Artificial Intelligence in Predictive Maintenance: A

Comprehensive Review of Methods, Applications

and Future Directions
Tharak Nagaveti1 Mundlapati Hem Charan1 Aditya Sontena1 Tejesh Varma1

Abstract—Predictive Maintenance (PdM) has become a critical

component of modern industrial operations, leveraging advance-
ments in Artificial Intelligence (AI) and Machine Learning
(ML) to anticipate equipment failures and optimize maintenance
strategies. This paper explores cutting-edge ML techniques such
as meta-learning, federated learning, generative AI, and graph
neural networks (GNNs) to enhance PdM’s predictive capabili-
ties. The integration of ML with physics-based models, including
Physics-Informed Neural Networks (PINNs) and AI-enhanced
Finite Element Analysis (FEA), is examined to improve reliability
in high-stakes industries like aerospace, energy, and automotive.
A key challenge in PdM is data scarcity, which is addressed
through generative AI-based synthetic failure data augmentation
and federated learning for decentralized model training. Explain-
able AI (XAI) methods such as SHAP, LIME, and attention-based
models are explored to enhance model interpretability and foster
trust in AI-driven maintenance decisions. Furthermore, Edge AI
and TinyML are analyzed for their potential to enable real-time,
low-latency fault detection in resource-constrained environments.
The paper also highlights the economic and operational ben-
efits of AI-driven PdM through case studies, demonstrating sig-
nificant reductions in downtime, maintenance costs, and energy
consumption. Ethical and regulatory considerations, including Fig. 1. Taxonomy of Machine Learning approaches used in Predictive
data privacy and AI bias, are discussed alongside emerging Maintenance, showing the hierarchy from Artificial Intelligence through var-
solutions like federated learning and blockchain for secure, ious learning paradigms to specific algorithms. The framework encompasses
supervised learning (classification and regression), unsupervised learning
collaborative PdM. Finally, future research directions emphasize (clustering), and reinforcement learning methods.
the role of next-generation digital twins, quantum computing, and
self-healing AI systems in further advancing PdM technologies.
Keywords— Predictive Maintenance, Artificial Intelligence, Machine
Learning, Industry 4.0, Digital Twins, Federated Learning
in fostering trust and transparency in AI-driven maintenance
I. I NTRODUCTION decisions, as well as the potential of Edge AI and TinyML for
real-time, low-latency fault prediction.
Predictive maintenance (PdM) has emerged as a cornerstone In addition to technological advancements, this paper high-
of modern industrial operations [21], particularly in the era of lights the importance of cross-industry collaboration, standard-
Industry 4.0 [22]. The integration of advanced technologies ization, and open data sharing in driving the development of
such as Artificial Intelligence (AI) and Machine Learning more robust and generalizable AI models. We also address
(ML) [20] is revolutionizing traditional maintenance practices. the ethical, regulatory, and sustainability challenges associated
Recent studies show that AI-driven PdM can reduce mainte- with AI-driven PdM, emphasizing the need for carbon-aware
nance costs by up to 30% [19] and increase equipment uptime maintenance scheduling and the role of predictive models in
by 20% [1]. promoting a circular economy. Through a series of case studies
This paper seeks to address these challenges by exploring and benchmarking analyses, we demonstrate the practical
the latest innovations in ML and AI for PdM, including applications and economic benefits of AI-driven PdM, while
meta-learning, federated learning, generative AI, and graph also identifying key areas for future research.
neural networks (GNNs). We also delve into the integration of
ML with physics-based approaches, such as Physics-Informed II. M ACHINE L EARNING I NNOVATIONS FOR P D M
Neural Networks (PINNs) and AI-enhanced Finite Element Predictive maintenance (PdM) relies on machine learning
Analysis (FEA), which are critical for ensuring the reliability (ML) models trained with historical failure data to forecast
and safety of predictive models in high-stakes environments. equipment breakdowns. As illustrated in Figure 1, various ML
Furthermore, we examine the role of Explainable AI (XAI) approaches, from traditional supervised learning to advanced
 B. Federated Learning for Decentralized PdM
Federated learning (FL) has emerged as a key technique
for industrial PdM by addressing privacy and data ownership
concerns. Unlike traditional centralized ML approaches that
require data aggregation in a single repository, FL enables
multiple industrial sites to collaboratively train a global model
while keeping their data localized. This ensures compliance
with data protection regulations and reduces communication
overhead in Industrial Internet of Things (IIoT) environments
[29].
Chen et al. [29] proposed a federated meta-learning frame-
work for fault diagnosis, which leverages meta-learning to
adapt the global model to new machinery with minimal data.
The framework enhances convergence speed and model accu-
racy in non-independent and identically distributed (non-IID)
Fig. 2. Federated meta-learning framework for industrial PdM showing the data scenarios. The study’s empirical evaluation on real-world
interaction between central server and production line clients. The process industrial datasets demonstrated that federated meta-learning
involves: (1) sending global meta-learner to clients, (2) local meta-learner significantly improves fault detection accuracy while preserv-
updates through two-stage update process, (3) uploading local meta-learners
to server, and (4) aggregating global meta-learner. ing data privacy. However, challenges such as communication
costs, model aggregation techniques, and computational over-
head in edge devices remain areas for future research.
A major limitation of federated learning in PdM is the
deep learning methods, are employed in modern PdM systems. heterogeneity of industrial environments. Differences in sensor
Each approach offers unique advantages for different mainte- types, operating conditions, and failure modes across sites can
nance scenarios and data availability conditions. lead to disparities in local model updates. To mitigate this,
adaptive federated learning techniques that assign dynamic
A. Meta-Learning for Few-Shot Failure Prediction weighting to client updates based on data relevance and
Predictive maintenance (PdM) relies on machine learning reliability have been explored. Future work should focus on
(ML) models trained with historical failure data to fore- optimizing these techniques to ensure robust model perfor-
cast equipment breakdowns. However, traditional ML mod- mance across diverse industrial settings.
els require large labeled datasets, which may not always
be available in industrial settings. Meta-learning, particularly III. H YBRID M ODELS : I NTEGRATING M ACHINE
Model-Agnostic Meta-Learning (MAML), has emerged as a L EARNING WITH P HYSICS -BASED A PPROACHES
promising approach to address the few-shot learning challenge
A. Why Pure ML Models Fail in High-Stakes Industries
in PdM. By training models to adapt quickly to new tasks
with minimal labeled data, meta-learning enables rapid fault Machine learning (ML) has revolutionized predictive main-
detection with high accuracy [30]. tenance and failure detection in various industries, yet its
A key advantage of meta-learning is its ability to generalize application in high-stakes domains such as aerospace, civil
across different fault types. Wu et al. [30] demonstrated engineering, and advanced manufacturing remains limited due
that meta-learning models trained on limited instances of to inherent shortcomings. Traditional ML models operate as
rotating machinery faults outperformed conventional deep black-box predictors, relying on large volumes of labeled data
learning models in scenarios with scarce labeled data. The to detect patterns and anomalies. However, in industries where
few-shot learning framework, when combined with episodic failures are rare but catastrophic, such as aerospace and nuclear
training, enhances the model’s ability to recognize anomalies power, obtaining sufficient labeled data is impractical. Conse-
in different operational conditions. Furthermore, integrating quently, pure ML models often suffer from poor generalization
transfer learning with meta-learning allows PdM models to when applied to unseen failure conditions.
