system. Econ Res-Ekon Istraživanja
10.​1016/j.​cie.​2013.​07.​011 33:331–360. [Link]
232. Afonso M, Gabriel AT, Godina R (2022) Proposal of an inno- 248. Yu C-M, Huang T-H, Chen K-S, Huang T-Y (2022) Construct six
vative ergonomic SMED model in an automotive steel springs sigma DMAIC improvement model for manufacturing process
industrial unit. Adv Ind Manuf Eng 4:100075. [Link] oi.o​ rg/1​ 0.​ quality of multi-characteristic products. Mathematics 10:814.
1016/j.​aime.​2022.​100075 [Link]
233. Kemal Karasu M, Cakmakci M, Cakiroglu MB et al (2014) 249. e Silva LC, de Carvalho César Sobrinho ÁA, Cordeiro TD et al
Improvement of changeover times via Taguchi empowered (2023) Applications of convolutional neural networks in educa-
SMED/case study on injection molding production. Measure- tion: a systematic literature review. Expert Syst Appl. [Link] oi.​
ment 47:741–748. [Link] org/​10.​1016/j.​eswa.​2023.​120621
09.​035 250. Hireš M, Bugata P, Gazda M et al (2022) Brief overview of
234. Perico P, Mattioli J (2020) Empowering process and control in neural networks for medical applications. Acta Electrotech Inf
Lean 4.0 with artificial intelligence. 2020 Third Int Conf Artif 22:34–44. [Link]
5416 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

251. Wu T-W, Zhang H, Peng W et al (2023) Applications of convo- 267. Elhabashy A e. (1), Wells L j. (2), Camelio J a. (3) (2019) Cyber-
lutional neural networks for intelligent waste identification and physical security research efforts in manufacturing - a literature
recycling: a review. Resour Conserv Recycl 190. [Link] review. In: Procedia Manufacturing. Elsevier B.V., pp 921–931
10.​1016/j.​resco​nrec.​2022.​106813 268. Shahin M (2022) Evaluating the fidelity and efficiency of network
252. Abade A, Ferreira PA, de Barros Vidal F (2021) Plant diseases intrusion detection systems via deep learning, machine learning,
recognition on images using convolutional neural networks: a and deep hybrid learning in Industrial IoT Devices. Doctoral dis-
systematic review. Comput Electron Agric 185. [Link] sertation, The University of Texas at San Antonio. [Link]
10.​1016/j.​compag.​2021.​106125 proqu​est.​com/​openv​iew/​8e00e​0660e​08524​54c37​5526f​03291​
253. Zhiyu G, Bailin L, Hongxian G, Jing M (2020) Research on the 35/1?​pq-​origs​ite=​gscho​lar&​cbl=​18750​&​diss=y. Accessed 28
application of convolutional neural networks in the image recog- Aug 2023
nition. Int J Adv Netw Monit Controls 5:31–38. [Link] 269. Significant Cyber Incidents | Center for Strategic and Interna-
10.​21307/​ijanmc-​2020-​014 tional Studies [Link]
254. (2023) Using machine learning and deep learning algorithms ologi​es-​progr​am/​signi​ficant-​cyber-​incid​ents. Accessed 25 Feb
for downtime minimization in manufacturing systems: an early 2022
failure detection diagnostic service. [Link] 270. O’Reilly P, Rigopoulos K, Feldman L, Witte G (2021) 2020
com. Accessed 3 Jun 2023 Cybersecurity and privacy annual report. National Institute of
255. Chriskos P, Frantzidis CA, Papanastasiou E, Bamidis PD (2021) Standards and Technology
Applications of convolutional neural networks in neurodegen- 271. Koroniotis N, Moustafa N, Sitnikova E, Turnbull B (2019)
eration and physiological aging. Int J Psychophysiol 159:1–10. owards the development of realistic botnet dataset in the internet
[Link] of things for network forensic analytics: Bot-IoT dataset. Future
256. Shahin M, Frank Chen F, Bouzary H, Hosseinzadeh A (2023) Gener Comput Syst 100:779–796. [Link]
