Received 27 May 2024; revised 26 June 2024 and 9 July 2024; accepted 12 July 2024.

Date of publication 19 July 2024;

date of current version 30 September 2024. The review of this article was arranged by Associate Editor S. Karnouskos.
Digital Object Identifier 10.1109/OJIES.2024.3431240

Machine Learning Applications in

Manufacturing—Challenges, Trends,
and Future Directions
ALEXANDRE MANTA-COSTA 1,2,4 , SARA OLEIRO ARAÚJO 1,3,4 , RICARDO SILVA PERES 1,2,4 ,

AND JOSÉ BARATA 1,2,4 (Member, IEEE)

1
UNINOVA—Centre of Technology and Systems (CTS), FCT Campus, 2829-516 Caparica, Portugal
2
Department of Electrical and Computer Engineering, NOVA School of Science and Technology, NOVA University of Lisbon, 2829-516 Caparica, Portugal
3
Earth Sciences Department (DCT), NOVA School of Science and Technology, NOVA University of Lisbon, 2829-516 Caparica, Portugal
4
Associated Laboratory of Intelligent Systems (LASI), NOVA University of Lisbon, 2829-516 Caparica, Portugal
CORRESPONDING AUTHOR: ALEXANDRE MANTA-COSTA (e-mail: [email protected]).
This work supported by Fundação para Ciência e Tecnologia through the program under Grant UIDB/00066/2020 and Center of Technology and Systems (CTS).

ABSTRACT The emergence of Industry 4.0 (I4.0) has significantly transformed manufacturing landscapes,
introducing interconnected, dynamic, and data-rich environments. This article focuses on the application
of industrial machine learning (I-ML) within these evolving manufacturing contexts, exploring both the
challenges and future prospects of its integration. A systematic literature review, following the preferred
reporting items for systematic reviews and meta-analyzes (PRISMA) guidelines, forms the foundation of our
analysis, characterizing the role of machine learning (ML) in modern manufacturing, its current challenges,
and future trends. This research delves into the implications of I-ML in various manufacturing scenarios, in-
cluding predictive maintenance, anomaly detection, and quality control, providing a comprehensive overview
of practical applications along with an identification of related emerging technologies and trends. We also
address the critical need for sustainable, reproducible, and reliable performance in industrial applications
and explore strategies for overcoming barriers to ML adoption in the industry. Recommendations for future
research directions are provided, aiming to bridge the gap between ML advancements and their practical,
scalable implementation in industrial settings, paving the way to future research in the field. Lastly, we
aim to contribute to the identification of challenges and future research directions for the ongoing digital
transformation of manufacturing industries, offering insights into how ML can be effectively leveraged in
the era of I4.0.

INDEX TERMS Industrial artificial intelligence (I-AI), industrial machine learning (I-ML), Industry 4.0
(I4.0), machine learning (ML), manufacturing, systematic review.

NOMENCLATURE CAE Convolutional autoencoder.

AD Anomaly detection. CART Classification and regression trees.
AI Artificial intelligence. CMEANS Fuzzy C-means clustering.
AR Augmented reality. CNN Convolutional neural network.
AT Analytics technologies. CPS Cyber-physical systems.
BB Balanced bagging. DAGM Deutsche Arbeitsgemeinschaft für Muster-
BDT Bounded decision tree. erkennung dataset.
BiFPN Weighted Bi-directional feature pyramid net- DIANA DIvisive ANAlysis Clustering.
work. DNN Deep neural network.
BiLSTM Bidirectional long short-term memory. DPCA Principal components analysis.
BRF Balanced random forest. DRN Deep recurrent network.

© 2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see
VOLUME 5, 2024 https://creativecommons.org/licenses/by-nc-nd/4.0/ 1085
MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

DT Data technologies. XGBoost EXtreme gradient boosting.

DRL Deep reinforcement learning.
ESSM Ensemble sparse supervised model. I. INTRODUCTION
E-STGCN Electrical spatio-temporal graph convolu- With the advent of the fourth industrial revolution, commonly
tional network. known as Industry 4.0 (I4.0), the employment of CPS has be-
FRCNN Faster R-CNN. come widely spread as a source of automation and information
GAN Generative adversarial networks. sharing. Leveraging the capacities of integration provided by
GCN Graph convolution network. the CPS, the concept of smart manufacturing has emerged,
GMM Gaussian mixture model. representing a fully integrated approach to the manufacturing
GPR Gaussian process region. processes. This integration has been made feasible through
GRU Gated recurrent unit. the application of key enabling technologies, such as AI,
HC Hierarchical clustering. cloud computing, big data, advanced analytics, and IoT [1].
HMT Human–machine technologies. Manufacturers have created and collected significant volumes
HRI Human–robot interaction. of data from these technologies and implemented actionable
HRNet High-resolution network. insights. New plants capable of automating data collection,
I-AI Industrial artificial intelligence. analysis, and decision-making have been labeled smart fac-
I-ML Industrial machine learning. tories, aiming to achieve more efficient production through
I4.0 Industry 4.0. real-time monitoring, predictive analysis, and reduced waste
ICT Information and communication technolo- and downtime [2].
gies. AI is considered a field of computer science that aims to
IoT Internet of Things. provide computers with the ability to make things that require
K-NN K-nearest neighbor. human intelligence. In the context of I4.0, AI has faced sig-
KCF Kernelized correlation filter. nificant improvements mainly due to the accessibility to more
KM k-means. powerful computation and, more importantly, to the quantities
LDA Linear discriminant analysis. of data produced by the CPS. In I4.0, AI aims to provide
LR Linear Regression. intelligent capabilities to the CPS by leveraging tremendous
LSTM Long short-term memory network. amounts of data generated at the enterprise level, crossing
LSTMCSMC LSTM with chained sequence memory cells. manufacturing activities and equipment, to labor activities,
MLOps Machine learning operations. and even environmental conditions [3]. Notably, generated
ML Machine learning. data in manufacturing systems is rapidly increasing, with an
MLP Multilayer perceptron. estimated 1000 EB collected in 2015 and an expected 20-fold
NLP Natural language processing. growth by 2025 [4].
NMF Nonnegative matrix factorization. I-AI focuses on overcoming the barriers imposed by the
NN Neural network. industry in adopting AI methods. Lee [5] defines I-AI as a sys-
OT Operation technologies. tematic discipline that emphasizes developing, validating, and
PAM Partition around medoids. deploying AI algorithms in industrial applications with the
PdM Predictive maintenance. sustainable performance. This discipline provides the ability
PPS Production planning and scheduling. to solve industry problems effectively, repeatedly, and reli-
PRISMA Preferred reporting items for systematic re- ably by combining academic research outcomes with industry
views and meta-analysis. expertise [6], [7].
PT Platform technologies. In the manufacturing domain, AI has faced several barri-
QI Quality inspection. ers [6]. For instance, the absence of demonstrated repeata-
ResNets Residual neural networks. bility and consistency poses challenges for establishing its
RF Random forest. usability in industrial settings. In addition, the dependability
R-CNN Region-based CNN. of AI, influenced by developer experience, limits its broader
RFR Random forest regressor. applicability in this context.
RL Reinforcement learning. ML, a subdomain within the broader field of AI, is gaining
SAE Siamese autoencoder. traction in various industrial applications [8]. Both terms are
SML Supervised machine learning. often used synonymously since a wide spectrum of research
SLSTM Stack long short-term memory network. into AI in the manufacturing industry generally refers to using
SoS System of systems. ML algorithms [9], [10].
SVC Support vector classifier. While the field of I-AI encompasses various techniques
TR-TRS Tensor ring decomposed transformer. beyond just ML, delving into a subfield of I-AI specifically
TRL Technology readiness level. focused on ML applications can unveil solutions applicable
UML Unsupervised machine learning. to a broader context. In ML, there are already techniques
VR Virtual reality. that might contribute to the lifecycle and applicability of I-AI

1086 VOLUME 5, 2024

in manufacturing. For instance, MLOps is an engineering The rest of this article is organized as follows. Section II
practice that addresses issues in designing and maintaining provides a general overview of I-AI, and Section III explains
ML models in production scenarios [11]. This field can be the survey methodology. Section IV provides an overview
applied to the I-AI paradigm to enhance it further and apply and characterization of the current landscape of AI imple-
it to new scenarios. mentation in the manufacturing industry. Section V proposes
I-ML emerges as a subdiscipline of I-AI, inheriting not only a possible taxonomy for applying I-ML. In Section VI, a
its capabilities but also its challenges. In order to further delve discussion of the found result is made. In Section VII, the
into the realm of I-ML, it is important to understand the I- limitations of this work are elaborated. Finally, Section VII
AI taxonomy and then focus on the ML applications within concludes this article.
it. However, many challenges remain in the industry’s current
applications of ML. II. INDUSTRIAL ARTIFICIAL INTELLIGENCE
To bridge the gap between the current ML implementations A. GENERAL OVERVIEW
in an industrial setting and the concept of I-AI, a system- I-AI is a systematic discipline that focuses on developing,
atic literature review was conducted following the preferred validating, deploying, and maintaining AI solutions in indus-
reporting items for systematic reviews and meta-analyzes trial applications and aims to improve the reliability, precision,
(PRISMA) guidelines in order to evaluate the current ML efficiency, and future optimization of engineering systems [6],
implementations and identify common flaws and understand [12]. It is an interdisciplinary field that bridges academic
how can I-ML be characterized and applied in. research with industrial practice and seamlessly integrates
To guide this literature review, the following research ques- advanced AI methodologies, including ML, natural language
tions were defined. processing (NLP), and robotics, into the heart of industrial
RQ1: How can applications of ML in manufacturing be operations [13].
characterized in the context of I-ML? I-AI distinguishes itself from generic AI in the following
To answer this research question, an evaluation of the state five dimensions [13].
of the art from an I-ML perspective is going to be conducted. 1) Infrastructure: I-AI emphasizes real-time processing,
In order to characterize ML applications from an I-ML per- robustness, security, and high-level interconnectivity,
spective, an assessment will be made of the technologies used, ensuring that systems can handle the demanding re-
the state of the solutions presented in relation to their TRL, quirements of industrial environments.
and the inclusion of humans in the solutions presented. The 2) Data Handling: I-AI requires large volumes of data with
aim of this research question is to characterize the state of great variety and high velocity, originating from several
current ML applications and identify the gap between them units, products, and regimes.
and the I-ML goals. 3) Algorithms: I-AI goes beyond traditional AI algorithms
RQ2: What challenges and limitations are commonly iden- by integrating physical, digital, and heuristic knowl-
tified in the implementation of I-ML? edge, taking advantage of the nuances of industrial
The design principles will be evaluated and complemented operations and allowing for practical and effective so-
with an evaluation of the challenges for each application sce- lutions.
nario identified. With this research question, an assessment 4) Decision Making: The margin for errors in an indus-
of the crucial point of failure in current ML implementations trial setting is typically very low. I-AI aids in efficient
will be made in order to identify possible solutions for future decision-making, emphasizing handling uncertainties
research. and optimizing large-scale problems.
RQ3: How can future I-ML systems be designed to meet 5) Objectives: Compared to general AI, I-AI focuses on a
industrial requirements? small range of value creation. Whether through waste
Based on the challenges identified in the literature review, reduction, quality enhancement, operator performance
we will propose possible research venues to address the bar- boosting, or process speed, the main goal is to enhance
riers in current ML applications and advance the I-AI vision. tangible, real-world industrial outputs
These solutions will be designed to guide future research and Furthermore, a significant aspect of I-AI is the enhancement
the application of I-ML, ensuring adherence to the design of the workers’ capabilities rather than their replacement [14],
principles of I-ML applications. [15]. As the industry marches toward more autonomous sce-
In this context, the main contributions of this work can be narios, the predominant sentiment is to leverage AI to improve
summarized as follows: human performance in the industrial setting.
1) a literature review following PRISMA guidelines; I-AI serves as a bridge between AI technologies and in-
2) the characterization of existing I-AI implementations; dustrial processes. By integrating AI techniques into several
3) identification of key employed technologies and appli- aspects of the industrial value chain, from research and de-
cations; velopment (R&D) to production and service, this subject
4) development of a taxonomy of research challenges, op- offers an innovative approach to traditional industrial opera-
portunities, and future trends in I-AI. tions [16].