leverage knowledge from related domains, improving fault Moreover, conventional ML approaches lack interpretability,
diagnosis accuracy and reducing dependency on extensive making them unsuitable for applications that require stringent
labeled datasets. safety validations. Engineers and regulators in high-stakes
Figure 2 illustrates the federated meta-learning framework fields demand models that align with established physical prin-
for industrial PdM, where multiple production lines collaborate ciples to ensure robustness and reliability. Without physics-
to build a robust predictive model while maintaining data based constraints, purely data-driven ML models may produce
privacy. This approach enables rapid adaptation to new fault unrealistic predictions that contradict known engineering prin-
types with minimal labeled data, while leveraging the collec- ciples. Additionally, computational efficiency remains a con-
tive experience across different production environments. cern; high-fidelity simulations, such as finite element analysis
(FEA), are accurate but computationally expensive, while data-
driven ML methods, though faster, often fail to capture the
underlying physics governing structural behavior [31].

B. Physics-Informed Neural Networks (PINNs) for Real-Time

Failure Simulation
To overcome the limitations of pure ML models, researchers
have developed Physics-Informed Neural Networks (PINNs), Fig. 3. XAI workflow demonstrating the process from data input to decision-
which embed physical laws as constraints within deep learning making. The workflow shows how raw data is processed through a learning
models. Unlike traditional ML, PINNs incorporate governing model, which provides predictions along with explanations using interpretable
indicators, enabling informed decision-making by human operators.
equations such as partial differential equations (PDEs) into the
loss function, ensuring that predictions adhere to established
physical principles [32].
PINNs have shown promise in structural health monitoring
different materials and loading conditions. By integrating ML-
and failure simulation, particularly in dynamic stress analysis.
based feature extraction with physics-based wear simulations,
For example, [35] introduced a PINN-Stress model capable of
researchers have developed predictive maintenance models
predicting stress distributions in structural components under
capable of dynamically updating wear predictions based on
dynamic loading conditions. This model outperforms purely
real-time sensor data. This approach has proven particularly
data-driven ML approaches by leveraging PDE constraints to
effective in aerospace applications, where component degra-
enforce physically consistent stress predictions. Compared to
dation must be predicted with high precision to ensure flight
traditional finite element solvers, PINN-Stress models achieve
safety [32].
near real-time predictions while maintaining accuracy, making
them ideal for applications such as seismic analysis and impact D. Challenges and Future Directions in Hybrid Models
resistance assessments.
A key advantage of PINNs is their ability to generalize Despite their advantages, hybrid ML-physics models face
to new geometries and loading conditions without requiring several challenges that must be addressed to facilitate
extensive retraining. By incorporating boundary conditions and widespread industrial adoption. One significant limitation
physics-based loss functions, PINNs provide robust failure is the computational cost of training PINNs, as enforcing
predictions for structures experiencing extreme environmental physics-based constraints increases optimization complexity.
loads. However, challenges such as hyperparameter tuning and Additionally, data scarcity remains an issue, particularly for
convergence stability must be addressed to optimize PINNs for failure-critical applications where labeled failure data is rare.
large-scale industrial applications. To address this, researchers are exploring transfer learning
techniques, allowing models trained on simulated data to
C. AI-Enhanced Finite Element Analysis (FEA) for Industrial generalize to real-world conditions.
Wear Prediction Another challenge is the lack of standardized methodologies
Finite Element Analysis (FEA) is a cornerstone of compu- for integrating ML with physics-based solvers. Different indus-
tational mechanics, enabling engineers to model stress, strain, tries employ diverse simulation frameworks, necessitating the
and deformation in complex structures. Despite its accuracy, development of interoperable AI-driven simulation pipelines.
FEA is computationally expensive, often requiring hours or Collaborative efforts between domain experts and ML re-
days to simulate high-fidelity structural behaviors. Recent searchers are crucial to ensuring that hybrid models align with
advancements in hybrid ML-FEA models aim to accelerate engineering best practices and regulatory requirements.
these simulations while maintaining predictive accuracy [31]. Finally, the interpretability of hybrid models must be im-
Hybrid models integrating ML with FEA leverage data- proved to gain trust from engineers and decision-makers.
driven surrogates to replace or augment computationally in- While PINNs offer greater transparency than black-box ML
tensive FEA simulations. For example, convolutional neural models, further research is needed to develop explainable
networks (CNNs) can be trained on high-fidelity FEA sim- AI (XAI) techniques that provide insights into how hybrid
ulations to approximate stress distributions in new structural models make predictions. By enhancing interpretability, hybrid
designs, significantly reducing computation time. [31] devel- models can facilitate AI-assisted decision-making in high-
oped a hybrid PINN-FEA model, combining the efficiency stakes engineering applications.
of ML with the robustness of physics-based solvers. This
IV. E XPLAINABLE AI (XAI) FOR I NDUSTRIAL A DOPTION
model replaces automatic differentiation in traditional PINNs
with finite volume approximations, improving computational The integration of Artificial Intelligence (AI) in predictive
efficiency and convergence rates. maintenance (PdM) has emerged as a pivotal strategy for
Another application of AI-enhanced FEA is in wear predic- reducing carbon footprints in industrial settings. As illustrated
tion for industrial machinery. Conventional wear models rely in Figure 3, XAI systems provide transparent decision-making
on empirical equations that may not generalize well across processes by not only generating predictions but also offering
 30% of AI-predicted failures were biased toward premium
machinery, leading to avoidable losses of $50 million annually
due to suboptimal maintenance scheduling [23].

B. SHAP, LIME, and Attention-Based XAI for PdM Models

To address the limitations of black-box PdM models, var-
ious explainable AI (XAI) methodologies have been imple-
mented. Three of the most widely adopted techniques include
SHapley Additive exPlanations (SHAP), Local Interpretable
Model-Agnostic Explanations (LIME), and Attention-Based
XAI Models.
Fig. 4. Schematic representation of black-box model interpretation in SHAP (SHapley Additive Explanations) is a game-theoretic
maintenance decision-making. The diagram shows the flow from input data
through the black-box model to decision-making, highlighting the need for
approach that quantifies the contribution of each input feature
interpretable elements and explanation methods to transform opaque decisions to the AI model’s decision. In a multi-factory PdM case
into informative motivations. study, SHAP revealed that temperature fluctuations accounted
for 63% of all failure predictions, while vibration levels
contributed 24%. This insight allowed predictive maintenance
clear explanations for those predictions, making AI systems
teams to adjust operational parameters, reducing unexpected
more trustworthy and actionable for industrial operators.
failures by 38% over six months [24].
A. Risks of Black-Box Models in Maintenance Decision- LIME (Local Interpretable Model-Agnostic Explanations)
Making provides localized, instance-specific explanations by perturb-
ing input data and analyzing resulting changes in predictions.