Deploying convolutional neural network to reduce waste in pro- future.​2019.​05.​041
duction system. Manuf Lett 35:1187–1195. [Link] 272. Shahin M, Chen F, Bouzary H et al (2023) Classification and
1016/j.​mfglet.​2023.​08.​127 detection of malicious attacks in industrial IoT devices via
257. Lecun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based machine learning. In: Kim K-Y, Monplaisir L, Rickli J (eds) Flex-
learning applied to document recognition. Proc IEEE 86:2278– ible automation and Intelligent Manufacturing: the human-data-
2324. [Link] technology Nexus. Springer International Publishing, Cham, pp
258. Khan A, Sohail A, Zahoora U, Qureshi AS (2020) A survey 99–106
of the recent architectures of deep convolutional neural net- 273. Moustafa N (2021) A new distributed architecture for evaluating
works. Artif Intell Rev 53:5455–5516. [Link] AI-based security systems at the edge: network TON_IoT data-
s10462-​020-​09825-6 sets. Sustain Cities Soc 72:102994. [Link]
259. Alom MZ, Taha TM, Yakopcic C et al (2019) A state-of-the-art 2021.​102994
survey on deep learning theory and architectures. Electronics 274. Booij TM, Chiscop I, Meeuwissen E et al (2021) ToN_IoT: the
8:292. [Link] role of heterogeneity and the need for standardization of features
260. Chen L-C, Papandreou G, Kokkinos I et al (2018) DeepLab: and attack types in IoT network intrusion datasets. IEEE Internet
semantic image segmentation with deep convolutional nets, Things J 1–1. [Link]
atrous convolution, and fully connected CRFs. IEEE Trans 275. Alsaedi A, Moustafa N, Tari Z et al (2020) IEEE Access
Pattern Anal Mach Intell 40:834–848. [Link] 8:165130–165150. [Link] oi.o​ rg/1​ 0.1​ 109/A
​ CCESS.2​ 020.3​ 0228​
TPAMI.​2017.​26991​84 62. TON_IoT telemetry dataset: a new generation dataset of IoT
261. Badrinarayanan V, Kendall A, Cipolla R (2017) SegNet: a deep and IIoT for data-driven intrusion detection systems
convolutional encoder-decoder architecture for image segmenta- 276. Moustafa N, Keshky M, Debiez E, Janicke H (2020) Federated
tion. IEEE Trans Pattern Anal Mach Intell 39:2481–2495. [Link] TON_IoT windows datasets for evaluating AI-based security
doi.​org/​10.​1109/​TPAMI.​2016.​26446​15 applications. In: 2020 IEEE 19th International Conference on
262. Chen W, Li Y, Tian Z, Zhang F (2023) 2D and 3D object detec- Trust, Security and Privacy in Computing and Communications
tion algorithms from images: a survey. Array 19:100305. [Link] (TrustCom), pp 848–855
doi.​org/​10.​1016/j.​array.​2023.​100305 277. Moustafa N, Ahmed M, Ahmed S (2020) Data analytics-enabled
263. Majchrowska S, Mikołajczyk A, Ferlin M et al (2022) Deep intrusion detection: evaluations of ToN_IoT linux datasets. In:
learning-based waste detection in natural and urban environ- 2020 IEEE 19th International Conference on Trust, Security
ments. Waste Manag 138:274–284. [Link] and Privacy in Computing and Communications (TrustCom),
wasman.​2021.​12.​001 pp 727–735
264. Shahin M, Chen F, Bouzary H et al (2023) Implementation of a 278. Moustafa N, New generations of internet of things datasets for
novel fully convolutional network approach to detect and classify cybersecurity applications based machine learning : TON_IoT
cyber-attacks on IoT devices in smart manufacturing systems. In: datasets. In: Res. Data Aust. [Link]
Kim K-Y, Monplaisir L, Rickli J (eds) Flexible automation and gener​ations-​inter​net-​toniot-​datas​ets/​14259​41. Accessed 11 Dec