VOLUME 5, 2024 1087

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

B. TAXONOMY data analytics and statistical methods to uncover pat-

I-AI was a term coined by Dr. Jay Lee to describe the applica- terns, correlations, and insights from manufacturing
tions of AI in industrial scenarios in a systematic and sustain- data. This information supports tasks such as predictive
able way [6]. As a recent discipline, I-AI is under active devel- maintenance (PdM) and quality inspection (QI), helping
opment, and as of today, there are still several flawed designs, manufacturers detect early flaws and manage machine
configurations, and implementations [17]. To better under- health. AT integrates directly gathered data with in-
stand I-AI approaches and to guide the design of I-AI systems, formation from other relevant technologies to improve
it is essential to understand some aspects of its taxonomy. manufacturing systems, allowing manufacturers to iden-
tify bottlenecks or underperforming components. In
addition, AT emphasizes data visualization with user-
1) KEY ELEMENTS friendly interfaces to ensure that all personnel, from
As a systems subject, I-AI requires the integration of the shop-floor operators to managers, can easily interpret
following six key elements [12]. the information and gain enhanced insights.
1) Analytics: Assumes the core algorithms of AI, their 3) Platform Technologies (PTs) [5], [6], [12]: PTs are es-
adaptation to specific scenarios, and their alignment sential for achieving smart manufacturing goals, such
with their predefined objectives. as agility and complex processing. They come in the
2) Big Data: Ensures the availability and accessibility of following three main configurations.
relevant data that can be processed and transformed into a) Stand-alone systems operating independently.
actionable insights. Industrial big data needs to consider b) Fog/edge systems where most industrial data are
three main challenges: bad quality data, broken data, generated and analytics are deployed.
and context. c) Cloud systems using remote servers for data stor-
3) Computational Platform: Encompasses cloud, fog, and age and processing, offering faster deployments,
edge computing integration based on embedded intelli- customization, knowledge integration, and scala-
gence. It also includes integrating distributed systems bility without high infrastructure costs.
with centralized systems and creating quick platform PTs are crucial in I-AI, supporting various layers of
reconfiguration. CPS architecture from signal acquisition and compu-
4) Domain Knowledge: I-AI feeds not only on big data tation to data storage, digital twin integration, and
analytics but also on expert knowledge of the specific robot integration. They ensure connectivity across the
sector/industry where the I-AI is applied. shop floor, enabling equipment self-organization, and
5) Evidence: Real-world information that indicates the self-configuration on production lines. Security and in-
state of the system at a given time and provides valuable terconnectivity are key considerations, requiring secure
insights to aid decision-making processes. access and maintaining system privacy and integrity.
6) Feedback: Link with control systems to establish a 4) Operation Technologies (OT) [5], [6], [12]: Based on
closed-loop connection between decision-making and analytics insights, aim to achieve enterprise control
operations. and optimization through systems like product lifecycle
management, enterprise resource planning, manufactur-
ing execution systems, customer relationship manage-
2) ENABLING TECHNOLOGIES ment, and supply chain management. They enhance
As a systematic multidisciplinary field, I-AI is backed by manufacturing performance, intelligence, and efficiency
several enabling technologies that can be categorized into the by using information to optimize processes and facili-
following five dimensions. tate machine-to-machine collaboration across different
1) Data Technologies (DT) [5], [6], [12]: Encapsulate the shop floors or locations, adjusting production based on
tools and methodologies needed to facilitate successful resource availability and insights.
data collection and transformation into actionable infor- At a higher level, OT improves control and optimiza-
mation and knowledge to ensure optimal performance in tion of enterprise systems, transforming smart manu-
smart manufacturing. Each stage of the manufacturing facturing from experience-driven to data-driven. This
process generates vast datasets, which can be structured, introduces self-X capabilities, such as self-awareness,
unstructured, or a mix of both, from components, ma- self-prediction, self-configuration, and self-comparison.
chines, or the entire shop floor. Digital twins play a Integrating OT with other technologies enables ac-
crucial role in monitoring, managing, and document- tionable decision-making, optimizing maintenance, and
ing high-speed, real-time manufacturing data. Given the overall manufacturing management.
varied formats and data sources, providing context and 5) Human–Machine Technologies (HMTs) [5], [6], [12]:
establishing standardized protocols for data acquisition The rise of industrial intelligence has significantly
is essential. influenced the interaction between humans and
2) Analytics Technologies (AT) [5], [6], [12]: AT trans- manufacturing systems, with data volumes exceeding
form sensory data into actionable information using human processing capabilities. HMTs are essential

1088 VOLUME 5, 2024

 for integrating production processes with machine numerous machines interact with other machines and even
intelligence, enhancing productivity, and operational workers, the application of AI cannot interfere with previous
efficiency. Technologies such as augmented reality solutions, causing entropy to the system.
(AR) and virtual reality (VR) can assist workers in A variety of challenges rely on data. As a technology heav-
tasks, such as diagnosis and maintenance. HMT will ily dependent on data, it is fundamental that the fed data
transform how workers interact with manufacturing for the algorithms’ training is clean and unbiased, producing
systems, leveraging data, and insights generated by the best quality results for a given task. Also, data avail-
these systems. Workforce training is critical for the ability cannot be assumed. Many industrial scenarios find it
effective use of new technologies, ensuring seamless very costly to produce the necessary data, such as defective
integration and maximizing the potential of HMT. products, to develop reliable AI algorithms [19]. Creating
synthetic data might help with dataset balancing, providing
3) PRINCIPLES a satisfactory dataset for a given problem [20]. In several
Finally, introducing I-AI in industrial systems must fulfill five situations, resorting to techniques, such as transfer learning
basic principles for adapting to industrial scenarios [12]. allows the knowledge of other domains to be transferred into
Systematic approach: A structured architecture that in- another where data might be scarce [21]. Cybersecurity is a
tegrates the associated technologies is necessary. It should key feature in data-dependent systems, with more intercon-
include a standardized interface and system of systems en- nections, as they become more susceptible to cyber threats.
compassing I-AI with industrial systems. New emerging techniques, such as federated learning that pro-
Speed: The velocity at which customer demands change motes decentralization by training models with data locally
should be met as I-AI prioritizes fast systems development, and only transferring the parameter to a central server to be
modeling, validation, and deployment. aggregated [22], [23] may help mitigate such vulnerabilities
Streamline: I-AI must focus on goal-oriented convergent without comprising the model performance [24].
processes, guaranteeing goal certainty, and harmonizing ex- Finally, one big challenge of I-AI is governance. I-AI
isting industrial processes. systems must comply with transparency rules and be under-
Standards: There is a necessity for standardization of data, standable for all stakeholders to create trust and encourage
analysis, formats, and operations in industrial systems, al- adoption. Also, the employed algorithms must be audited and
lowing for easier integration of AI in industrial systems and corrected to prevent biases and misconceptions, which can
between employed technologies. depend on the application scenario.
Sustainability: The integration of AI should ensure that the
III. METHODOLOGY
deployed models are maintained and updated with quality and
consistent performance, including uncertainty management The present literature review relied on the PRISMA method-
ology (see Fig. 1) for analyzing and evaluating scientific
publications under the topic of study. PRISMA1 was chosen
C. DESIGN PRINCIPLES as it is well-suited for systematic reviews and meta-analyzes.
To implement I-AI in manufacturing settings, it is funda- It provides a rigorous framework for conducting literature as-
mental to understand the design principles required for such sessments, ensuring a high methodological rigor and reducing
implementation. As described in Peres et al. [13], and in potential bias in the evidence synthesis.
Ghobakhloo et al. [18], I-AI shares several design principles The search was conducted on 19 June 2023, in three dif-
with I4.0, for instance, decentralization, modularity, real-time ferent digital databases [web of science (WoS), Scopus, and
capacity, and cybersecurity, and privacy preservation. science direct (SDirect)] using the following search string.
The role of human resources in the manufacturing setting Search String = (“Artificial Intelligence” OR “machine
is crucial in most I-AI applications. Despite the necessity of learning” OR “deep learning”) AND (“predictive maintenance”
this role, there is still a barrier between humans and AI. To OR “quality control” OR “personal* manufacturing”) AND
further advance the relationship between both parts, there is (manufacturing OR “industr* 4.0” OR “industr* Intelligence”).
a necessity for a deeper understanding of the reasoning and This research string was adapted to the syntax and specifici-
mechanisms behind AI-based decisions. ties of each database. Criteria for exclusion and inclusion were
Furthermore, adopting standardized practices and fostering set to assess the quality of the articles under consideration,
a data-driven culture will facilitate replication and scaling of namely, the following.
AI applications within CPS. The designs of CPS are reliant on 1) Screening Criteria: Years from 2019 to 2023; Only Q1
modularity, implying that AI modules shall have standardized articles (according to Clarivate); Only articles with im-
interfaces that allow flexible integration with CPS, enhancing pact factor above 9.14 (average of the impact factors of
its capabilities. the 8512 records); titles with the terms “implemented,”
“proposed technique,” “proposed method,” “real data,”
D. CHALLENGES “real environment”; full-text articles.
Despite being very promising for the manufacturing sector,
I-AI faces many challenges. In an industry environment where 1 [Online]. Available: http://www.prisma-statement.org/

VOLUME 5, 2024 1089

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

FIGURE 2. Distribution of the selected publications by research area (after

PRISMA).

FIGURE 3. Distribution of the top 5 journals (after PRISMA).