Artificial Intelligence (AI) has revolutionized predictive
This method is particularly useful when maintenance teams
maintenance (PdM) by enabling early failure detection, opti-
need to verify why a specific failure prediction was made. For
mizing asset utilization, and reducing unexpected downtimes.
instance, in an oil refinery maintenance study, LIME identified
However, despite its high accuracy, AI remains a black-box
that corrosion levels in pipelines were incorrectly flagged
system in many industrial applications, making it difficult for
as critical in 17% of cases due to sensor calibration errors.
engineers and decision-makers to understand its predictions.
By correcting these misclassifications, the refinery reduced
As illustrated in Figure 4, while black-box models can
unnecessary shutdowns by 22% [24].
achieve high accuracy in predictions, their lack of interpretabil-
ity presents significant risks, particularly in sectors such as Attention-Based XAI Models leverage deep learning ar-
aerospace, automotive, and manufacturing, where safety and chitectures to highlight the most influential data patterns in
compliance are paramount [23]. real-time PdM applications. Unlike SHAP and LIME, which
are post-hoc explainability techniques, attention-based models
One of the primary concerns is the trust deficit in AI-
natively integrate interpretability into neural networks. This is
generated predictions. Studies have shown that 82% of indus-
particularly beneficial in high-speed manufacturing environ-
trial engineers are reluctant to adopt AI-based maintenance
ments, where real-time decision-making is essential. Research
recommendations without a clear explanation of the decision-
in automated production lines found that using attention-based
making process. In critical applications, such as aircraft engine
AI improved fault detection rates by 12%, while reducing
monitoring, an AI model predicting an impending failure with
false alarms by 19%, leading to a 26% increase in production
97% accuracy is insufficient if the model cannot justify the
efficiency [25].
decision. This results in over 45% of AI-generated main-
tenance recommendations being disregarded due to lack of
C. Case Study: Implementing XAI in Automotive Manufactur-
transparency [33].
ing
Moreover, regulatory compliance issues are exacerbated by
the opacity of black-box models. The European Union’s Gen- The automotive industry has been at the forefront of
eral Data Protection Regulation (GDPR) mandates explain- AI-driven predictive maintenance, utilizing machine learning
ability, requiring that AI-driven decisions impacting safety or models to optimize maintenance schedules, detect anomalies,
financial stability be interpretable [34]. Failure to comply with and improve equipment uptime. However, the lack of inter-
these regulations has already resulted in legal disputes and pretability in AI-driven PdM systems has hindered widespread
fines, with estimated compliance costs exceeding $3.5 billion adoption. A case study conducted in an automotive assembly
annually across industrial sectors. plant showcases the impact of integrating XAI techniques into
A significant challenge is bias and systemic errors in AI- AI-driven PdM [25].
driven PdM. Historical data inconsistencies can cause the AI The facility initially deployed a deep learning PdM system
model to prioritize maintenance for high-value machinery, to predict robotic arm failures in the welding and assembly
inadvertently neglecting equally critical but lower-cost assets. processes. Although the model achieved 95% accuracy, main-
A study conducted in automotive manufacturing revealed that tenance teams were reluctant to act on its recommendations
due to its lack of transparency. To address this, engineers (CNNs), recurrent neural networks (RNNs), and long short-
integrated SHAP and LIME into the PdM workflow. term memory (LSTM) networks have been the most effective.
By applying SHAP, the team discovered that abnormal Studies have shown that employing CNNs for fault detection in
temperature fluctuations in robotic joints were responsible for rotating machinery improves accuracy by 20% while reducing
71% of failure predictions. This insight prompted a redesign of computational time by 40% [37].
the cooling systems, leading to a 21% reduction in overheating A comparative analysis of different ML models on edge
incidents and extending the operational lifespan of robotic devices revealed the following execution times per inference
arms by 14%. cycle:
LIME provided further interpretability by explaining spe-
• Neural Networks (NNs): 65 ms (Jetson Nano) [36]
cific failure instances. It was found that sensor noise con-
• Decision Trees: 80 ms (Jetson Nano) [36]
tributed to false alarms in 9% of cases, resulting in unnecessary
• SVM (Support Vector Machines): 40 ms (CPU) [36]
maintenance actions. By implementing sensor recalibration
procedures, the company reduced false maintenance alerts by These results illustrate that NN-based models offer the best
18%. trade-off between speed and accuracy, particularly when opti-
To enhance real-time monitoring, an attention-based XAI mized for GPU acceleration on embedded devices.
model was introduced, which prioritized sensor readings
linked to imminent failures. This approach led to a 30% D. Feature Engineering for High-Speed Fault Prediction
increase in PdM adoption across the facility, improving overall
equipment effectiveness (OEE) by 17% and reducing down- Effective PdM relies on accurate feature extraction and
time by 12%. selection from sensor data. Key extracted features include:
• Energy of the centered signal
V. AI FOR R EAL -T IME , L OW-L ATENCY P REDICTIVE
• Energy of the derivative signal
M AINTENANCE (P D M)
• Shannon entropy for anomaly detection
A. The Advent of AI in Predictive Maintenance
Using these features, AI models can classify equipment
The advent of Artificial Intelligence (AI) in Predictive states into categories such as normal, unbalanced, belt loose-
Maintenance (PdM) has revolutionized industrial operations, ness, and bearing fault with an accuracy of 92% [36].
particularly with real-time, low-latency fault detection and
anomaly prediction. AI-based PdM integrates edge computing,
E. Case Study: Edge AI for Smart Factory PdM
deep learning algorithms, and real-time data analytics to ensure
rapid response and high operational efficiency. By leveraging A recent implementation of edge AI for PdM in a smart
AI at the edge, industries can significantly reduce latency factory setting demonstrated significant operational improve-
issues inherent in cloud-based solutions, enhancing predictive ments. The deployed AI system utilized edge-based CNN
accuracy while ensuring minimal downtime and cost savings. models for real-time analysis of sensor data from industrial
motors. The results were:
B. Edge Computing in PdM: Enhancing Real-Time Capabili-
ties • 30% faster anomaly detection compared to traditional
Traditional PdM relies heavily on cloud computing, leading cloud-based PdM [37]
to latency issues due to data transmission and processing • 20% reduction in maintenance costs due to early failure
delays. Edge AI addresses these challenges by processing prediction [37]
data locally on industrial IoT (IIoT) devices. According to • Improvement in production efficiency by 15% through
recent studies, implementing edge-based PdM reduces latency optimized scheduling [37]
by 70% compared to cloud-centric models [36]. Additionally, Additionally, leveraging federated learning in this setup
using embedded AI models on devices such as NVIDIA Jetson allowed multiple manufacturing units to collaboratively train
Nano and Raspberry Pi has resulted in a 30% improvement in AI models without data sharing, ensuring privacy compliance
real-time failure detection accuracy [36]. while enhancing model robustness [37].
For instance, a study on industrial screw compressors
demonstrated that real-time vibration data processed at the F. Challenges and Future Directions
edge led to an 85% reduction in false alarms and an increase in
system uptime by 25% [36]. This improvement was achieved Despite the success of AI in real-time PdM, several chal-
through on-device signal processing and feature extraction, lenges remain. Scalability remains an issue, particularly for
eliminating the dependency on high-latency cloud-based an- large-scale deployments across multiple factories. Further re-
alytics. search into lightweight AI models, such as TinyML, is needed
to enable real-time inference on resource-constrained devices.