Intelligent Manufacturing: the human-data-technology Nexus. 2021
Springer International Publishing, Cham, pp 107–114 279. Moustafa N (2019) A systemic IoT-fog-cloud architecture for
265. Cerqueira V, Torgo L, Mozetič I (2020) Evaluating time series big-data analytics and cyber security systems: a review of fog
forecasting models: an empirical study on performance estima- computing. ArXiv190601055 Cs
tion methods. Mach Learn 109:1997–2028. [Link] 280. Ashraf J, Keshk M, Moustafa N, Abdel-Basset M, Khurshid H,
1007/​s10994-​020-​05910-7 Bakhshi AD, Mostafa RR (2021) IoTBoT-IDS: a novel statistical
266. Shahin M, Chen FF, Bouzary H, Zarreh A (2020) Frameworks learning-enabled botnet detection framework for protecting net-
proposed to address the threat of cyber-physical attacks to lean works of smart cities. Sustain Cities and Soc 72:103041. [Link]
4.0 systems. Procedia Manuf 51:1184–1191. [Link] doi.​org/​10.​1016/j.​scs.​2021.​103041
1016/j.​promfg.​2020.​10.​166 281. Zargar ST, Joshi J, Tipper D (2013) A survey of defense mecha-
nisms against distributed denial of service (DDoS) flooding
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5417

attacks. IEEE Commun Surv Tutor 15:2046–2069. [Link] 297. Zamzam AH, Hasikin K, Wahab AKA (2023) Integrated fail-
org/​10.​1109/​SURV.​2013.​031413.​00127 ure analysis using machine learning predictive system for smart
282. Krupp J, Backes M, Rossow C (2016) Identifying the scan and management of medical equipment maintenance. Eng Appl Artif
attack infrastructures behind amplification DDoS attacks. In: Intell 125:106715. [Link]
Proceedings of the 2016 ACM SIGSAC Conference on Com- 106715
puter and Communications Security. Association for Computing 298. Ahmed AAA, Mahalakshmi A, ArulRajan K et al (2023) Inte-
Machinery, New York, pp 1426–1437 grated artificial intelligence effect on crisis management and
283. Al-Rimy BAS, Maarof MA, Shaid SZM (2018) Ransomware lean production: structural equation modelling frame work. Int
threat success factors, taxonomy, and countermeasures: a survey J Syst Assur Eng Manag 14:220–227. [Link]
and research directions. Comput Secur 74:144–166. [Link] s13198-​022-​01679-1
org/​10.​1016/j.​cose.​2018.​01.​001 299. Explainable Artificial Intelligence for Predictive: Results.
284. Al-Hawawreh M, Hartog FD, Sitnikova E (2019) Targeted ran- [Link]
somware: a new cyber threat to edge system of brownfield indus- sid=​E BSCO%​3 aeds​e ee&​genre=​b ook&​i ssn=​e dseee.​I EEEC​
trial internet of things. IEEE Internet Things J. [Link] onfer​enc&​ISBN=​97817​28187​013&​volume=​&​issue=​&​date=​
10.​1109/​JIOT.​2019.​29143​90 & ​ s page= ​ 6 9& ​ p ages= ​ 6 9- ​ 7 4& ​ t itle= ​ 2 020+ ​ T hird+ ​ I nter ​ n atio​
285. Kolias C, Kambourakis G, Stavrou A, Gritzalis S (2016) Intru- nal+​Confe​rence+​on+​Artif​i cial+​Intel​ligen​ce+​for+​Indus​tries+​
sion detection in 802.11 networks: empirical evaluation of threats (AI4I)%​2c+​Artif​i cial+​Intel​ligen​ce+​for+​Indus​tries+​(AI4I)%​
and a public dataset. IEEE Commun Surv Tutor. [Link] 2c+​2020+​Third+​Inter ​natio​nal+​Confe​rence+​on%​2c+​AI4I&​
10.​1109/​COMST.​2015.​24021​61 atitle=​Expla​inable+​Artif​i cial+​Intel​ligen​ce+​for+​Predi​ctive+​
286. Zolanvari M, Teixeira MA, Gupta L et al (2019) Machine learn- Maint​enance+​Appli​catio​ns&​aulast=​Matzka%​2c+​Steph​an&​
ing-based network vulnerability analysis of industrial internet of id=​DOI%​3a10.​1109%​2fAI4​I49448.​2020.​00023​&​site=​ftf-​live.