Furthermore, Fig. 3 illustrates the distribution of publica-

tions among the top five journals. The journal “Computers in
FIGURE 1. Flowchart for document search and selection using PRISMA Industry” leads with ten publications, underlining its signifi-
methodology. cant influence in the field. Following are “Robotics Computer
Integrated Manufacturing” and “IEEE Transactions on Indus-
2) Eligibility Criteria: Duplicated papers; not an AI appli- trial Informatics,” with seven and five publications, respec-
cation; not a real-world application; out of the manufac- tively. The “Journal of Manufacturing Systems” maintains
turing scope. its position with five publications, and, lastly, the “Additive
A total of 9072 records were retrieved from the databases Manufacturing” journal contributes to the list with a singular
mentioned above, with 560 of them identified as duplicate publication.
studies. After removing duplicates, 8512 unique studies re- Several pieces of information from the eligible articles re-
mained. Based on the screening criteria, 8409 records were trieved from the systematic review were analyzed to answer
removed in the screening phase. Moving to the eligibility the proposed research questions.
phase, a statistical analysis of the impact factor was made From the assessment of Table 1 present in Appendix A, it
on the remaining 103 studies. The results revealed a median is possible to evaluate the employed technologies, TRLs, and
impact factor of 10.96 with a standard deviation of 1.36, a inclusion of humans in the system. Also, an analysis of AI’s
25th percentile of 10.10, and a 75th percentile of 11.25. In application in each scenario can be made.
this phase, 31 articles were excluded as they did not address
an AI application, 29 were not a manufacturing environment A. EMPLOYED TECHNOLOGIES
application, and 14 were out of the manufacturing setting. In Regarding the technologies employed in the studied literature,
the end, 29 articles were selected to conduct a more in-depth there is a clear highlight in the usage of ML and deep learning
analysis. (DL) technologies and frameworks, as well as technologies
related to computer vision. The technologies and frameworks
IV. CHARACTERIZATION OF THE IMPLEMENTATIONS OF of machine learning were studied, resulting in the assessment
I-AI in Fig. 4.
Fig. 2 shows the distribution of selected publications across As it is possible to assess, many studies do not disclose
various research areas. Engineering stands out with 28 publi- the used technologies, making it challenging to replicate the
cations, followed by computer science with 22 publications. proposed implementations. However, it is also possible to
The study also includes seven publications in robotics, six verify that the most used technologies in the studies presented
publications in automation control systems, and five publi- in Table 1 are Tensorflow, Pytorch, and OpenCV, which is
cations in operations research management science, although expected as they are the most used frameworks for DL and
these appear less frequently. Materials science is represented computer vision.
by one publication. This distribution highlights the reach of These technologies can kickstart an overview of the en-
AI and ML in diverse fields of scientific research in the scope abling technologies, building an ego graph (see Fig. 5)
of manufacturing. based on those terms. Tensorflow [25] and Pytorch [26] are

1090 VOLUME 5, 2024

 of AI applications, a common gap among most manufactur-
ers [5]. However, other possibilities might come into question,
such as the difficulty of testing specific experiments in a real
environment, as many of these experiments might require
interruptions in production. It is possible to observe that the
solutions with a higher TRL are built around a framework,
though not following any guidelines or standardization, the
care for a system for data collection treatment and AI applica-
tion exists. It is also possible to verify that the application of
AI for task automation is predominant in higher TRLs, which
makes sense since its application relies more on the intended
environment and considers all the variables found in such an
environment. On the other hand, it is possible to verify that
anomaly detection (AD) and quality control scenarios have
FIGURE 4. ML related technologies mostly used in the studied papers. broader TRLs, which may indicate the capability of these
scenarios being tested in a lab environment and real-world
environments. However, the found studies reflect the lack of
two well-known open-source platforms for building machine a “guideline/framework” for I-ML applications as their TRLs
learning models focusing on DL. Both platforms support are mainly affected more by their application scenario instead
GPU acceleration through compute unified device architec- of their use case.
ture (CUDA) [27] drivers. While Tensorflow and Pytorch
provide similar functionalities, some differences in usability C. HUMAN-IN-THE-LOOP
can be found. Keras [28] is a Python library commonly used as Human-in-the-loop refers to integrating humans and hu-
an interface for TensorFlow in Python. Through the years, this man actions in the management and control of production
interface has evolved to be one of the most used for artificial processes. With the introduction of enabling information and
neural networks. So far, Keras has a modular backend support- communication technologies, the main goal of taking into
ing TensorFlow, JAX [29], and PyTorch. Another develop- account human-in-the-loop and human-in-the-loop technolo-
ment tool focused on machine learning is the Scikit-learn [30] gies is to not only rely on the contributions that humans can
library for Python. This library provides functionalities of provide to manufacturing systems but also the augmentation
supervised and unsupervised learning, ready to implement. of human capabilities in the manufacturing site [35].
Regarding computer vision, OpenCV [31] is an open- The overall percentage of approaches that contemplate hu-
source library that provides an infrastructure for computer man work to aid the task in question or even utilize I-AI to
vision applications, ranging from image processing to includ- aid human decision-making is meager (see Table 1). Of the
ing ML algorithms for computer vision. On the subtopic of DL 27 studies, only seven considered having humans in the loop.
for computer vision, YOLO [32] architecture is highlighted. Most studies were conducted in a lab environment without
YOLO is an object detection algorithm based on convolutional considering how workers might affect the system. Although
neural network (CNN). Currently, in its eighth version, the QI scenarios might not require as much assistance from hu-
capabilities of YOLO extend from object detection, classifi- man experts in the deployment phase, other procedures, such
cation, segmentation, and tracking to pose estimation. as PdM I-AI, can aid human action and vice versa.

D. APPLICATION SCENARIOS
B. TECHNOLOGY READINESS LEVEL
TRL is a measurement system that verifies the level of ma- The analysis of the papers presented in Table 1 highlighted
turity of a given technology on a scale from one to nine, five application scenarios [PdM, AD, production planning and
where one is a basic scientific principle, and nine is an actual scheduling (PPS), QI, and human–robot interaction (HRI)]
system that was proven in an operational environment. The where AI is applied in industrial settings. In this section, a
TRL system makes it consistent and systematic to compare characterization per application scenario is made.
the maturity of different types of technology [33].
From the assessment of Table 1, it is possible to identify 1) PREDICTIVE MAINTENANCE
that the TRLs are in their majority in TRLs up to 5 (2/3) (pro- PdM leverages the power of statistics and AI to predict
totyping stage). It was verified that most research was mainly trends, behavior patterns, and correlations to improve the
done in a laboratory environment, with very few cases being process of decision-making in maintenance activities. The
validated in a real-world environment, which is the common main objectives of PdM are decreasing machine downtime,
practice in the research of AI in manufacturing [34]. This early cost reduction, and increasing control, quality, and produc-
stage of readiness might be due to the lack of a structured tion [36]. Some authors consider PdM part of prognostics and
framework for implementation and guidelines for the lifecycle health management [37]. From the analyzed articles, it was

VOLUME 5, 2024 1091

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

FIGURE 5. Ego graph of ML technologies landscape based on Google’s autocomplete queries originating from the search words “tensorflow,” “pytorch,”
and “opencv,” the most common popular ML technologies found in the literature review. Search Depth = 3.

possible to verify that the main focus of ML in PdM is to 3) PRODUCTION PLANNING AND SCHEDULING
test the remaining useful life of a particular machine in lab The I4.0 phenomenon gave the ability to retrieve large
environments. Although some cases considered a complete quantities of data regarding product analysis and production
manufacturing scenario [38], [39], most only consider an iso- measurements, providing ML with the capability to enhance
lated machine. production [46]. PPSs are the processes that establish the
In PdM, the most common types of ML found were un- production rates needed to fulfill the commercial goals. These
supervised ML and DL. Several authors used custom DL processes bring efficiency to the supply chain and have a direct
algorithms to tackle specific problems, being the most com- impact on strategic-level decision-making [47]. Furthermore,
monly used CNNs. DL algorithms’ usage presupposes larger it includes monitoring, managing, and scheduling the produc-
data quantities corresponding to the high sensor data usage. tion processes, allowing real-time coordination of resources
and product customization [48].
In PPS, the algorithms used range from DL and supervised
machine learning (SML) to reinforcement learning (RL) ar-
2) ANOMALY DETECTION eas. Due to being a field where the objectives are enhanced
AD is the process of detecting instances in data that deviate production or even reduced unnecessary downtimes, it is a
significantly from the majority of data instances [40]. In man- good scenario for exploratory methods of ML. However, cus-
ufacturing, this technique has increased in popularity [41], tom DL algorithms were found, and support vector machines
[42], where anomalies often signify product defects. Contrary (SVMs) were also present in PPS studies.
to PdM, AD in the literature is often related to the use of
semisupervised learning and unsupervised learning, such as 4) QUALITY INSPECTION
clustering to detect new and/or anomalous patterns in the data, QI is mainly derived from the inability of human operators
which can mean either faults and defects or simply different to keep up with inspection tasks in today’s manufacturing
modes of operation [43]. Thus, AD may be a Predictive Main- processes. Automation through machine learning made the
tenance (PdM) enabler. process of defect detection easier. Although there are several
Preventing or detecting failures earlier allows industries manners of automated quality control, such as sensor-based
to reduce costs and conduct maintenance earlier [44]. The systems, magnetic particle inspection, and ultrasound testing,
analyzed articles confirm that AD is a rising subject, being image-based systems are one of the most popular as they are
one of the most frequently studied, from early TRLs where one of the easiest to implement with a lower cost [49].
the experiments are in a lab environment to more advanced It is possible to characterize QI on the following attributes:
TRLs where AD was applied in an industrial scenario in an type, whether in real-time or after the manufacturing process
assembly line. is completed and strategy, if the inspection is done in all
In AD also, DL surges as the major ML technique used, processes or certain parts of the process [50].
with CNNs being the most common types of algorithm [45]. In ML, a large variety of machine learning algorithms and
In this scenario, a large variety of algorithms is used as data subfields were found. SML, unsupervised machine learning
from diverse sensors and cameras, making this scenario rich (UML), and DL algorithms can be found, being DL algo-
in conditions for ML experimentation. rithms mostly used in applications with image.

1092 VOLUME 5, 2024

5) HUMAN–ROBOT INTERACTION principles. However, the application of these principles is still
In the era of I4.0, the role of AI and ML extends beyond au- shallow, falling behind half of the chosen articles. The most
tomation and optimization. For instance, collaborative robots unconsidered design principles were, in decreasing order, de-
(or cobots) are AI-powered systems designed to work along- centralization, continuous engineering, and cybersecurity. The
side human operators. This synergy facilitates agile and astute need for more consideration for continuous engineering is
maintenance operations carried out by human personnel and aligned with the absence of the sustainability principle as
imbues cobots with the capacity for dependable and repetitive both consider maintenance and continuity of the employed
maintenance maneuvers [51]. Moreover, AR and VR appli- technologies.
cations provide immersive training environments, allowing When considering cybersecurity as a missing design prin-
workers to practice tasks in a virtual setting before executing ciple, it is necessary to consider the work’s scope carefully.
them in the real world. Most of the researched articles are methodologies or early-
The HRI scenario intensely relied on image data, thus, the stage frameworks where the focus is primarily on the vali-
chosen DL algorithms. However, it is possible to verify that dation of the ML model or the execution of the framework.
exploratory deep reinforcement learning (DRL) algorithms However, there is a growing concern about cybersecurity, and
were also used to improve the quality of the overall solution. in I-ML, most data regarding industrial processes may require
certain levels of privacy and security [60]. The principle of
decentralization goes hand in hand with the principle of cy-
V. CHALLENGES AND FUTURE TRENDS bersecurity, as security must be a principal component of a
The advances in the application of ML in industrial scenarios decentralized framework.
have been evolving in recent years, mainly due to the advance- Considering the challenges of I-ML mentioned in Sec-
ments in enabling technologies, such as IoT, cloud computing, tion II-D, such as data cleanliness, availability, quantity,
and ML itself. However, several challenges must be overcome cybersecurity, and governance, it was found that the reviewed
to attain the full potential in I-ML. In this section, a character- papers mostly face these problems or do not even consider
ization of the main challenges and future trends, withdrawn them. This lack of consideration could be from an unaware-
from the researched literature and complemented with ad- ness of such challenges or a deliberate choice considering the
ditional readings, is presented, as well as a more detailed scope of each work. However, these challenges still exist, and
characterization for each application scenario to serve as a although most studied papers do not consider them evident,
guideline for future research. they are present.
Upon analyzing the state of the art of machine learning In Fig. 6, it is possible to see that the challenges and future
applications in manufacturing (see Table 1), it is necessary to trends are divided into four essential categories: infrastructure,
evaluate it from an I-ML perspective. The application of I-ML software, data, and governance. However, it is also possible to
considers its design principles and fundamental principles. infer that different priorities exist based on the coverage of
It carries several challenges mentioned in Section II-D. To those applications. Focusing on the challenges, infrastructure
characterize the studied applications, it is necessary to verify is considered of high priority with 68% of the challenges
if they comply with the design principles of I-ML and if the found, followed by data with 25% considered a medium pri-
challenges found are aligned with the ones stated by I-ML ority, and software and governance considered a low priority
taxonomy. Finally, confirming the primary discrepancies con- with 8% and 4%, respectively.
cerning the I-ML concept is also necessary. The rest of this section will cover a more in-depth overview
The analysis showed that the fundamental principles of of the found challenges and future trends in the application of
I-ML are generally unmet. Although some articles propose I-ML in specific application scenarios that can be verified in
methodologies and frameworks that comply with existing Fig. 6.
industrial processes [39], [52], [53], [54], fulfilling the stream-
line principle, the majority are not compliant. The most unmet A. PREDICTIVE MAINTENANCE
principle is sustainability. None of the articles refers to model As PdM has as its main goal to model and predict failure
maintenance or as a methodology for model updates to keep patterns in machines and processes [61], it is still possible
quality and consistent performance. The second most unmet to identify key challenges in its implementation. The PdM
principle is the standards principle, as most of the studied arti- process is still isolated and not integrated with other manufac-
cles do not contemplate industrial standards nor data analytics turing processes. This shows a lack of integration that removes
standards on the used data. context from the PdM process, not allowing it to adapt to new
Regarding the design principles, the most commonly used variants and processes within the manufacturing process [62].
are real time, context awareness, and integration. Although Major challenges in applying ML algorithms in the PdM
most papers do not consider these principles, the ones that procedure are still related to big data and ML models. Re-
do mainly have an industrial context in mind, thus consid- garding big data, there is still a necessity for a distributed
ering I4.0 principles [51], [52], [54], [55], [56], [57], [58], computing scenario. The coexistence of edge and cloud com-
[59]. The most common principles fall on a common ground puting is fundamental for reducing the latency and velocity
between I4.0 and I-ML, where both concepts share these of predictions in real-time scenarios. Concerning the ML