C. AI Models for Low-Latency PdM Additionally, integrating 5G and 6G networks can further
Various AI models have been employed to enhance the ef- reduce latency and improve connectivity between edge devices
ficiency of PdM. Among these, convolutional neural networks [37].
 C ROSS -I NDUSTRY C OLLABORATION : S TANDARDIZATION 1) Standardization: Shared datasets often come with stan-
& O PEN DATA dardized formats and annotations, making it easier for
Automotive, Healthcare, and Energy Sector PdM Comparison researchers to compare different models and approaches
[4].
Predictive maintenance (PdM) has become a cornerstone
2) Reproducibility: Shared datasets enable researchers to
of modern industrial operations, with its application spanning
reproduce and validate results, which is critical for
across various sectors, including automotive, healthcare, and
advancing the field [15].
energy. Each sector presents unique challenges and opportu-
3) Cross-Industry Learning: Shared datasets facilitate
nities for PdM, driven by the nature of the equipment, the
cross-industry learning, where insights gained from one
criticality of failures, and the availability of data.
sector can be applied to another [16].
In the automotive sector, PdM is primarily focused on
4) Robustness: Shared datasets often include a wide range
ensuring the reliability and safety of vehicles, particularly
of operating conditions and failure modes, making the
in critical components such as engines, transmissions, and
resulting models more robust [17].
braking systems. The automotive industry has been a pioneer
in adopting PdM, leveraging advanced sensors and IoT tech- Case Study: NASA’s Turbofan Engine Dataset and Its Impact
nologies to monitor vehicle health in real-time. For instance, on PdM Research
the C-MAPSS dataset, which simulates turbofan engine degra-
The NASA C-MAPSS (Commercial Modular Aero-
dation, has been extensively used to develop and validate PdM
Propulsion System Simulation) dataset has had a profound
models. The dataset includes 21 sensor measurements and
impact on PdM research, particularly in the development of
three operational settings, providing a rich source of data for
data-driven models for predicting the remaining useful life
training deep learning models. The automotive sector’s focus
(RUL) of turbofan engines. The dataset, which was released
on high-volume production and stringent safety standards has
in 2008, simulates the degradation of turbofan engines under
driven the need for accurate and timely PdM solutions.
different operating conditions and fault modes. It includes 21
In the healthcare sector, PdM is applied to medical equip-
sensor measurements and three operational settings, providing
ment such as MRI machines, ventilators, and surgical robots.
a rich source of data for training and validating PdM models.
The failure of such equipment can have dire consequences,
making PdM critical for ensuring patient safety and minimiz-
ing downtime. However, the healthcare sector faces unique
challenges, including the need for highly accurate predictions
due to the life-critical nature of the equipment and the com-
plexity of the data. Unlike the automotive sector, where data is
often collected from controlled environments, healthcare data
is often noisy and incomplete, requiring more sophisticated
data preprocessing techniques. For example, in a study on MRI
machine maintenance, researchers found that incorporating
domain knowledge from healthcare professionals significantly
improved the accuracy of PdM models, reducing false alarms
by 15
The energy sector, particularly in power generation and
distribution, also heavily relies on PdM to ensure the reliability
of critical infrastructure such as turbines, transformers, and
power grids. The energy sector faces challenges related to the
harsh operating conditions of equipment, which can lead to
rapid degradation and unexpected failures. For example, in
wind turbines, the failure of a single component can lead to
Fig. 5. Modern turbofan engine in a maintenance facility, representing
significant downtime and repair costs. A study on wind turbine the type of complex aerospace system for which the NASA C-MAPSS
maintenance found that implementing PdM reduced unplanned dataset was developed. The engine’s sophisticated sensor network and multiple
downtime by 20% and maintenance costs by 15%. The energy subsystems demonstrate the complexity of predictive maintenance applications
in aerospace.
sector also benefits from the availability of large datasets, such
as those collected from SCADA systems, which provide real-
The impact of the C-MAPSS dataset on PdM research can
time monitoring of equipment health.
be quantified in several ways:
Benefits of Shared Datasets for More Robust AI Models 1) Model Performance: The dataset has enabled re-
One of the key enablers of PdM across these sectors is searchers to develop models with increasingly accurate
the availability of shared datasets, which allow researchers RUL predictions. For example, the RMSE of models
and practitioners to develop and validate models on real-world developed using the C-MAPSS dataset has decreased
data. Shared datasets provide several benefits, including: from 18.44 in 2016 (Babu et al., 2016) to 7.78 in 2022
 (Asif et al., 2022), representing a 58% improvement in Algorithm 1 Proposed Methodology for Turbofan Engine
accuracy. RUL Prediction
2) Algorithm Diversity: The dataset has been used to Training Data, T = (x11 , y11 ), (x12 , y21 )...(xE E
N , yN )
develop a wide range of algorithms, including traditional Parameters: N → No. of sensor variables, E → No. of
machine learning models, deep learning models, and Engines, N → Filtered data, C → Correlated sensors,
hybrid models. For example, a study by Zhang et al. M → No. of correlated sensors, D → Normalized
(2017) used a multi-objective evolutionary algorithm data, P → Piece-wise linear RUL function, LN →
to develop a deep belief network for RUL prediction, No. of layers in LSTM, LD → Dropout layers, RL →
achieving an RMSE of 16.14 on the FD001 sub-dataset. Regression layer, LF → No. of neuron in FC layer,
Another study by Wang et al. (2018) used a hybrid W → Window Length, η → Learning rate, GD →
model combining LSTM and gradient boosting regres- Gradient descent optimizer, V s → Validation set,
sion (GBR) to achieve an RMSE of 15.68 on the same Starc → Network architecture parameters Output:
sub-dataset. Performance Parameter, Opp [RMSE, Score] i = 1
3) Cross-Industry Applications: The insights gained from to E j = 1 to N C = Correlation Analysis(xij , yji )
the C-MAPSS dataset have been applied to other in- k = 1 to M N = Median Filter(kC , W )
dustries, such as energy and healthcare. For example, a D = Data Normalization(kN , mean(N ), SD(N ))
study by Liu et al. (2019) applied techniques developed P = Piece-wise linear function(kC ) Starc ←
using the C-MAPSS dataset to predict the RUL of wind (LN, LH, LF, η, GD, LD, V s, RL, D, P ) l = 1 to
turbines, achieving an RMSE of 8.12. Similarly, a study T F D = Forward Pass(Starc ) Et = Error(F D) Bp =
by Elattar et al. (2018) applied these techniques to Back Propagation(Et ) Wnew = Wold + η ∗ derivative(Bp)
predict the RUL of medical equipment, achieving an bnew = bold + η ∗ derivative(Bp) Calculate RMSE using
RMSE of 9.32. equation 14 Calculate score function using equation 15
4) Research Output: The C-MAPSS dataset has been cited
in over 500 research papers, making it one of the most This algorithm demonstrates a systematic approach to pro-
widely used datasets in PdM research. The dataset has cessing the C-MAPSS dataset, incorporating correlation anal-
also been used in several international competitions, ysis, data normalization, and LSTM-based deep learning to
such as the PHM08 Data Challenge, which has further predict RUL. The methodology has achieved an RMSE of 7.78
accelerated research in the field. on the FD001 sub-dataset, representing one of the best results
in the field.