things. IEEE Internet Things J 6:6822–6834. [Link] Accessed 6 Jul 2022
1109/​JIOT.​2019.​29120​22 300. Matzka S (2020). AI4I 2020 predictive maintenance dataset. UCI
287. Chaabouni N, Mosbah M, Zemmari A et al (2019) Network Machine Learning Repository. University of California, School
Intrusion detection for IoT security based on learning techniques. of Information and Computer Science, Irvine, CA
IEEE Commun Surv Tutor 21:2671–2701. [Link] 301. Indrawati S, Azzam A, Ramdani AC (2019) Manufacturing
1109/​COMST.​2019.​28963​80 efficiency improvement through lean manufacturing approach:
288. Shahin M, Chen FF, Hosseinzadeh A et al (2024) An AI-powered a case study in a steel processing industry. IOP Conf Ser: Mater
network intrusion detection system in industrial IoT devices via Sci Eng 598(1):012062. IOP Publishing
deep learning. In: Silva FJG, Ferreira LP, Sá JC (eds) Flexible 302. Wahab ANA, Mukhtar M, Sulaiman R (2013) A conceptual
automation and Intelligent Manufacturing: establishing bridges model of lean manufacturing dimensions. Procedia Technol
for more sustainable Manufacturing systems. Springer Nature 11:1292–1298
Switzerland, Cham, pp 1149–1156 303. Qian K (2019) Automated detection of steel defects via machine
289. Zheng A, Casari A (2018) Feature engineering for machine learn- learning based on real-time semantic segmentation. Proc 3rd Int
ing: principles and techniques for data scientists, First edition. Conf Video Image Process 42–46. [Link] oi.o​ rg/1​ 0.1​ 145/3​ 3760​
O’Reilly Media 67.​33761​13
290. Witten IH, Frank E (2002) Data mining: practical machine learn- 304. Luo Q, Fang X, Liu L et al (2020) Automated visual defect detec-
ing tools and techniques with Java implementations. Acm Sig- tion for flat steel surface: a survey. IEEE Trans Instrum Meas
mod Record 31(1):76–77 69:626–644. [Link]
291. He Y, Gu C, Chen Z, Han X (2017) Integrated predictive main- 305. Guo B, Wang Y, Zhen S et al SPEED: semantic prior and
tenance strategy for manufacturing systems by combining extremely efficient dilated convolution network for real-time
quality control and mission reliability analysis. Int J Prod Res metal surface defects detection. IEEE Trans Ind Inf 1–11. [Link]
55(19):5841–5862. [Link] doi.​org/​10.​1109/​TII.​2022.​32336​74
13468​43 306. (2023) Waste reduction via computer vision-based inspection:
292. Froger A, Gendreau J, Mendoza E (2016) Maintenance schedul- towards lean systems in metal production. [Link]
ing in the electricity industry: a literature review. Eur J Oper Res rchsq​uare.​com. Accessed 3 Jun 2023
251:695–706 307. Akhyar F, Lin C-Y, Muchtar K et al (2019) High efficient single-
293. Ding S, Kamaruddin S (2014) Maintenance policy optimiza- stage steel surface defect detection. 2019 16th IEEE Int Conf
tion – literature review and directions. Int J Adv Manuf Technol Adv Video Signal Based Surveill AVSS Adv Video Signal Based
76:1263–1283 Surveill AVSS 2019 16th IEEE Int Conf On 1–4. [Link]
294. Meissner R, Rahn A, Wicke K (2021) Developing prescriptive 10.​1109/​AVSS.​2019.​89098​34
maintenance strategies in the aviation industry based on a dis- 308. Utsunomiya H, Hara K, Matsumoto R, Azushima A (2014) For-
crete-event simulation framework for post-prognostics decision mation mechanism of surface scale defects in hot rolling process.
making. Reliab Eng Syst Saf 214. [Link] CIRP Ann 63:261–264. [Link]
2021.​107812 022
295. An Y, Chen X, Hu J et al (2023) Integrated optimization of condi- 309. Bao Y, Song K, Liu J et al (2021) Triplet-graph reasoning net-
tion-based preventive maintenance and production rescheduling work for few-shot metal generic surface defect segmentation.
with multi-phase processing speed selection and old machine IEEE Trans Instrum Meas 70. [Link] oi.o​ rg/1​ 0.1​ 109/T ​ IM.2​ 021.​
scrap. Reliab Eng Syst Saf 238 30835​61
296. Boumallessa Z, Chouikhi H, Elleuch M, Bentaher H (2023) Mod- 310. Althubiti SA, Alenezi F, Shitharth S et al (2022) Circuit manu-
eling and optimizing the maintenance schedule using dynamic facturing defect detection using VGG16 convolutional neural
quality and machine condition monitors in an unreliable single networks. Wirel Commun Mob Comput 1–10. [Link]
production system. Reliab Eng Syst Saf 235:109216. [Link] 10.​1155/​2022/​10704​05
org/​10.​1016/j.​ress.​2023.​109216 311. Cancer Facts and Statistics [Link] ​ ww.c​ ancer.o​ rg/r​ esear​ ch/c​ an-
cer-​facts-​stati​stics.​html. Accessed 9 Oct 2022
5418 The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419

312. Guy JGP, Machlin SR, Ekwueme DU, Yabroff KR (2015) Preva- 329. Alam TM, Shaukat K, Khan WA et al (2022) An efficient deep
lence and costs of skin cancer treatment in the U.S., 2002 – 2006 learning-based skin cancer classifier for an Imbalanced dataset.