VOLUME 5, 2024 1093

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

FIGURE 6. Challenges and future trends identified for five application scenarios regarding the implementation of I-ML in manufacturing scenarios.

models used, verifying the lack or excess of heterogeneity is ML model training and application in an industrial scenario
still necessary as it deeply influences the model performance with the required instrumentation, software, and specialized
[63]. personnel [67]. Also, the usage of federated learning can be
Another challenge that may relate to the isolation of the a possible advantage for creating models collaboratively pre-
PdM process is its incompatibility with current industrial stan- serving the privacy and security of data. The missing data can
dards [39], [64], [65]. Adapting the current methodologies to be complemented by generating synthetic data from genera-
the used standards would allow for more straightforward in- tive adversarial networks (GANs) or even in simulation [20],
tegration and further interoperability between state-of-the-art [68]. Nowadays, game design software presents the capabil-
approaches to PdM and industrial settings. ities for creating photorealistic simulations for diverse envi-
When concerning data for PdM, it is also necessary to ronments, which can produce a large amount of synthetic data
verify the datasets’ dimensions [55]. Many PdM techniques that can be used in the training of ML algorithms. However,
rely on DL, which is heavily dependent on data. However, synthetic data generation still faces several challenges as well.
not only the quantity matters but also the relation between For instance, in sim2real technologies, no simulation can per-
classes in labeled data to have a sufficient amount for training; fectly replicate reality, leading to AI exploitations in simulated
however, a failure is not a common or desirable event, creating scenarios [69]. To achieve photorealism, object characteristics
an imbalanced dataset in most cases [66]. need to be replicated with maximum detail, and the character-
Another major challenge in applying PdM is the scarcity of istics of the generated scene also need to be congruent with the
resources to create a PdM infrastructure from data collection, scenario being simulated. Despite these challenges, several

1094 VOLUME 5, 2024

authors have succeeded in the creation of photorealistic data Manufacturing environments using AD also face
through the usage of game engines [70], [71]. some challenges. It is necessary to involve experts for
Digital twins can be a great advantage in deploying models. reinspection and feedback on new events or potential
To fully deploy a model, testing it first in a secure environment misclassifications [39], [77]. In addition, sometimes, the
that does not tamper with the real application environment cause for less data is limited sensor use in AD systems [80].
is essential [72]. Digital twins can present a possibility for Finally, testing ML models in real-world settings is crucial
testing new ML models and approaches in PdM. for adapting and enhancing their performance.
It is also essential to consider the role of humans in PdM Upon analyzing the challenges found in AD, possible
in future research. The ML algorithms are limited to the data research trends can be encountered. Similarly to PdM, a
used for training, which leads to two major research paths. possibility to reduce data scarcity and class defect may
The inclusion of humans in the process of PdM, either in data pass through collaborative ML techniques, such as federated
analysis or even in the detection of maintenance scenarios, learning. Also, in the sequence of the lack of explain-
failed by the ML algorithms and the maintenance of the ML ability/interpretability to attain better insights from detected
infrastructure. Organizations that rely on ML models need to anomalies and better include human expertise in the detection
maintain a solid infrastructure. In order to maintain a reliable loop, the usage of explainable AI is essential as a research
and efficient ML infrastructure, various methodologies have trend in this area. Techniques such as transparency by design
been studied over the years. One of the most prominent among and post hoc explainability methodologies can be employed
them is MLOps [73]. MLOps is a software development in the continuous development and integration of ML mod-
paradigm that is derived from the principles of development els [81]. Furthermore, a research venue that is interesting
and operations (DevOps) [74]. It involves using cloud tech- to explore is self-diagnosis. Although sensorization is very
nologies and integrating different software and middleware to present in manufacturing, collecting large quantities of AD
perform data intake, transfer, and modification. MLOps also data, there are still very few studies on self-diagnosis and
enables ongoing ML model retraining and continuous model self-detection of sensor status to guarantee data quality [82].
redeployment, making it an ideal choice for organizations The price for deployment and maintenance of AD systems
that require continuous integration and development of ML is still very high, thus being inaccessible to many manufactur-
models [75]. ers to create an infrastructure capable of fitting their needs.
The last future trend considered worth exploring is PdM The utilization of infrastructure and software services may
as a service. As mentioned before, the costs to implement an help reduce costs to manufacturers, allowing the collection of
end-to-end PdM infrastructure are very high, which is a sig- data that further improves AD models and technologies.
nificant setback to several manufacturers. A possibility relies Another future trend in AD is the implementation of tech-
on acquiring PdM services and expertise to reduce costs for nologies for continuous development and integration of ML
the manufacturer. models. With techniques such as MLOps, ML models used in
AD can be maintained and improved over time, lasting longer,
B. ANOMALY DETECTION and reducing manufacturing costs while performing with sim-
Although it has been a research subject since the 1960s [40], ilar quality. Finally, the introduction of ML on edge will help
AD still presents itself as a growing field. The introduction decrease latency in the usage of ML for AD, providing faster
of enabling technologies, including machine learning, opens a classifications and fitting the velocity of today’s factories.
variety of challenges, and opportunities worth exploring.
The challenges found in the application of ML in AD are
presented in several focus points. The lack of explainability to C. PRODUCTION PLANNING AND SCHEDULING
anomaly classification is still a challenge for manufacturers as The research in the ML applications for PPS revealed a
it presents a large barrier to the adoption of AD techniques. As scenario with several challenges and, consequently, some
of today, explainable AD has become a regulatory and ethical research trends. The first challenge found was the central-
requirement in its applications [76]. ization of decisions and schedules without consideration for
In addition, there is still a challenge in classifying anoma- a distributed approach with CPSs or even agent-based sys-
lies in noisy data, presenting a problem in data quality [77]. tems [52]. However, not only are distribution challenges found
Also, detecting complex or even small anomalies still presents in PPS, but the need for definitive and practical solutions for
a significant challenge to the application of ML in this addressing flexibility, agility, and reactiveness requirements
field [40]. In addition, subsequent dimensionality shifts in data still present a barrier for the PPS when new products or urgent
and parameter occlusion, for instance, due to sensor failure, changes appear. Concerning data produced for PPS, a substan-
also contribute to failed classifications [78]. tial amount is being received; it is necessary to have a capable
Not only in data relies on the challenges of AD, but in man- infrastructure to handle such large quantities of data. The data
ufacturing environments, approaches often fail to adjust their problems are not only due to data quantity, schedules, and
configurations over time, allied to the lack of self-diagnosis dispatching rules generated by that data, which are still not
and failure of overtime improvement, leading to outdated adjusted in real time, contributing to a slower adaptation of
models very quickly and a reduction of effectiveness [79]. the manufacturing system to changes [83].

VOLUME 5, 2024 1095

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

When concerning the manufacturing infrastructure, two First, it is necessary to research the continuous adaptation
major challenges are highlighted: the lack of self-awareness of ML algorithms. For that, technologies such as MLOPs
of the machines and the factory as a whole to seamlessly should be considered to continuously adapt and maintain
adapt to changes and provide self-diagnosis and a necessity the ML models and algorithms used in QI. However, it is
to employ systems and methodologies that provide scalability necessary to create mechanisms for ML model selection, con-
to the manufacturing system locally and globally [84]. sidering the speed of change in a manufacturing environment;
The following research trends are proposed to solve the ML model training might not adapt to such fast changes.
challenges found in PPS. Modular manufacturing systems It is necessary to create highly adaptable infrastructures
should be developed to easily rearrange the whole system as where ML models can be used and adapted to high rates of
necessary when, for instance, a new product arrives or even change.
the manufacturing speed needs to be altered. To attain modu- Also, deploying ML models cannot be done directly in
larity, it is essential to implement sensing systems adaptable real-world environments due to possible downtimes caused
to different types of machines, even legacy ones that can re- by unadapted models. It is necessary to implement virtual
trieve reliable data. Also, creating data models would provide environments such as digital twins to deploy models in safe
easy access and better interoperability between devices and environments before adapting to the real world [89].
processes, allowing a seeminglyless communication system. Another future research direction is the inclusion of humans
Not only is interoperability important when dealing with data in the QI process. Expert feedback is necessary to further
for PPSs, but the velocity of how data and information are improve QI in highly changeable environments.
processed and edge computing are possibilities. Also, when
concerning edge computing, the capabilities of self-awareness
and self-diagnosis should be considered to be computed on E. HUMAN–ROBOT INTERACTION
edge to mitigate data dimensionality change due to occluded The challenges in the HRI field are vast. In manufacturing
parameters. environments, collaboration between humans and machines
is still limited, and even HRI is still in the early stages of
research. As such, several challenges can be found. When col-
D. QUALITY INSPECTION laborating with a robotic entity, there is still much distrust by
QI is a major application scenario when concerning manu- the human part about the robot’s behavior; the lack of compre-
facturing. Although proven very fruitful, applying ML into hension may be due to the fear of injuries or even unexpected
QI has brought several challenges and future possibilities to behaviors [90]. Also, the naturality with which robotic sys-
explore. tems interact with humans is still very low, intensifying the
From the literature reviewed, it is possible to draw a few resistance in the human–robot synergy. From the machine
challenges. The necessity to evaluate further ML models in the side, it is still challenging to interpret human behavior and
context of QI becomes evident as it is crucial to develop this intentions, making it hard to act accordingly. This fact may be
scenario’s quality further [53]. Also, there is a necessity for a due to the lack of real-world datasets where human behavior
virtual environment where further QI algorithms and mecha- and more minor interactions between humans and robots are
nisms can be designed and tested before entering a real-world studied. There is also a limitation regarding environmental
scenario [85]. The necessity for a real-time component was isolation. In HRI, there is a need for an “open-world” scenario
verified [86]. In several manufacturing scenarios, the speed where out-of-context inputs might occur. When considering
of the process vastly surpasses the capacity of QI processes, ML specifically for HRI, two significant challenges are found:
causing difficulties in obtaining quality results. a necessity for systems not to be fully controlled by ML and
The inadaptability of current QI processes to product a lack of robustness of current ML models regarding environ-
changes becomes evident in the current literature, limiting mental noises [91].
the application of these processes in real manufacturing en- Concerning the challenges mentioned above, some research
vironments. To improve the speed of the QI process, it is also trends have arisen as follows.
essential to consider the automation of data transfer processes 1) There should be an emphasis on researching robot sen-
either in-house or between edge and cloud, attaining a better sitization to collect more data about the surrounding
processing speed and, consequently, faster QI results [87]. environment and the humans working with them. An-
In addition to the difficulties in adapting to new products, other research path would be a photorealistic simulation
QI systems should also acquire the capacity to detect flaws for dataset generation and workforce training.
and defects in newer products and adapt to new standards 2) In order to improve trustworthiness in HMI, it is
and demands over time [88]. However, the human role in the necessary to train workers in high-fidelity simulation
interaction between humans and CPSs, which are key in QI, environments so that there is more knowledge and trust
still needs to be determined. Further research on the role of the in automated environments and machines.
human is still necessary to improve further QI mechanisms. 3) A standard physical/visual language for humans and
Several trends in the field are described below to cope with robots can be the target of research to facilitate human–
the challenges presented above. machine communication.