The C-MAPSS dataset has been used to develop various
algorithmic approaches for RUL prediction. Figure 6 shows VI. S USTAINABILITY AND C IRCULAR E CONOMY IN P D M
the heatmap visualization of the four sub-datasets (FD001- A. AI-driven Carbon Footprint Reduction in Industrial Main-
FD004), illustrating the correlation between sensor features tenance
and Remaining Useful Life (RUL) predictions. Algorithm 2 The integration of Artificial Intelligence (AI) in predictive
maintenance (PdM) has emerged as a pivotal strategy for re-
ducing carbon footprints in industrial settings. AI-driven PdM
leverages advanced data analytics and machine learning algo-
rithms to optimize maintenance schedules, thereby minimizing
energy consumption and reducing waste. According to the
KYKLOS 4.0 project, AI-based predictive maintenance can
significantly extend the lifecycle of machinery and equipment,
which directly contributes to reducing the carbon footprint as-
sociated with manufacturing processes. For instance, the KYK-
LOS 4.0 DL toolkit, which employs deep learning techniques,
has demonstrated the ability to predict machine failures and
degraded performance with high accuracy, allowing for timely
interventions that prevent unnecessary energy consumption
and material waste.
Fig. 6. Heat map visualization of the C-MAPSS dataset showing correlation
In the context of heavy industries, such as shipyards, the
factors across different operational sensors (s-1 to s-21) for all four sub- implementation of AI-driven PdM has shown promising re-
datasets (FD001-FD004). The color intensity represents the strength of cor- sults. The Astander use case, as described in the KYKLOS 4.0
relation with RUL prediction, where darker green indicates lower correlation
and yellow indicates higher correlation
project, involved data collected from a shipyard crane over a
year. The toolkit analyzed 110 non-constant features, including
80 binary values and 30 numeric values, sampled once a
presents a comprehensive methodology that has achieved state- minute. By aggregating data over an hour and applying prepro-
of-the-art results on this dataset: cessing techniques, the toolkit was able to predict anomalies
 environments, including manufacturing, and has shown the
ability to detect anomalies and predict material degradation
with high accuracy.
In the Astander use case, the toolkit analyzed data from a
shipyard crane, focusing on components such as the translation
system, main and auxiliary hooks, rotation movement, and
arm extension. By applying Principal Component Analysis
(PCA), the toolkit identified that only around 30 principal
components were necessary to describe over 80% of the
variance in the data. This indicates that many features are
correlated, and multiple features can be described by a single
principal component, simplifying the predictive model.
The predictive model for material degradation can be ex-
pressed as:
Fig. 7. Correlation matrix of selected features from the translation system.
The matrix shows the relationships between different sensor measurements
and operational parameters, with values ranging from -1 (strong negative RUL = Initial Lifespan − Time Elapsed − Degradation Factor
correlation, dark blue) to 1 (strong positive correlation, yellow).
(3)
Where the Degradation Factor is a function of operational
conditions, such as load, temperature, and usage frequency.
in the crane’s translation system with an anomaly score of 0.3, For example, if a component has an initial lifespan of 10,000
indicating a potential degradation in performance. This early hours, 5,000 hours have elapsed, and the Degradation Factor
detection allowed for maintenance to be scheduled before a is 2,000 hours, the RUL would be:
failure occurred, thereby reducing downtime and energy waste.
The carbon footprint reduction can be quantified using the 10, 000 − 5, 000 − 2, 000 = 3, 000 hours (4)
following formula:
This prediction allows for the timely replacement or re-
furbishment of the component, ensuring its reuse in the
Carbon Footprint Reduction = Energy Saved×Carbon Intensity production cycle and reducing the need for virgin materials.
(1)
For example, if the AI-driven PdM system saves 1000 kWh C. Case Study: Carbon-Aware Maintenance Scheduling in
of energy annually and the carbon intensity of the energy Heavy Industries
source is 0.5 kg CO2 /kWh, the carbon footprint reduction A practical example of carbon-aware maintenance schedul-
would be: ing can be seen in heavy industries, where the environmental
impact of maintenance activities is significant. The KYKLOS
1000 × 0.5 = 500 kg CO2 reduction annually (2) 4.0 project provides a case study involving a shipyard crane in
Astander, Spain. The crane’s data was collected over a year,
This reduction is significant, especially when scaled across
and the toolkit was used to predict anomalies and schedule
multiple machines and industries.
maintenance activities in a carbon-aware manner.
Figure 7 illustrates the correlation between different features
The toolkit employed an autoencoder based on Long Short-
of the translation system. Strong positive correlations (yellow)
Term Memory (LSTM) cells to learn the normal behavior of
indicate synchronized behavior between features, while strong
the crane. The evaluation algorithm measured the error of
negative correlations (dark blue) suggest inverse relationships.
predicted values using the squared error formula:
This analysis helps identify redundant features and key rela-
tionships for optimizing energy consumption and maintenance
errori = (ytruei − ypredictedi )2 (5)
scheduling.
B. Predictive Models for Material Degradation A threshold was defined from the training error, and anoma-
Reuse in a Circular Economy lies were detected when the error exceeded this threshold. The
anomaly score was calculated using a windowing technique:
Predictive models for material degradation and reuse are
essential components of a circular economy (CE). These
t
models enable industries to predict the remaining useful life 1 X
anomaly scoret = low level anomaliest (6)
(RUL) of materials and components, facilitating their reuse, WS
j=t−WS
refurbishment, or recycling. The KYKLOS 4.0 project has
developed a deep learning-based toolkit that monitors the Where WS is the window size. For example, if the window
conditions of machines and predicts potential breakdowns size is 10 and 3 anomalies are detected in the last 10 timesteps,
or RUL. This toolkit has been tested in various industrial the anomaly score would be:
 B. Benchmarking for PdM through ML Libraries
3 Benchmarking ML-based PdM systems is crucial for eval-
anomaly scoret = = 0.3 (7)
10 uating their performance and comparing them with traditional
This score indicates the likelihood of an anomaly, allowing approaches. The study by Morovati et al. (2023) introduces de-
for maintenance to be scheduled before a failure occurs. By fect4ML, a faultload benchmark for ML-based systems, which
reducing unplanned downtime and optimizing maintenance includes 100 bugs from GitHub and Stack Overflow (SO)
schedules, the carbon footprint associated with maintenance related to TensorFlow and Keras frameworks. This benchmark
activities is minimized. addresses several challenges in Software Reliability Engineer-
ing (SRE) of ML-based systems, including fast changes in ML
VII. C ASE S TUDIES & B ENCHMARKING frameworks, code portability, bug reproducibility, and lack of
detailed information on bugs.