and 2007 – 2011. Am J Prev Med 48:183–187. [Link] oi.o​ rg/1​ 0.​ Diagn 2075–4418 12:[Link] [Link]
1016/j.​amepre.​2014.​08.​036 diagn​ostic​s1209​2115
313. Alam TM, Khan MMA, Iqbal MA et al (2019) Cervical cancer 330. (2022) JAMA Oncology. JAMA Oncol 8:1085. [Link]
prediction through different screening methods using data min- 10.​1001/​jamao​ncol.​2021.​5511
ing. Int J Adv Comput Sci Appl 10:388–396. [Link] 331. Khandpur S, Ramam M (2012) Skin tumours. J Cutan Aesthetic
14569/​ijacsa.​2019.​01002​51 Surg 5:159–162. [Link]
314. Siegel RL, Miller KD, Jemal A (2016) Cancer statistics, 2016. 332. Tschandl P, Rinner C, Apalla Z et al (2020) Human–computer
CA Cancer J Clin 66:7–30. [Link] collaboration for skin cancer recognition. Nat Med 26:1229–
315. Tavakolpour S (2017) 1 2), Daneshpazhooh M(2), Mahmoudi 1234. [Link]
H(2). Skin cancer: genetics, immunology, treatments, and psy- 333. Sabanci K, Aslan MF, Ropelewska E, Unlersen MF (2022) A
chological care. Springer International Publishing convolutional neural network-based comparative study for pep-
316. Dascalu A, David EO (2019) Skin cancer detection by deep per seed classification: analysis of selected deep features with
learning and sound analysis algorithms: a prospective clinical support vector machine. J Food Process Eng 45:e13955. [Link]
study of an elementary dermoscope. EBioMedicine 43:107–113. doi.​org/​10.​1111/​jfpe.​13955
[Link] 334. Johnson JM, Khoshgoftaar TM (2019) Survey on deep learning
317. Shahin M, Maghanaki M, Chen FF, Hosseinzadeh A (2024) with class imbalance. J Big Data 6:27. [Link]
Integrating lean healthcare and machine learning for cancer risk s40537-​019-​0192-5
prediction. In: Daimi K, Al Sadoon A (eds) Proceedings of the 335. Saini M, Susan S (2020) Deep transfer with minority data aug-
Second International Conference on Advances in Computing mentation for imbalanced breast cancer dataset. Appl Soft Com-
Research (ACR’24). Springer Nature Switzerland, Cham, pp put 97:106759. [Link]
373–381 336. Bhimavarapu U, Battineni G (2022) Skin lesion analysis for
318. Niino M, Matsuda T (2021) Age-specific skin cancer incidence melanoma detection using the novel deep learning model fuzzy
rate in the world. Jpn J Clin Oncol 51:848–849. [Link] GC-SCNN. Healthcare 10:962. [Link]
10.​1093/​jjco/​hyab0​57 hcare​10050​962
319. Petrie T, Samatham R, Witkowski AM et al (2019) Melanoma 337. Batool D, Shahbaz M, Shahzad Asif H et al (2022) A hybrid
early detection: big data, bigger picture. J Invest Dermatol approach to tea crop yield prediction using simulation models
139:25–30. [Link] and machine learning. Plants 11:1925 [Link]