1096 VOLUME 5, 2024

VI. DISCUSSION HRI focuses on collaborative robots and AI-powered
This section will provide an answer to the three guiding systems working alongside humans, utilizing DL and
research questions, summarizing the findings of this re- exploratory DRL algorithms to enhance overall solutions,
view, and offering a summary of possible future research particularly in immersive training environments using AR
venues. and VR applications.
To address the first RQ (How applications of ML in man- The analysis of Table 1 also proved that although lack of
ufacturing can be characterized in the context of I-ML?), an data is still a significant setback in the application of DL,
extensive literature review was conducted. the advantages of applying DL to different use cases are
The literature review on the application of ML and AI clear and significant (82% of the studied articles applied DL
in manufacturing reveals a predominant use of technologies models), which further reinforces the need to support digital
related to ML and DL, particularly TensorFlow, Pytorch, transformation and facilitate the adoption of I-AI practices in
and OpenCV. These frameworks, widely acknowledged in a systematic and verifiable way.
the field of DL and computer vision, play a pivotal role in Regarding RQ2 (What challenges and limitations are com-
implementing ML models. However, the need for disclosure monly identified in the implementation of I-ML?), through an
of technologies used in some studies poses challenges for analysis of the challenges and limitations of the research liter-
replication. Tensorflow and Pytorch, both supporting GPU ature, it was possible to draw a baseline to research further the
acceleration, are instrumental, while Keras serves as a pop- current challenges in the implementation of I-ML in several
ular interface for TensorFlow. Scikit-learn is another notable manufacturing scenarios.
Python library for ML, emphasizing supervised and unsuper- Although the necessity of implementation in a real-world
vised learning. environment is transversal to every application scenario to
In terms of TRL, the majority of assessed applications are mitigate unexpected challenges, many challenges are more
in TRL up to 5 (prototyping stage). This early-stage readiness focused on each application scenario.
may be attributed to the need for a structured framework and In the domain of PdM, challenges include the isolation of
guidelines for AI application lifecycle. Notably, applications PdM processes, the lack of integration with other manufac-
with higher TRL are often built around frameworks, show- turing processes, issues related to big data and ML models,
casing a degree of systematization. Task automation using AI and incompatibility with industrial standards [39], [55], [92].
tends to dominate higher TRL, reflecting a reliance on the AD presents challenges, such as explainability, classification
intended environment’s variables. On the contrary, AD and in noisy data, and detecting complex anomalies [77]. Ad-
quality control scenarios exhibit broader TRLs, suggesting dressing these challenges can be facilitated by collaborative
testing in both lab and real-world environments. The absence ML techniques, explainable AI, self-diagnosis, and infras-
of a standardized framework for I-ML applications manifests tructure and software services. In PPS, challenges include
in TRLs being more influenced by the application scenario centralized decision-making, flexibility, agility, and data han-
rather than the use case. dling issues [52]. Proposed research trends involve modular
Regarding human involvement, the percentage of ap- manufacturing systems, sensing systems, data models, inter-
proaches incorporating human input or using I-AI for operability, edge computing, and self-awareness for machines
decision-making is limited. Among 27 studies, only seven and factories. QI faces challenges related to evaluating ML
consider human involvement, with many studies conducted models, adapting to product changes, and real-time processing
in lab environments without addressing the impact of human speed [85], [86]. HRI encounters challenges in human distrust,
interaction on the system. While QI scenarios may require robotic comprehension of human behavior, and environmental
less human intervention, procedures like PdM I-AI can benefit isolation. Overall, these challenges and trends provide valu-
from human collaboration, emphasizing the need for more able insights for guiding future research in the application of
consideration of human factors in the deployment phase. ML in industrial contexts.
The reviewed literature also identified various application In this way, it was possible to address the third RQ (How
scenarios of ML in manufacturing. can future I-ML systems be designed to guarantee industrial
In PdM, ML, particularly UML and DL with a focus on requirements?)
CNNs, is primarily used to assess the remaining useful life of The assessment of future trends and research venues re-
machines. vealed that future trends in the application of I-ML in the
AD involves detecting deviations in data, and in manu- studied application scenarios are divided into three main cat-
facturing, it is crucial for identifying product defects. DL, egories: infrastructure, software, data, and governance.
especially CNN, dominates in this scenario, where diverse When concerning PdM, future trends involve exploring
sensor and camera data are utilized. transfer learning, federated learning, synthetic data genera-
PPS benefits from algorithms ranging from DL and SML to tion, digital twins, and the role of humans in PdM.
RL, enhancing production rates and minimizing downtimes. The high deployment costs of AD systems could be
QI leverages ML (including SML, UML, and DL al- mitigated by utilizing infrastructure and software services,
gorithms) for automated defect detection in manufacturing while continuous development and integration of ML models
processes. through MLOps can enhance long-term efficiency.

VOLUME 5, 2024 1097

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

In QI, continuous adaptation of ML algorithms through to consider journals with slightly lower impact factors yet
MLOps, virtual environments for development and testing, still significant insights would enhance the inclusivity of the
and human inclusion in the QI process are suggested trends. review.
Finally, research trends in the scenario of HRI include
sensitizing robots, photorealistic simulations for dataset gen-
eration, worker training in simulation environments, and VIII. CONCLUSION
establishing a standard physical/visual language for human– In this article, a systematic literature review following the
robot communication. PRISMA guidelines was conducted in journal publications
Although many future trends were identified in specific ap- of WoS, Scopus, and SDirect databases to characterize the
plication scenarios, transversal challenges, and future trends current research landscape in I-ML applications in manu-
are worth mentioning. For instance, it is essential to highlight facturing. From the chosen literature, the applications were
the need for a framework or guidelines for implementing I- characterized, considering the employed technologies, TRL,
ML. This gap makes it very difficult to replicate and further presence of humans in the loop and application scenarios. An
improve current applications of ML, making the transfer of analysis of the current challenges of the application scenarios
technology from research environments to production envi- was conducted in order to determine which research directions
ronments a challenging task. It is also relevant to mention are viable to pursue in order to improve further the field of I-
that many manufacturers may naturally show some adversity ML. The main challenges found were categorized in four main
to adopting I-ML without guidelines or frameworks proven subjects transversal to the five studied application scenarios.
to be an advantage in real scenarios, imposing a signifi- Focusing on the applications of machine learning in industrial
cant barrier to the evolution of I-ML solutions. Furthermore, scenarios it was found that model’s lifecycle and security is
legacy systems threaten the introduction of I-ML as many often neglected. Also, transparency and explainability need
manufacturers, especially small and medium enterprises, have further research in order to improve understandability among
difficulties supporting the financial and human resource in- manufacturers. Furthermore, the existence of legacy systems
vestments required to upgrade their traditional manufacturing pose a challenge regarding small-medium enterprises invest-
sites [93]. In such scenarios, digital retrofitting with IoT de- ment costs of infrastructure digitization leaving behind lower
vices can provide a more economical approach to improve financial power manufacturers.
current infrastructure [94]. Toward future research, it is recommended to delve into
the application of ML in real industrial scenarios, tackling
VII. LIMITATIONS four essential key points: data collection frameworks, edge
The limitations of this systematic literature review primarily computing devices, newer methodologies of serving the ML
stem from the predefined inclusion criteria outlined in the functionalities in manufacturing scenarios, and frameworks
PRISMA methodology. The first limitation lies in its exclusive for continuous development and integration of ML. From
focus on Q1 articles, representing high-impact journals in the insights gathered in Fig. 6, it is possible to identify that
the realm of I-AI. While this emphasis on reputable sources MLOps should be one of the main focuses as it can be the
ensures a quality baseline, it introduces a potential constraint most effective method to be implemented with visible results
by excluding valuable perspectives, innovative findings, and in the continuous development and integration of ML models
diverse methodologies present in journals with lower impact in manufacturing environments. Digital twins can also prove
factors. Consequently, the review might present a somewhat to be very effective in simulating and testing ML models
skewed representation of the I-AI landscape, limiting its abil- and scenarios before production deployment. The continuous
ity to encompass emerging trends, practical implications, and research in the field of digital twins proves that its usage
context-specific insights found in journals with specific sub- brings numerous advantages in the manufacturing field. In
domains or interdisciplinary perspectives. Moreover, applying addition, synthetic data and edge ML are also proven to be
a high impact factor inclusion criterion (above 9.4) was in- very effective in manufacturing scenarios, improving ML im-
tended to prioritize articles from highly regarded journals. plementation and data gathering.
However, this criterion can affect the study’s representative- In contrast, although proven to be effective, technologies
ness. The strict adherence to a high impact factor threshold and methods such as photorealistic simulation are difficult
may inadvertently overlook contributions from journals with to attain sustainably. Explainable AI is also a method that,
slightly lower impact factors that still offer substantial insights despite its necessity being irrefutable, is under constant devel-
into the multifaceted landscape of I-AI. This raises questions opment, and it is especially difficult to attain full transparency
about the completeness and diversity of the literature con- for complex models. Furthermore, the creation of services
sidered in the review, emphasizing the need for a nuanced surrounding the scenarios presented in Fig. 6 (as, for instance,
approach in the selection criteria to ensure a comprehensive PdM as a service) involves the creation of newer business
exploration of the field. Following studies should have selec- models, a significant initial investment that imposes a barrier
tion criteria beyond Q1 journals to encompass a more diverse to its adoption. Even though with higher difficulty, it is advan-
array of publications, ensuring a better understanding of I-AI. tageous to implement these techniques as they were proven to
In addition, adjusting the impact factor threshold for inclusion enhance the application of I-ML in manufacturing scenarios.