A. Failure Analysis of ML-Based PdM in Production Environ- The defect4ML benchmark is designed to satisfy stan-
ments dard benchmark criteria, including relevance, reproducibility,
fairness, verifiability, and usability. The benchmark includes
Predictive Maintenance (PdM) using Machine Learning
bugs from different versions of ML frameworks, ensuring
(ML) has gained significant traction in industrial settings, par-
that it remains relevant as frameworks evolve. For example,
ticularly in semiconductor manufacturing, where unscheduled
the benchmark includes bugs from TensorFlow versions 1.10
downtime can lead to substantial productivity losses. In the
to 2.0, which experienced significant changes in 26% of
case study presented by Kalir et al. (2023), the authors high-
operations. This ensures that the benchmark can be used to
light the application of Long Short-Term Memory (LSTM) and
evaluate ML-based PdM systems across different framework
Random Forest (RF) models to address unscheduled down-
versions.
time and process time (PT) variation in a complex chamber
The benchmark also addresses the challenge of bug repro-
processing tool. The study demonstrates that ML-based PdM
ducibility by providing complete information on dependencies,
can significantly reduce unscheduled downtime by detecting
Python versions, and datasets required to reproduce each bug.
tool states and attributes, thereby improving tool availability
This is particularly important for ML-based PdM systems,
and run-rate.
where the behavior of the system can vary significantly
One of the key findings from the case study is the reduction depending on the input data and model configuration. The
in the standard deviation of weekly availability during Cold benchmark includes 62 bugs from GitHub and 38 from SO, all
Trap (CT) Preventive Maintenance (PM) weeks from 5.9% of which are fully reproducible. This allows researchers and
to 3.2%. This improvement was achieved by using LSTM practitioners to evaluate their PdM systems using real-world
to predict CT pressure trends, allowing engineers to perform bugs, ensuring that their models are robust and reliable.
PM activities more effectively. The Mean Time Between In terms of verifiability, the benchmark includes test cases
Failures (MTBF) also increased from 4.3 days to 16.67 days, for 10 out of 62 GitHub bugs, allowing users to verify the
indicating a substantial reduction in unscheduled downtime. correctness of bug fixes. However, none of the SO bugs
The LSTM model was trained on 27 chamber-years of data, include test cases, highlighting a gap in the current state
which provided a robust dataset for learning the tool’s behavior of ML-based PdM systems. This underscores the need for
and predicting potential failures. better testing practices in ML-based systems, particularly in
However, the study also reveals challenges in implementing production environments where the cost of failure can be high.
ML-based PdM in production environments. One of the pri-
mary issues is the complexity of the tools and the variability C. Cost-Benefit Analysis of AI-Driven PdM vs. Traditional
in their behavior. The semiconductor manufacturing process Approaches
involves thousands of process steps and hundreds of tools, The cost-benefit analysis of AI-driven PdM versus tradi-
each with its own set of parameters and control limits. This tional approaches is a critical aspect of evaluating the ef-
complexity makes it difficult to generalize ML models across fectiveness of ML-based maintenance strategies. In the semi-
different tools and processes. Additionally, the study notes that conductor manufacturing case study, the implementation of
traditional PdM methods, such as XGBoost, were insufficient LSTM and RF models led to significant improvements in tool
for the complex chamber tools, necessitating the use of more availability and run-rate, resulting in a 4% increase in capacity.
advanced ML techniques like LSTM. This improvement directly translates to cost savings, as higher
Another challenge is the need for large amounts of high- tool availability reduces the need for additional tools to meet
quality data to train ML models effectively. In the case production targets.
study, the LSTM model required 27 chamber-years of data to The study also highlights the cost savings associated with
achieve accurate predictions. This highlights the importance of reducing unscheduled downtime. By extending the MTBF
data collection and preprocessing in ML-based PdM. Without from 4.3 days to 16.67 days, the ML-based PdM system
sufficient data, ML models may fail to capture the underlying significantly reduced the frequency of unscheduled downtime
patterns in tool behavior, leading to inaccurate predictions and events. This reduction in downtime not only improves produc-
ineffective maintenance strategies. tivity but also reduces the costs associated with emergency
maintenance, such as overtime labor and expedited parts
delivery.
However, the implementation of AI-driven PdM systems
also incurs costs, particularly in terms of data collection,
model training, and maintenance. The LSTM model in the
case study required 27 chamber-years of data for training,
which represents a significant investment in data collection
and preprocessing. Additionally, the complexity of ML models
like LSTM and RF requires specialized expertise, which can
increase labor costs.
To quantify the cost-benefit ratio, we can use the following
formula:

Total Benefits
Cost-Benefit Ratio = (8)
Total Costs
In the case study, the total benefits include the increase
in tool availability (from 84.5% to 87.1%) and the reduction
in unscheduled downtime (MTBF increase from 4.3 days to
16.67 days). The total costs include the investment in data
collection, model training, and maintenance. Assuming that
the cost of unscheduled downtime is $X per hour and the cost Fig. 8. Architecture of next-generation Digital Twin integrating AI and
of implementing the ML-based PdM system is $Y, the cost- physics-based models. The framework shows the interaction between physical
space (sensors, actuators) and virtual/digital space, unified through AI-ML
benefit ratio can be calculated as follows: and big data analytics components for comprehensive predictive maintenance
capabilities.

(87.1% − 84.5%) × Production Value

+ (16.67 − 4.3) × Cost of Downtime
Cost-Benefit Ratio =
Y A. Next-Generation Digital Twins: AI and Physics-Based
(9) Model Integration
This formula provides a quantitative measure of the cost-
effectiveness of AI-driven PdM compared to traditional ap- Digital Twins (DTs) are becoming a foundational technol-
proaches. In the case study, the cost-benefit ratio is likely ogy for advanced PdM by combining real-time data analytics
to be favorable, given the significant improvements in tool with physics-based simulations [20]. While traditional PdM
availability and MTBF. relies on either AI-driven analytics or physics-based models,
next-generation DTs integrate both to improve precision and
In conclusion, the case studies and benchmarking efforts
reliability.
highlight the potential of AI-driven PdM to improve main-
By fusing AI with physics-based models, DTs can simulate
tenance strategies in production environments. However, the
operational conditions in real-time and optimize maintenance
implementation of these systems requires careful consideration
schedules, as shown in Figure 8. Research shows that hybrid
of the associated costs and challenges, particularly in terms
AI-physics DTs increase failure prediction accuracy by 35%
of data collection, model complexity, and testing practices.
compared to standalone AI models [26].
By addressing these challenges, organizations can realize the
1) Enhancing Failure Prediction Accuracy: By fusing AI
full potential of AI-driven PdM and achieve significant cost
with physics-based models, DTs can simulate operational
savings and productivity improvements.
conditions in real-time and optimize maintenance schedules.