320. Marks R (2000) Epidemiology of melanoma. Clin Exp Dermatol plant​s1115​1925.
25:459–463. [Link] 338. Franses PH, Welz M (2022) Forecasting real GDP growth for
321. Elina Reponen R, Jokela JC, Blodgett et al (2021) Validation of Africa. Econometrics 10:3. [Link]
the lean healthcare implementation self-assessment instrument cs100​10003
(LHISI) in the Finnish healthcare context. BMC Health Serv Res 339. Kroos P, Schabus M, Verbeeten FHM (2021) The relation
21:1–9. [Link] between internal forecasting sophistication and accounting mis-
322. Maghanaki M, Shahin M, Chen FF, Hosseinzadeh A (2024) reporting. J Manag Acc Res 34:51–73. [Link]
Improving early diagnosis: the intersection of lean healthcare JMAR-​2020-​072
and computer vision in cancer detection. In: Daimi K, Al Sadoon 340. de Prince D, Marçal EF, Valls Pereira PL (2022) Forecasting
A (eds) Proceedings of the Second International Conference on industrial production using its aggregated and disaggregated
Advances in Computing Research (ACR’24). Springer Nature series or a combination of both: evidence from one emerging
Switzerland, Cham, pp 404–413 market economy. Econometrics 10:27. [Link]
323. Shahin M, Chen FF, Hosseinzadeh A et al (2023) A smartphone- econo​metri​cs100​20027
based application for an early skin disease prognosis: towards a 341. Zuo Y, Hu Y, Liu X, Zhao H (2022) Corporate misconduct and
lean healthcare system via computer-based vision. Adv Eng Inf analyst forecasting accuracy: evidence from China. Emerg Mark
57:102036. [Link] Finance Trade 58:3006–3022. [Link] oi.o​ rg/1​ 0.1​ 080/1​ 54049​ 6X.​
324. Shahin M, Chen FF, Hosseinzadeh A, Maghanaki M (2024) 2022.​20572​20
Deploying deep convolutional neural network to the battle 342. Conceição J, de Souza J, Rossini EG et al (2021) Implementation
against cancer: towards flexible healthcare systems. Inf Med of inventory management in footwear industry. J Ind Eng Manag
Unlocked 47:101494. [Link] oi.o​ rg/1​ 0.1​ 016/j.i​ mu.2​ 024.1​ 01494 14:360–375. [Link]
325. Martins Drei S, Sérgio de Arruda Ignácio P (2022) Lean health- 343. Guijarro E, Babiloni E, Canós-Darós MJ et al (2020) Fuzzy
care applied systematically in a medium-sized medical clinic modeling approach to on-hand stock levels estimation in (R, S)
hospitalization. J Health Organ Manag 36:666–689. [Link] inventory system with lost sales. J Ind Eng Manag 13:464–474.
org/​10.​1108/​JHOM-​05-​2021-​0194 [Link]
326. Tlapa D, Tortorella G, Fogliatto F et al (2022) Effects of lean 344. Semenchuk K, Shutenko T (2021) Research of material and tech-
interventions supported by digital technologies on healthcare nical resources inventory management in projects supply chains
services: a systematic review. Int J Environ Res Public Health 345. Bi X, Adomavicius G, Li W, Qu A (2022) Improving sales fore-
19:9018. [Link] casting accuracy: a tensor factorization approach with demand
327. Tschandl P, Rosendahl C, Kittler H (2018) The HAM10000 data- awareness. INFORMS J Comput 34(3):1644–1660. [Link]
set, a large collection of multi-source dermatoscopic images of org/​10.​1287/​ijoc.​2021.​1147
common pigmented skin lesions. Sci Data 5:180161. [Link] 346. Hoseinzade E, Haratizadeh S (2019) CNNpred: CNN-based stock
org/​10.​1038/​sdata.​2018.​161 market prediction using a diverse set of variables. Expert Syst
328. Han SS, Kim MS, Lim W et al (2018) Classification of the clini- Appl 129:273–285. [Link]
cal images for benign and malignant cutaneous tumors using 347. Zhang L, Yang F, Daniel Zhang Y, Zhu YJ (2016) Road crack
a deep learning algorithm. J Invest Dermatol 138:1529–1538. detection using deep convolutional neural network. In: 2016
[Link] IEEE International Conference on Image Processing (ICIP), pp
3708–3712
The International Journal of Advanced Manufacturing Technology (2024) 133:5343–5419 5419

348. Özgenel ÇF, Sorguç AG (2018) Performance comparison of 361. Kusiak A (2017) Smart manufacturing must embrace big data.
pretrained convolutional neural networks on crack detection in Nature 544:23–25. [Link]
buildings. In: Isarc Proceedings of the international symposium 362. Ayvaz S, Alpay K (2021) Predictive maintenance system for
on automation and robotics in construction 35:1–8. IAARC production lines in manufacturing: a machine learning approach
Publications using IoT data in real-time. Expert Syst Appl Int J 173. [Link]
349. Sánchez-Núñez P, Cobo MJ, De Las Heras-Pedrosa C, Peláez JI, doi.​org/​10.​1016/j.​eswa.​2021.​114598
Herrera-Viedma E (2020) Opinion mining, sentiment analysis 363. Proteau A, Zemouri R, Tahan A, Thomas M (2020) Dimension
and emotion understanding in advertising: a bibliometric analy- reduction and 2D-visualization for early change of state detection
sis. IEEE Access 8:134563–134576. [Link] in a machining process with a variational autoencoder approach.