1098 VOLUME 5, 2024

APPENDIX A
OVERVIEW OF I-ML APPLICATIONS
TABLE 1. Overview of Eligible Applications of I-ML in Manufacturing Environments

VOLUME 5, 2024 1099

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

TABLE 1. (Continued)

REFERENCES [17] J. Leng et al., “Unlocking the power of industrial artificial intelligence
[1] K. Zhou, T. Liu, and L. Zhou, “Industry 4.0: Towards future industrial towards industry 5.0: Insights, pathways, and challenges,” J. Manuf.
opportunities and challenges,” in Proc. IEEE 12th Int. Conf. Fuzzy Syst., vol. 73, pp. 349–363, 2024, doi: 10.1016/j.jmsy.2024.02.010.
Syst. Knowl. Discov., 2015, pp. 2147–2152. [Online]. Available: https://www.sciencedirect.com/science/article/
[2] S. Sahoo and C.-Y. Lo, “Smart manufacturing powered by recent pii/S0278612524000335
technological advancements: A review,” J. Manuf. Syst., vol. 64, [18] M. Ghobakhloo, “The future of manufacturing industry: A. strategic
pp. 236–250, Jul. 2022, doi: 10.1016/j.jmsy.2022.06.008. roadmap toward Industry 4.0,” J. Manuf. Technol. Manage., vol. 29,
[3] T. Kotsiopoulos, P. Sarigiannidis, D. Ioannidis, and D. Tzovaras, “Ma- no. 6, pp. 910–936, Jul. 2018, doi: 10.1108/JMTM-02-2018-0057.
chine learning and deep learning in smart manufacturing: The smart [19] L. Pahren, P. Thomas, X. Jia, and J. Lee, “A novel method in intelli-
grid paradigm,” Comput. Sci. Rev., vol. 40, May 2021, Art. no. 100341, gent synthetic data creation for machine learning-based manufacturing
doi: 10.1016/j.cosrev.2020.100341. quality control,” IFAC-PapersOnLine, vol. 55, no. 19, pp. 73–78,
[4] K. Xu et al., “Advanced data collection and analysis in data-driven 2022.
manufacturing process,” Chin. J. Mech. Eng., vol. 33, pp. 1–21, 2020. [20] R. S. Peres, M. Azevedo, S. O. Araújo, M. Guedes, F. Miranda, and
[5] J. Lee, J. Singh, and M. Azamfar, “Industrial artificial intelligence,” J. Barata, “Generative adversarial networks for data augmentation
2019, arXiv:1908.02150. in structural adhesive inspection,” Appl. Sci., vol. 11, no. 7, 2021,
[6] J. Lee, H. Davari, J. Singh, and V. Pandhare, “Industrial artificial intel- Art. no. 3086.
ligence for industry 4.0-based manufacturing systems,” Manuf. Lett., [21] F. Zhuang et al., “A comprehensive survey on transfer learning,” Proc.
vol. 18, pp. 20–23, Oct. 2018, doi: 10.1016/j.mfglet.2018.09.002. IEEE, vol. 109, no. 1, pp. 43–76, Jan. 2021.
[7] J. Lee, Introduction: The Development and Application [22] M. Aledhari, R. Razzak, R. M. Parizi, and F. Saeed, “Federated learn-
of AI Technology. Singapore: Springer, 2020, pp. 1–4, ing: A survey on enabling technologies, protocols, and applications,”
doi: 10.1007/978-981-15-2144-7_1. IEEE Access, vol. 8, pp. 140699–140725, 2020, doi: 10.1109/AC-
[8] G. A. Susto, A. Schirru, S. Pampuri, S. McLoone, and A. Beghi, CESS.2020.3013541.
“Machine learning for predictive maintenance: A multiple classifier [23] L. Li, Y. Fan, M. Tse, and K.-Y. Lin, “A review of applications in fed-
approach,” IEEE Trans. Ind. Informat., vol. 11, no. 3, pp. 812–820, erated learning,” Comput. Ind. Eng., vol. 149, 2020, Art. no. 106854,
Jun. 2015, doi: 10.1109/TII.2014.2349359. doi: 10.1016/j.cie.2020.106854. [Online]. Available: https://www.
[9] Z. Jan et al., “Artificial intelligence for industry 4.0: Sys- sciencedirect.com/science/article/pii/S0360835220305532
tematic review of applications, challenges, and opportunities,” [24] C. Zhang, Y. Xie, H. Bai, B. Yu, W. Li, and Y. Gao, “A survey on fed-
Expert Syst. Appl., vol. 216, Apr. 2023, Art. no. 119456, erated learning,” Knowl.-Based Syst., vol. 216, 2021, Art. no. 106775.
doi: 10.1016/j.eswa.2022.119456. [25] M. Abadi et al., “{TensorFlow}: A. system for {Large-Scale} machine
[10] S. W. Kim, J. H. Kong, S. W. Lee, and S. Lee, “Recent advances of learning,” in Proc. 12th USENIX Symp. Operating Syst. Des. Imple-
artificial intelligence in manufacturing industrial sectors: A review,” mentation, 2016, pp. 265–283.
Int. J. Precis. Eng. Manuf., vol. 23, no. 1, pp. 111–129, Jan. 2022, [26] A. Paszke et al., “Pytorch: An imperative style, high-performance deep
doi: 10.1007/s12541-021-00600-3. learning library,” in Proc. Int. Conf. Adv. Neural Inf. Process. Syst.,
[11] D. Kreuzberger, N. Kühl, and S. Hirschl, “Machine learn- 2019, vol. 32, pp. 8026–8037.
ing operations (MLOps): Overview, definition, and architecture,” [27] J. Nickolls, I. Buck, M. Garland, and K. Skadron, “Scalable parallel
IEEE Access, vol. 11, pp. 31866–31879, 2023, doi: 10.1109/AC- programming with CUDA: Is CUDA the parallel programming model
CESS.2023.3262138. that application developers have been waiting for?,” Queue, vol. 6,
[12] J. Lee, Industrial AI: Applications With Sustainable Performance. Sin- no. 2, pp. 40–53, 2008.
gapore: Springer, 2020. [28] N. Ketkar, “Introduction to Keras,” in Deep Learning With Python: A
[13] R. S. Peres, X. Jia, J. Lee, K. Sun, A. W. Colombo, and J. Barata, Hands-on Introduction. Berlin, Germany: Springer, 2017, pp. 97–111.
“Industrial artificial intelligence in Industry 4.0—systematic review, [29] J. Bradbury et al., “JAX: Composable transformations of python
challenges and outlook,” IEEE Access, vol. 8, pp. 220121–220139, NumPy programs,” 2018. [Online]. Available: http://github.com/
2020, doi: 10.1109/ACCESS.2020.3042874. google/jax
[14] J. Lee, Definition and Meaning of Industrial AI. Singapore: Springer, [30] O. Kramer, “Scikit-learn,” in Machine Learning for Evolution Strate-
2020, pp. 33–61, doi: 10.1007/978-981-15-2144-7_3. gies. Berlin, Germany: Springer, 2016, pp. 45–53.
[15] J. Lee, Killer Applications of Industrial AI. Singapore: Springer, 2020, [31] G. Bradski, “The openCV library,” Dr Dobb’s J.: Softw. Tools Profes-
pp. 63–117, doi: 10.1007/978-981-15-2144-7_4. sional Programmer, vol. 25, no. 11, pp. 120–123, 2000.
[16] X. Zhang, X. Ming, Z. Liu, D. Yin, Z. Chen, and Y. Chang, [32] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look
“A reference framework and overall planning of industrial arti- once: Unified, real-time object detection,” in Proc. IEEE Conf. Com-
ficial intelligence (I-AI) for new application scenarios,” Int. J. put. Vis. Pattern Recognit., 2016, pp. 779–788.
Adv. Manuf. Technol., vol. 101, no. 9, pp. 2367–2389, Apr. 2019, [33] J. C. Mankins, “Technology readiness levels,” National Aeronautics
doi: 10.1007/s00170-018-3106-3. and Space Administration (NASA), White Paper, Apr. 6, 1995.