Research shows that hybrid AI-physics DTs increase failure
VIII. F UTURE R ESEARCH D IRECTIONS IN P REDICTIVE prediction accuracy by 35% compared to standalone AI mod-
M AINTENANCE (P D M) els [26]. This improvement stems from the ability of physics-
based models to generalize across unseen failure scenarios,
Predictive Maintenance (PdM) is evolving rapidly, driven
reducing dependence on historical data.
by advanced technologies like digital twins (DTs), artificial
Mathematically, the PdM process in a hybrid DT system
intelligence (AI), and blockchain [1]. Future research in PdM
can be represented as:
aims to enhance prediction accuracy, optimize maintenance
scheduling, and ensure secure data sharing. This section
outlines three major research directions: AI-integrated digital P (f |t, D) = α × PAI (f |t, D) + (1 − α) × Pphy (f |t, D) (10)
twins, AI-driven automated scheduling, and federated AI with
blockchain for privacy-preserving PdM. Where:
 • P (f |t, D) is the probability of failure at time t given 1) Federated Learning for Collaborative PdM: Federated
operational data D. learning (FL) allows multiple industrial sites to train AI mod-
• PAI (f |t, D) represents the AI-driven failure prediction. els collaboratively without sharing raw data. This approach
• Pphy (f |t, D) represents the physics-based model estima- mitigates data breach risks while maintaining model accuracy.
tion. Key advantages of FL in PdM include:
• α is a weighting factor optimized for prediction accuracy. • 15% higher model accuracy compared to isolated train-
2) Case Study: Aerospace Industry Implementation: In the ing.
aerospace sector, next-gen DTs for turbofan engines have led • 50% reduction in data transmission costs.
to a 20% reduction in unscheduled maintenance and a 15% • Enhanced compliance with data privacy regulations
increase in component lifespan. This translates to an annual (GDPR, HIPAA) [1].
saving of $2 million per aircraft [27]. Such implementations The FL model update process is defined as:
demonstrate the practical advantages of AI-physics hybrid
N
models in critical industries. X
wt+1 = wt − η ∇Li (wt ) (12)
i=1
B. AI-Driven Automated Maintenance Scheduling
Where:
AI-driven automated maintenance scheduling is a crucial t
• w is the global model at iteration t.
enabler of zero-downtime manufacturing. Unlike traditional • η is the learning rate.
static, rule-based methods, AI-driven scheduling dynamically • Li represents the local loss function at site i.
adapts to real-time sensor data and operational constraints [28]. • N is the number of participating sites.
1) Optimizing Scheduling with Deep Reinforcement Learn- 2) Blockchain for Data Integrity and Security: Blockchain
ing (DRL): AI-based scheduling uses deep reinforcement technology ensures tamper-proof PdM data sharing through a
learning (DRL) to predict optimal maintenance windows, decentralized, immutable ledger. In a recent study, integrating
minimizing downtime while ensuring asset reliability. A study blockchain with PdM systems resulted in:
on manufacturing flowshop scheduling found that AI-driven
• 98% reduction in data integrity concerns.
models:
• 75% fewer unauthorized access incidents.
• Reduced maintenance-induced downtime by 30%. • Secure multi-party model training with cryptographic
• Improved production efficiency by 25%. guarantees [1].
• Lowered operational costs by 20% [28]. A blockchain ledger entry can be represented as:
The scheduling optimization can be expressed as:
H(Bk ) = h(Bk−1 , Dk ) (13)
T
Where:
X
min [Cm (t) + Ld (t)] (11)
t=0 • H(Bk ) is the hash of the current block.
• h is a cryptographic hash function.
Where: • Bk−1 is the previous block.
• Cm (t) represents the maintenance cost at time t. • Dk is the data stored in the current block.
• Ld (t) is the loss due to downtime at time t. 3) Case Study: Smart Factory Deployment: In an automo-
• T is the scheduling horizon. tive factory, a federated AI + blockchain PdM system achieved
By training AI models on historical maintenance data and the following outcomes:
real-time logs, companies can achieve near-optimal schedul- • 30% faster failure detection.
ing, minimizing human intervention and maximizing opera- • 20% improvement in supply chain efficiency.
tional uptime. • $1.2 million annual cost savings in maintenance opera-
2) Energy-Efficient Maintenance Planning: Energy effi- tions [1].
ciency is another critical dimension of AI-driven scheduling.
IX. C ONCLUSION
Studies show that energy-aware maintenance scheduling re-
duces energy consumption by 18% in industrial PdM applica- This paper presents a comprehensive analysis of AI and
tions, leading to annual savings of $500,000 per factory [28]. ML’s transformative role in predictive maintenance within
Industry 4.0. By leveraging advanced techniques such as
meta-learning, federated learning, generative AI, and physics-
C. Federated AI and Blockchain for Secure PdM
informed neural networks, we address critical challenges in
With the rise of data-driven PdM, ensuring privacy and data PdM, including data scarcity, model interpretability, and com-
security is paramount. Federated AI and blockchain technolo- putational efficiency [21].
gies offer robust solutions for privacy-preserving collaborative Moreover, the adoption of Explainable AI (XAI) techniques
PdM without exposing sensitive industrial data [1]. enhances trust and transparency, fostering broader industrial
adoption. Case studies underscore tangible benefits, from mini- [16] Q. Yang, Y. Liu, T. Chen and Y. Tong, ”Federated Machine Learning:
mizing unplanned downtime to optimizing equipment lifespan. Concept and Applications,” ACM Transactions on Intelligent Systems
and Technology, vol. 10, no. 2, pp. 1-19, 2019.
However, challenges such as data quality, model overfitting, [17] K. Bonawitz et al., ”Towards Federated Learning at Scale: System
and the lack of standardized AI-physics integration persist, Design,” in Conference on Systems and Machine Learning, pp. 1-15,
alongside pressing ethical and regulatory concerns. 2019.
[18] T. Li, A. K. Sahu, A. Talwalkar and V. Smith, ”Federated Learning:
Emerging solutions, including federated learning and Challenges, Methods, and Future Directions,” IEEE Signal Processing
blockchain, offer pathways for secure, privacy-preserving col- Magazine, vol. 37, no. 3, pp. 50-60, 2020.
laborative learning. Future research should focus on AI-driven [19] M. Du, K. Wang, Y. Chen, X. Wang and Y. Sun, ”Big Data Privacy
Preserving in Multi-Access Edge Computing for Heterogeneous Internet
digital twins, quantum computing for failure simulations, and of Things,” IEEE Communications Magazine, vol. 56, no. 8, pp. 62-67,
scalable federated AI systems. Advancing these frontiers will 2018.
unlock AI’s full potential in predictive maintenance, driving ef- [20] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo and J. Zhang, ”Edge
Intelligence: Paving the Last Mile of Artificial Intelligence With Edge
ficiency, sustainability, and resilience in industrial operations. Computing,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1738-1762,
2019.
[21] J. Chen and X. Ran, ”Deep Learning With Edge Computing: A Review,”
R EFERENCES Proceedings of the IEEE, vol. 107, no. 8, pp. 1655-1674, 2019.
[22] Y. Mao, C. You, J. Zhang, K. Huang and K. B. Letaief, ”A Survey
[1] Y. Lu, X. Huang, Y. Dai, S. Maharjan and Y. Zhang, ”Communication- on Mobile Edge Computing: The Communication Perspective,” IEEE
Efficient Federated Learning and Permissioned Blockchain for Digital Communications Surveys Tutorials, vol. 19, no. 4, pp. 2322-2358, 2017.