ACCESS.​2020.​30094​82 Int J Adv Manuf Technol 111:3597–3611. [Link]
350. Wang H, Hou M (2023) Quantum-like implicit sentiment analysis 1007/​s00170-​020-​06338-y
with sememes knowledge. Expert Syst Appl 232:120720. [Link] 364. Ou J, Li H, Huang G, Zhou Q (2020) A novel order analysis and
doi.​org/​10.​1016/j.​eswa.​2023.​120720 stacked sparse auto-encoder feature learning method for milling
351. Sawale GJ, Rawat MK (2022) Stock market prediction using tool wear condition monitoring. Sensors 20:2878. [Link] oi.o​ rg/​
sentiment analysis and machine learning approach. In: 2022 4th 10.​3390/​s2010​2878
International Conference on Smart Systems and Inventive Tech- 365. Taiebat M, Sassani F (2017) Distinguishing sensor faults from
nology (ICSSIT), pp 1–6 system faults by utilizing minimum sensor redundancy. Trans
352. Ireland R, Liu A (2018) Application of data analytics for product Can Soc Mech Eng 41:469–487. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 139/​
design: Sentiment analysis of online product reviews. CIRP J tcsme-​2017-​1033
Manuf Sci Technol 23:128–144. [Link] 366. Ferguson M, Ak R, Lee Y-TT, Law KH (2018) Detection and seg-
2018.​06.​003 mentation of manufacturing defects with Convolutional Neural
353. Customer Support on Twitter [Link] Networks and Transfer Learning
thoug​htvec​tor/​custo​mer-​suppo​r t-​on-​twitt​er. Accessed 7 May 367. Imoto K, Nakai T, Ike T, Haruki K, Sato Y (2019) A CNN-based
2023 transfer learning method for defect classification in semiconduc-
354. PointScrape by Thought Vector - Super Simple (2023) Super tor manufacturing. IEEE Trans Semicond Manuf 32(4):455–459.
scale datasets as a service. [Link] [Link]
scrape/. Accessed 7 368. Wu D, Liu S, Zhang L et al (2017) A fog computing-based frame-
355. Shahin M, Chen FF, Hosseinzadeh A et al (2024) A novel work for process monitoring and prognosis in cyber-manufactur-
approach to voice of customer extraction using GPT-3.5 Turbo: ing. J Manuf Syst 43:25–34. [Link] oi.o​ rg/1​ 0.1​ 016/j.j​ msy.2​ 017.​
linking advanced NLP and lean six Sigma 4.0. Int J Adv Manuf 02.​011
Technol. [Link] 369. Koodiani HK, Jafari E, Majlesi A et al (2024) Machine learning
356. Khan S, Yairi T (2018) A review on the application of deep tools to improve nonlinear modeling parameters of RC columns.
learning in system health management. Mech Syst Signal Process J Build Eng 84:108492. [Link]
107:241–265. [Link] 108492
357. Wang J, Ma Y, Zhang L et al (2018) Deep learning for smart 370. Shahin M, Chen FF, Hosseinzadeh A (2024) Harnessing custom-
manufacturing: methods and applications. J Manuf Syst 48:144– ized AI to create voice of customer via GPT3.5. Adv Eng Inf
156. [Link] 61:102462. [Link]
358. Misaka T, Herwan J, Kano S et al (2020) Deep neural network-
based cost function for metal cutting data assimilation. Int J Publisher’s Note Springer Nature remains neutral with regard to
Adv Manuf Technol 107:385–398. [Link] jurisdictional claims in published maps and institutional affiliations.
s00170-​020-​04984-w
359. Miotto R, Wang F, Wang S et al (2018) Deep learning for health- Springer Nature or its licensor (e.g. a society or other partner) holds
care: review, opportunities and challenges. Brief Bioinform exclusive rights to this article under a publishing agreement with the
19:1236–1246. [Link] author(s) or other rightsholder(s); author self-archiving of the accepted
360. Li D-C, Wen I-H, Chen W-C (2016) A novel data transformation manuscript version of this article is solely governed by the terms of
model for small data-set learning. Int J Prod Res 54:7453–7463. such publishing agreement and applicable law.
[Link]