1100 VOLUME 5, 2024

 [34] J. F. Arinez, Q. Chang, R. X. Gao, C. Xu, and J. Zhang, “Ar- [52] Y. Zhang, H. Zhu, D. Tang, T. Zhou, and Y. Gui, “Dynamic job
tificial intelligence in advanced manufacturing: Current status and shop scheduling based on deep reinforcement learning for multi-agent
future outlook,” J. Manuf. Sci. Eng., vol. 142, no. 11, 2020, manufacturing systems,” Robot. Comput.- Integr. Manuf., vol. 78,
Art. no. 110804. Dec. 2022, Art. no. 102412, doi: 10.1016/j.rcim.2022.102412.
[35] C. Cimini, F. Pirola, R. Pinto, and S. Cavalieri, “A human-in- [53] J. Wang, P. Fu, and R. X. Gao, “Machine vision intelli-
the-loop manufacturing control architecture for the next genera- gence for product defect inspection based on deep learning and
tion of production systems,” J. Manuf. Syst., vol. 54, pp. 258–271, hough transform,” J. Manuf. Syst., vol. 51, pp. 52–60, Apr. 2019,
2020. doi: 10.1016/j.jmsy.2019.03.002.
[36] E. Sezer, D. Romero, F. Guedea, M. Macchi, and C. Emmanouilidis, [54] X. Yang, Z. Zhou, J. H. Sørensen, C. B. Christensen, M. Ünalan,
“An Industry 4.0-Enabled low cost predictive maintenance approach and X. Zhang, “Automation of SME production with a Cobot system
for SMEs,” in Proc. IEEE Int. Conf. Eng., Technol. Innov., Jun. 2018, powered by learning-based vision,” Robot. Comput.- Integr. Manuf.,
pp. 1–8, doi: 10.1109/ICE.2018.8436307. vol. 83, Oct. 2023, Art. no. 102564, doi: 10.1016/j.rcim.2023.102564.
[37] J. Wang, C. Liu, M. Zhu, P. Guo, and Y. Hu, “Sensor data based [55] W. Yu, T. Dillon, F. Mostafa, W. Rahayu, and Y. Liu, “A global
system-level anomaly prediction for smart manufacturing,” in Proc. manufacturing big data ecosystem for fault detection in predictive
IEEE Int. Congr. Big Data, Jul. 2018, pp. 158–165, doi: 10.1109/Big- maintenance,” IEEE Trans. Ind. Informat., vol. 16, no. 1, pp. 183–192,
DataCongress.2018.00028. Jan. 2020, doi: 10.1109/TII.2019.2915846.
[38] M. L. H. Souza, C. A. D. Costa, G. D. O. Ramos, and R. D. R. Righi, [56] L. Leontaris et al., “A blockchain-enabled deep residual architecture
“A feature identification method to explain anomalies in condition for accountable, in-situ quality control in industry 4.0 with mini-
monitoring,” Comput. Ind., vol. 133, Dec. 2021, Art. no. 103528, mal latency,” Comput. Ind., vol. 149, Aug. 2023, Art. no. 103919,
doi: 10.1016/j.compind.2021.103528. doi: 10.1016/j.compind.2023.103919.
[39] M. L. H. Souza, C. A. D. Costa, and G. D. O. Ramos, “A machine- [57] A. Villalonga et al., “A decision-making framework for dy-
learning based data-oriented pipeline for prognosis and health man- namic scheduling of cyber-physical production systems based on
agement systems,” Comput. Ind., vol. 148, Jun. 2023, Art. no. 103903, digital twins,” Annu. Rev. Control, vol. 51, pp. 357–373, 2021,
doi: 10.1016/j.compind.2023.103903. doi: 10.1016/j.arcontrol.2021.04.008.
[40] G. Pang, C. Shen, L. Cao, and A. van den Hengel, “Deep learning [58] S. D’Avella, C. A. Avizzano, and P. Tripicchio, “ROS-Industrial based
for anomaly detection: A review,” ACM Comput. Surv., vol. 54, no. 2, robotic cell for industry 4.0: Eye-in-hand stereo camera and visual
pp. 1–38, Mar. 2022, doi: 10.1145/3439950. servoing for flexible, fast, and accurate picking and hooking in the
[41] J. Liu et al., “Anomaly detection in manufacturing systems using production line,” Robot. Comput.- Integr. Manuf., vol. 80, Apr. 2023,
structured neural networks,” in Proc. 13th World Congr. Intell. Control Art. no. 102453, doi: 10.1016/j.rcim.2022.102453.
Autom., Jul. 2018, pp. 175–180, doi: 10.1109/WCICA.2018.8630692. [59] J. Hu, G. Zhao, W. Xiao, and R. Li, “AR-based deep learn-
[42] F. Lopez et al., “Categorization of anomalies in smart manufac- ing for real-time inspection of cable brackets in aircraft,” Robot.
turing systems to support the selection of detection mechanisms,” Comput.- Integr. Manuf., vol. 83, Oct. 2023, Art. no. 102574,
IEEE Robot. Autom. Lett., vol. 2, no. 4, pp. 1885–1892, Oct. 2017, doi: 10.1016/j.rcim.2023.102574.
doi: 10.1109/LRA.2017.2714135. [60] A.-R. Sadeghi, C. Wachsmann, and M. Waidner, “Security and privacy
[43] L. Basora, X. Olive, and T. Dubot, “Recent advances in anomaly challenges in industrial Internet of Things,” in Proc. 52nd Annu. Des.
detection methods applied to aviation,” Aerospace, vol. 6, no. 11, 2019, Autom. Conf., 2015, pp. 1–6.
Art. no. 117, doi: 10.3390/aerospace6110117. [Online]. Available: [61] Y. Wen, M. Fashiar Rahman, H. Xu, and T.-L. B. Tseng, “Recent
https://www.mdpi.com/2226-4310/6/11/117 advances and trends of predictive maintenance from data-driven ma-
[44] R.-J. Hsieh, J. Chou, and C.-H. Ho, “Unsupervised online anomaly chine prognostics perspective,” Measurement, vol. 187, Jan. 2022,
detection on multivariate sensing time series data for smart manu- Art. no. 110276, doi: 10.1016/j.measurement.2021.110276.
facturing,” in Proc. IEEE 12th Conf. Serv.-Oriented Comput. Appl., [62] J.-Y. D. V. Ciancio, L. Homri, and A. Siadat, “Development of a
Nov. 2019, pp. 90–97, doi: 10.1109/SOCA.2019.00021. flexible data management system, to implement predictive mainte-
[45] P. Yan et al., “A comprehensive survey of deep transfer learning nance in the industry 4.0 context,” Int. J. Prod. Res., vol. 62, no. 6,
for anomaly detection in industrial time series: Methods, applica- pp. 2255–2271, 2024, doi: 10.1080/00207543.2023.2217293.
tions, and directions,” IEEE Access, vol. 12, pp. 3768–3789, 2024, [63] J. Dalzochio et al., “Machine learning and reasoning for pre-
doi: 10.1109/ACCESS.2023.3349132. dictive maintenance in Industry 4.0: Current status and chal-
[46] D. Weichert, P. Link, A. Stoll, S. Rüping, S. Ihlenfeldt, and S. Wrobel, lenges,” Comput. Ind., vol. 123, Dec. 2020, Art. no. 103298,
“A review of machine learning for the optimization of production pro- doi: 10.1016/j.compind.2020.103298.
cesses,” Int. J. Adv. Manuf. Technol., vol. 104, no. 5, pp. 1889–1902, [64] L. Ren, Z. Jia, X. Wang, J. Dong, and W. Wang, “A T2-Tensor-Aided
Oct. 2019, doi: 10.1007/s00170-019-03988-5. multiscale transformer for remaining useful life prediction in IIoT,”
[47] A. Ghasemi, F. Farajzadeh, C. Heavey, J. Fowler, and C. IEEE Trans. Ind. Informat., vol. 18, no. 11, pp. 8108–8118, Nov. 2022,
T. Papadopoulos, “Simulation optimization applied to production doi: 10.1109/TII.2022.3166790.
scheduling in the era of industry 4.0: A review and future [65] Q. Lin, Y. Zhang, S. Yang, S. Ma, T. Zhang, and Q. Xiao, “A
roadmap,” J. Ind. Inf. Integration, vol. 39, 2024, Art. no. 100599, self-learning and self-optimizing framework for the fault diagnosis
doi: 10.1016/j.jii.2024.100599. [Online]. Available: https://www. knowledge base in a workshop,” Robot. Comput.- Integr. Manuf.,
sciencedirect.com/science/article/pii/S2452414X24000438 vol. 65, Oct. 2020, Art. no. 101975, doi: 10.1016/j.rcim.2020.101975.
[48] J. P. U. Cadavid, S. Lamouri, B. Grabot, R. Pellerin, and A. Fortin, [66] Y. Ran, X. Zhou, P. Lin, Y. Wen, and R. Deng, “A survey of pre-
“Machine learning applied in production planning and control: A state- dictive maintenance: Systems, purposes and approaches,” Dec. 2019,
of-the-art in the era of industry 4.0,” J. Intell. Manuf., vol. 31, no. 6, arXiv:1912.07383.
pp. 1531–1558, Aug. 2020, doi: 10.1007/s10845-019-01531-7. [67] M. Compare, P. Baraldi, and E. Zio, “Challenges to IoT-Enabled pre-
[49] M. Babic, M. A. Farahani, and T. Wuest, “Image based qual- dictive maintenance for industry 4.0,” IEEE Internet Things J., vol. 7,
ity inspection in smart manufacturing systems: A literature no. 5, pp. 4585–4597, May 2020, doi: 10.1109/JIOT.2019.2957029.
review,” Procedia CIRP, vol. 103, pp. 262–267, Jan. 2021, [68] D. Cabrera et al., “Generative adversarial networks selection approach
doi: 10.1016/j.procir.2021.10.042. for extremely imbalanced fault diagnosis of reciprocating machinery,”
[50] G. Genta, M. Galetto, and F. Franceschini, “Inspection procedures IEEE Access, vol. 7, pp. 70643–70653, 2019.
in manufacturing processes: Recent studies and research perspec- [69] A. Kadian et al., “Sim2Real predictivity: Does evaluation in simulation
tives,” Int. J. Prod. Res., vol. 58, no. 15, pp. 4767–4788, Aug. 2020, predict real-world performance?,” IEEE Robot. Autom. Lett., vol. 5,
doi: 10.1080/00207543.2020.1766713. no. 4, pp. 6670–6677, Oct. 2020, doi: 10.1109/LRA.2020.3013848.
[51] C. Liu, Z. Zhang, D. Tang, Q. Nie, L. Zhang, and J. Song, [70] I. Rasmussen, S. Kvalsvik, P.-A. Andersen, T. N. Aune, and D. Hagen,
“A mixed perception-based human-robot collaborative maintenance “Development of a novel object detection system based on synthetic
approach driven by augmented reality and online deep reinforce- data generated from unreal game engine,” Appl. Sci., vol. 12, no. 17,
ment learning,” Robot. Comput.- Integr. Manuf., vol. 83, 2023, 2022, Art. no. 8534, doi: 10.3390/app12178534. [Online]. Available:
Art. no. 102568. https://www.mdpi.com/2076-3417/12/17/8534