Twin Edge Networks,” in IEEE Internet of Things Journal, vol. 8, no. [23] S. Saeed, et al., ”Explainable AI in Industrial Predictive Maintenance:
4, pp. 2276-2288, Feb. 15, 2021. A Systematic Review,” IEEE Access, vol. 11, pp. 12345-12367, 2023.
[2] K. Zhang, Y. Zhu, S. Maharjan and Y. Zhang, ”Edge Intelligence and [24] R. Gawde, et al., ”SHAP and LIME Applications in Industrial Mainte-
Blockchain Empowered 5G Beyond for Industrial Internet of Things,” nance: A Comparative Study,” Journal of Intelligent Manufacturing, vol.
IEEE Network, vol. 33, no. 5, pp. 12-19, 2019. 35, no. 1, pp. 78-95, 2024.
[3] J. Park, S. Samarakoon, M. Bennis and M. Debbah, ”Wireless Network [25] M. Zolanvari, et al., ”Attention-Based Models for Real-Time Fault De-
Intelligence at the Edge,” Proceedings of the IEEE, vol. 107, no. 11, pp. tection in Manufacturing,” IEEE Transactions on Industrial Informatics,
2204-2239, 2019. vol. 19, no. 8, pp. 5671-5683, 2023.
[4] W. Y. B. Lim et al., ”Federated Learning in Mobile Edge Networks: A [26] D. McWilliam, et al., ”Hybrid AI-Physics Digital Twins for Smart
Comprehensive Survey,” IEEE Communications Surveys Tutorials, vol. Manufacturing: A Review,” Journal of Manufacturing Systems, vol. 55,
22, no. 3, pp. 2031-2063, 2020. pp. 281-297, 2020.
[5] Y. Dai, D. Xu, S. Maharjan, Z. Chen and Q. He, ”Blockchain and Deep [27] S. Rathore, et al., ”Digital Twin Technology in Industrial Systems: A
Reinforcement Learning Empowered Intelligent 5G Beyond,” IEEE Review,” IEEE Internet of Things Journal, vol. 8, no. 12, pp. 9789-9803,
Network, vol. 33, no. 3, pp. 10-17, 2019. 2021.
[6] M. S. H. Abad, E. Ozfatura, D. Gunduz and O. Ercetin, ”Hierarchical [28] N. Danishvar, et al., ”AI-Driven Scheduling Optimization in Smart
Federated Learning Across Heterogeneous Cellular Networks,” in IEEE Manufacturing,” Computers in Industry, vol. 124, pp. 103324, 2021.
International Conference on Communications, pp. 1-6, 2020. [29] X. Chen, et al., ”Federated Learning for Industrial IoT: A Comprehensive
[7] Y. Lu, X. Huang, K. Zhang, S. Maharjan and Y. Zhang, ”Blockchain Survey,” IEEE Communications Surveys Tutorials, vol. 24, no. 2, pp.
Empowered Asynchronous Federated Learning for Secure Data Sharing 1-28, 2022.
in Internet of Vehicles,” IEEE Transactions on Vehicular Technology, [30] J. Wu, et al., ”Meta-Learning for Few-Shot Fault Diagnosis Under
vol. 69, no. 4, pp. 4298-4311, 2020. Different Working Conditions,” IEEE Transactions on Industrial Elec-
[8] H. Kim, J. Park, M. Bennis and S. Kim, ”Blockchained On-Device tronics, vol. 67, no. 11, pp. 9775-9784, 2020.
Federated Learning,” IEEE Communications Letters, vol. 24, no. 6, pp. [31] Y. Fang, et al., ”Deep Learning for Fault Diagnosis: A Review,” IEEE
1279-1283, 2020. Transactions on Neural Networks and Learning Systems, vol. 32, no. 9,
[9] Y. M. Saputra, D. T. Hoang, D. N. Nguyen, E. Dutkiewicz, M. D. Mueck pp. 3992-4008, 2021.
and S. Srikanteswara, ”Energy Demand Prediction with Federated Learn- [32] J. Zhu, et al., ”Physics-Informed Neural Networks for Real-Time Moni-
ing for Electric Vehicle Networks,” in IEEE Global Communications toring in Smart Manufacturing,” Journal of Manufacturing Systems, vol.
Conference, pp. 1-6, 2019. 58, pp. 262-276, 2021.
[10] J. Kang, Z. Xiong, D. Niyato, S. Xie and J. Zhang, ”Incentive Mecha- [33] F. Longo, et al., ”Trust and Adoption of AI-Based Maintenance Systems:
nism for Reliable Federated Learning: A Joint Optimization Approach An Industrial Perspective,” Reliability Engineering System Safety, vol.
to Combining Reputation and Contract Theory,” IEEE Internet of Things 235, pp. 109012, 2024.
Journal, vol. 6, no. 6, pp. 10700-10714, 2019. [34] E. Ortigossa, et al., ”Regulatory Compliance in AI-Driven Industrial
[11] X. Wang, Y. Han, C. Wang, Q. Zhao, X. Chen and M. Chen, ”In-Edge Maintenance,” IEEE Transactions on Industrial Informatics, vol. 20, no.
AI: Intelligentizing Mobile Edge Computing, Caching and Communica- 2, pp. 1123-1134, 2024.
tion by Federated Learning,” IEEE Network, vol. 33, no. 5, pp. 156-165, [35] H. Bolandi, et al., ”PINN-Stress: Physics-Informed Neural Networks
2019. for Structural Health Monitoring,” Engineering Structures, vol. 284, pp.
[12] M. Chen, O. Semiari, W. Saad, X. Liu and C. Yin, ”Federated Echo State 115217, 2023.
Learning for Minimizing Breaks in Presence in Wireless Virtual Reality [36] K. Zhang, et al., ”Edge Computing in Industrial IoT: A Review of Real-
Networks,” IEEE Transactions on Wireless Communications, vol. 19, Time Processing Requirements,” IEEE Internet of Things Journal, vol.
no. 1, pp. 177-191, 2020. 8, no. 4, pp. 2424-2440, 2021.
[13] S. R. Pandey, N. H. Tran, M. Bennis, Y. K. Tun, A. Manzoor and [37] L. Wang, et al., ”Smart Factory Implementation Challenges: A Compre-
C. S. Hong, ”A Crowdsourcing Framework for On-Device Federated hensive Review,” Journal of Manufacturing Systems, vol. 62, pp. 478-
Learning,” IEEE Transactions on Wireless Communications, vol. 19, 495, 2022.
no. 5, pp. 3241-3256, 2020.
[14] Y. Zhao, J. Zhao, L. Jiang, R. Tan, D. Niyato, Z. Li, L. Lyu and Y.
Liu, ”Privacy-Preserving Blockchain-Based Federated Learning for IoT
Devices,” IEEE Internet of Things Journal, vol. 8, no. 3, pp. 1817-1829,
2021.
[15] H. B. McMahan, E. Moore, D. Ramage, S. Hampson and B. A. y Arcas,
”Communication-Efficient Learning of Deep Networks from Decentral-
ized Data,” in International Conference on Artificial Intelligence and
Statistics, pp. 1273-1282, 2017.