VOLUME 5, 2024 1101

MANTA-COSTA ET AL.: MACHINE LEARNING APPLICATIONS IN MANUFACTURING—CHALLENGES, TRENDS, AND FUTURE DIRECTIONS

[71] M. Urgo, W. Terkaj, and G. Simonetti, “Monitoring manufactur- [89] C. Mecheri, Y. Ouazene, N.-Q. Nguyen, F. Yalaoui, T. Scaglia, and
ing systems using AI: A method based on a digital factory twin M. Gruss, “Optimizing quality inspection plans in knitting manufac-
to train CNNs on synthetic data,” CIRP J. Manuf. Sci. Tech- turing: A simulation-based approach with a real case study,” Int. J.
nol., vol. 50, pp. 249–268, 2024, doi: 10.1016/j.cirpj.2024.03.005. Adv. Manuf. Technol., vol. 131, no. 3, pp. 1167–1184, Mar. 2024,
[Online]. Available: https://www.sciencedirect.com/science/article/ doi: 10.1007/s00170-024-13075-z.
pii/S1755581724000361 [90] W. P. N. S. Panagou and F. Fruggiero, “A scoping review of hu-
[72] A. Sharma, E. Kosasih, J. Zhang, A. Brintrup, and A. Cali- man robot interaction research towards industry 5.0 human-centric
nescu, “Digital twins: State of the art theory and practice, chal- workplaces,” Int. J. Prod. Res., vol. 62, no. 3, pp. 974–990, 2024,
lenges, and open research questions,” J. Ind. Inf. Integration, doi: 10.1080/00207543.2023.2172473.
vol. 30, 2022, Art. no. 100383, doi: 10.1016/j.jii.2022.100383. [91] D. Mukherjee, K. Gupta, L. H. Chang, and H. Najjaran, “A survey
[Online]. Available: https://www.sciencedirect.com/science/article/ of robot learning strategies for human-robot collaboration in indus-
pii/S2452414X22000516 trial settings,” Robot. Comput.- Integr. Manuf., vol. 73, Feb. 2022,
[73] N. Sirisha, A. Kiran, M. Arshad, and M. M, “Automating ML models Art. no. 102231, doi: 10.1016/j.rcim.2021.102231.
using MLOps,” in Proc. Int. Conf. Advance. Smart, Secure, Intell. [92] Y. Jiang, P. Dai, P. Fang, R. Y. Zhong, and X. Cao, “Electrical-
Comput., 2024, pp. 1–5, doi: 10.1109/ASSIC60049.2024.10507923. STGCN: An electrical spatio-temporal graph convolutional network
[74] M. M. John, H. H. Olsson, and J. Bosch, “Towards MLOps: A frame- for intelligent predictive maintenance,” IEEE Trans. Ind. Informat.,
work and maturity model,” in Proc. IEEE 47th Euromicro Conf. Softw. vol. 18, no. 12, pp. 8509–8518, Dec. 2022, doi: 10.1109/TII.2022.
Eng. Adv. Appl., 2021, pp. 1–8. 3143148.
[75] D. A. Tamburri, “Sustainable MLOps: Trends and challenges,” in [93] D. Jaspert, M. Ebel, A. Eckhardt, and J. Poeppelbuss, “Smart
Proc. IEEE 22nd Int. Symp. Symbolic Numeric Algorithms Sci. Com- retrofitting in manufacturing: A systematic review,” J. Cleaner Prod.,
put., 2020, pp. 17–23. vol. 312, 2021, Art. no. 127555, doi: 10.1016/j.jclepro.2021.127555.
[76] Z. Li, Y. Zhu, and M. Van Leeuwen, “A survey on explainable anomaly [Online]. Available: https://www.sciencedirect.com/science/article/
detection,” ACM Trans. Knowl. Discov. Data, vol. 18, no. 1, pp. 1–54, pii/S095965262101773X
Sep. 2023, doi: 10.1145/3609333. [94] H. Alimam, G. Mazzuto, M. Ortenzi, F. E. Ciarapica, and M. Bevilac-
[77] A. Castellani, S. Schmitt, and S. Squartini, “Real-world anomaly de- qua, “Intelligent retrofitting paradigm for conventional machines
tection by using digital twin systems and weakly supervised learning,” towards the digital triplet hierarchy,” Sustainability, vol. 15, no. 2,
IEEE Trans. Ind. Informat., vol. 17, no. 7, pp. 4733–4742, Jul. 2021, 2023, Art. no. 1441, doi: 10.3390/su15021441. [Online]. Available:
doi: 10.1109/TII.2020.3019788. https://www.mdpi.com/2071-1050/15/2/1441
[78] R. Al-amri, R. K. Murugesan, M. Man, A. F. Abdulateef, M. A. Al- [95] A. Essien and C. Giannetti, “A deep learning model for smart man-
Sharafi, and A. A. Alkahtani, “A review of machine learning and deep ufacturing using convolutional LSTM neural network autoencoders,”
learning techniques for anomaly detection in IoT data,” Appl. Sci., IEEE Trans. Ind. Informat., vol. 16, no. 9, pp. 6069–6078, Sep. 2020,
vol. 11, no. 12, Jun. 2021, Art. no. 5320, doi: 10.3390/app11125320. doi: 10.1109/TII.2020.2967556.
[79] Y. Luo, Y. Xiao, L. Cheng, G. Peng, and D. D. Yao, “Deep learning- [96] I. A. Okaro, S. Jayasinghe, C. Sutcliffe, K. Black, P. Paoletti, and
based anomaly detection in cyber-physical systems: Progress and P. L. Green, “Automatic fault detection for laser powder-bed fusion
opportunities,” ACM Comput. Surv., vol. 54, no. 5, pp. 1–36, Jun. 2022, using semi-supervised machine learning,” Additive Manuf., vol. 27,
doi: 10.1145/3453155. pp. 42–53, May 2019, doi: 10.1016/j.addma.2019.01.006.
[80] B. Jung, H. You, and S. Lee, “Anomaly candidate extraction and [97] P. Calvo-Bascones and M. A. Sanz-Bobi, “Advanced progno-
detection for automatic quality inspection of metal casting products sis methodology based on behavioral indicators and chained se-
using high-resolution images,” J. Manuf. Syst., vol. 67, pp. 229–241, quential memory neural networks with a diesel engine appli-
Apr. 2023, doi: 10.1016/j.jmsy.2023.02.007. cation,” Comput. Ind., vol. 144, Jan. 2023, Art. no. 103771,
[81] V. Chamola, V. Hassija, A. R. Sulthana, D. Ghosh, D. Dhingra, doi: 10.1016/j.compind.2022.103771.
and B. Sikdar, “A review of trustworthy and explainable artificial [98] M.-A. Zamora-Hernández, J. A. Castro-Vargas, J. Azorin-Lopez, and
intelligence (Xai),” IEEE Access, vol. 11, pp. 78994–79015, 2023, J. Garcia-Rodriguez, “Deep learning-based visual control assistant
doi: 10.1109/ACCESS.2023.3294569. for assembly in industry 4.0,” Comput. Ind., vol. 131, Oct. 2021,
[82] M. Zhu, J. Li, W. Wang, and D. Chen, “Self-detection and self- Art. no. 103485, doi: 10.1016/j.compind.2021.103485.
diagnosis methods for sensors in intelligent integrated sensing sys- [99] D. Colucci, L. Morra, X. Zhang, D. Fissore, and F. Lamberti, “An au-
tem,” IEEE Sensors J., vol. 21, no. 17, pp. 19247–19254, Sep. 2021, tomatic computer vision pipeline for the in-line monitoring of freeze-
doi: 10.1109/JSEN.2021.3090990. drying processes,” Comput. Ind., vol. 115, Feb. 2020, Art. no. 103184,
[83] M. Parente, G. Figueira, P. Amorim, and A. Marques, “Produc- doi: 10.1016/j.compind.2019.103184.
tion scheduling in the context of Industry 4.0: Review and trends,” [100] Z. Li, Y. Wang, and K. Wang, “A deep learning driven method
Int. J. Prod. Res., vol. 58, no. 17, pp. 5401–5431, Sep. 2020, for fault classification and degradation assessment in mechan-
doi: 10.1080/00207543.2020.1718794. ical equipment,” Comput. Ind., vol. 104, pp. 1–10, Jan. 2019,
[84] O. E. Oluyisola, S. Bhalla, F. Sgarbossa, and J. O. Strandhagen, doi: 10.1016/j.compind.2018.07.002.
“Designing and developing smart production planning and con- [101] J. Božič, D. Tabernik, and D. Skočaj, “Mixed supervision for
trol systems in the industry 4.0 era: A methodology and case surface-defect detection: From weakly to fully supervised learn-
study,” J. Intell. Manuf., vol. 33, no. 1, pp. 311–332, Jan. 2022, ing,” Comput. Ind., vol. 129, Aug. 2021, Art. no. 103459,
doi: 10.1007/s10845-021-01808-w. doi: 10.1016/j.compind.2021.103459.
[85] Y. J. Cruz, M. Rivas, R. Quiza, R. E. Haber, F. Castaño, and [102] F. Zhang, J. Yan, P. Fu, J. Wang, and R. X. Gao, “Ensemble sparse
A. Villalonga, “A two-step machine learning approach for dy- supervised model for bearing fault diagnosis in smart manufacturing,”
namic model selection: A case study on a micro milling pro- Robot. Comput.- Integr. Manuf., vol. 65, Oct. 2020, Art. no. 101920,
cess,” Comput. Ind., vol. 143, Dec. 2022, Art. no. 103764, doi: 10.1016/j.rcim.2019.101920.
doi: 10.1016/j.compind.2022.103764. [103] S. Guo, M. Chen, A. Abolhassani, R. Kalamdani, and W. G.
[86] M. Sánchez, E. Exposito, and J. Aguilar, “Implementing self-* Guo, “Identifying manufacturing operational conditions by physics-
autonomic properties in self-coordinated manufacturing processes based feature extraction and ensemble clustering,” J. Manuf.
for the Industry 4.0 context,” Comput. Ind., vol. 121, Oct. 2020, Syst., vol. 60, pp. 162–175, Jul. 2021, doi: 10.1016/j.jmsy.2021.
Art. no. 103247, doi: 10.1016/j.compind.2020.103247. 05.005.
[87] J. Schmitt, J. Bönig, T. Borggräfe, G. Beitinger, and J. Deuse, “Pre- [104] L. Chen, G. Xu, S. Zhang, W. Yan, and Q. Wu, “Health indicator
dictive model-based quality inspection using machine learning and construction of machinery based on end-to-end trainable convolu-
edge cloud computing,” Adv. Eng. Informat., vol. 45, Aug. 2020, tion recurrent neural networks,” J. Manuf. Syst., vol. 54, pp. 1–11,
Art. no. 101101, doi: 10.1016/j.aei.2020.101101. Jan. 2020, doi: 10.1016/j.jmsy.2019.11.008.
[88] V. Azamfirei, A. Granlund, and Y. Lagrosen, “Lessons from adopt- [105] M. Mehta and C. Shao, “Federated learning-based semantic
ing robotic in-line quality inspection in the Swedish manufactur- segmentation for pixel-wise defect detection in additive man-
ing industry,” Procedia Comput. Sci., vol. 217, pp. 386–394, 2023, ufacturing,” J. Manuf. Syst., vol. 64, pp. 197–210, Jul. 2022,
doi: 10.1016/j.procs.2022.12.234. doi: 10.1016/j.jmsy.2022.06.010.

1102 VOLUME 5, 2024

 ALEXANDRE MANTA-COSTA received the B.S. RICARDO SILVA PERES received the M.Sc. and
and M.Sc. degrees in electrical and computer en- Ph.D. degrees in electrical and computer engi-
gineering in 2019 and 2022, respectively, from the neering from the NOVA School of Science and
NOVA School of Science and Technology (NOVA- Technology (NOVA-SST), Caparica, Portugal, in
SST), Caparica, Portugal, where he is currently 2015 and 2019, respectively, specializing in the
working toward the Ph.D. degree in cloud manu- application of AI and robotics in smart manufac-
facturing, industrial AI and cyber physical systems. turing.
Since 2021, he has been a Teaching Assistant He is currently an Assistant Professor with
with NOVA-SST. He is currently a Researcher with the Department of Electrical Engineering, NOVA-
UNINOVA-Centre of Technology and Systems, SST. Since 2014, he has been a Researcher with
Caparica, focusing on applying Industrial Artificial UNINOVA-Centre of Technology and Systems,
Intelligence in the manufacturing context. His research interests include in- Caparica, Portugal, working at the intersection of AI, robotics, and data
dustrial artificial intelligence, deep learning, Internet of Things, and I4.0. science. He has participated in several national and international research
projects. He is the author of several publications in high-ranking international
scientific journals and conference proceedings (peer-reviewed). His research
interests include Industrial AI, cyber-physical systems, and multiagent sys-
tems.
Dr. Peres has also been a member of the IEEE IES Technical Committee
on Industrial Agents since 2018 and of the IEEE Standards Association
P2805.1/2/3 Edge Computing Nodes Working Group since 2019.

JOSÉ BARATA (Member, IEEE) received the

Ph.D. degree in robotics and integrated manufac-
SARA OLEIRO ARAÚJO received the B.Sc. de- turing from the NOVA School of Science and
gree in chemical and biological engineering in Technology (NOVA-SST), Caparica, Portugal, in
2012 from the Portuguese Higher Education Poly- 2004.
technic Institution of Engineering and Technology He is currently a Full Professor with the Depart-
(ISEL), Lisbon, Portugal, and the M.Sc. degree ment of Electrical Engineering, Nova University of
in food technology and safety in 2015 from the Lisbon, Lisbon, Portugal, and a Senior Researcher
NOVA School of Science and Technology (NOVA- from the UNINOVA-Centre of Technology and
SST), Caparica, Portugal, where she is currently Systems, Caparica, Portugal. Since 2004, he has
working toward the Ph.D. degree in agro-industrial been leading the UNINOVA participation in EU
technologies. projects. He has authored or coauthored more than 100 original papers in
With a strong background in chemical engineer- international journals and international conferences. His main research inter-
ing, she has gained extensive experience in the field of pharmaceutical quality ests are in the area of intelligent manufacturing with a particular focus on
control. Since 2019, she has been integral to Portugal’s UNINOVA-Centre of complex adaptive systems involving intelligent manufacturing devices.
Technology and Systems team, Caparica, Portugal. Her research interests in- Dr. Barata is a member of the IEEE technical committees on Industrial
clude food technology, chemical engineering, digital agriculture, data science, Agents (IES), Self-Organisation and Cybernetics for Informatics (SMC), and
and artificial intelligence. Education in Engineering and Industrial Technologies (IES).

VOLUME 5, 2024 1103