A Comprehensive Review of Machine Learning Techniques for

Condition-Based Maintenance
Tyler Ward1, Kouroush Jenab2, Jorge Ortega-Moody3, Selva Staub4

1,2,3
Morehead State University, Morehead, Kentucky, 40531, United States
[email protected]
[email protected]
[email protected]
4
Bandırma Onyedi Eylül University, Bandırma, Balıkesir, Turkey
[email protected]

ABSTRACT driven by advancements in technology and the increased

importance of machinery in both military and industrial
While most industrial maintenance strategies are centered on
applications.
optimizing machine runtime and cost reduction, the
condition-based maintenance (CBM) strategy distinguishes Condition-based maintenance (CBM) is a maintenance
itself from others in its use of real-time operational data from strategy that seeks to extend the life of equipment and ensure
machines to help engineers make informed decisions. The operational effectiveness through the real-time assessment of
introduction of machine learning (ML) into a CBM strategy the operational condition of a machine. The primary objective
can increase its effectiveness, enabling more accurate of CBM is to perform maintenance only when warranted,
predictions and making the decision-making process more thereby adverting unnecessary downtime and reducing
efficient. In this review paper, we seek to provide a maintenance expenditures (Teixeria, Lopes, & Braga, 2020).
comprehensive overview of the role ML plays in modern The development of sophisticated diagnostic tools like
CBM systems, beginning by outlining the core concepts and vibration analysis and thermography in the 1970s and 80s
historical development of CBM and briefly introducing formed the basis for the earliest version of the CBM strategy,
various ML techniques being employed in industry today. alongside the development of the field of reliability
We then review numerous real-world cases where ML-based engineering, which emphasized system reliability and
CBM systems have been implemented and discuss some of maintenance optimization.
the technological, human, and ethical challenges faced by
The advent of computers and sensor technology towards the
organizations seeking to integrate sophisticated ML models
end of the 20th century significantly enhanced the CBM
into existing CBM systems. We end by highlighting some of
strategy by enabling real-time data collection and analysis.
the current limitations of ML-based CBM systems, paving
This era marked the start of the widespread integration of
the way for a discussion on emerging trends and future
CBM across various industries, as its potential to enhance
research directions in this area.
efficiency and reduce maintenance costs was realized. The
rise of Industry 4.0 has further propelled CBM to
1. INTRODUCTION
prominence, with the rapid growth of artificial intelligence
Prior to the Industrial Revolution, maintenance was (AI) technology revolutionizing the way that data collected
performed primarily reactively, occurring only after a from machines can be analyzed and interpreted, empowering
machine failed. As machinery became more complex, there CBM systems with impressive abilities to predict failures in
was a gradual shift toward more preventative maintenance a machine long before they occur, allowing maintenance
strategies, although these approaches were still primarily teams to deal with an issue before it ever becomes a problem
based on scheduled maintenance intervals. Following World (Fernandes, Reis, Melão, Teixeira, & Amorim, 2021).
War II, modern maintenance strategies began to take shape,
Machine learning (ML) is a subfield of AI that deals with the
_____________________ development of study of algorithms that can learn from data
Tyler Ward et al. This is an open-access article distributed under the terms
and generalize to unseen data. By leveraging patterns and
of the Creative Commons Attribution 3.0 United States License, which
permits unrestricted use, distribution, and reproduction in any medium, insights extracted from vast amounts of operational data, ML
provided the original author and source are credited. algorithms can identify subtle indicators of equipment health
https://doi.org/10.36001/IJPHM.2024.v15i2.3850

International Journal of Prognostics and Health Management, ISSN 2153-2648, 2024 1

 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

that may elude even the most experienced human experts, Adryan & Sastra 2021 Focused on the use of ML
making them a valuable asset in the predictive maintenance for PdM of aircraft engines
(PdM) toolkit (Carvalho et al., 2019). However, integrating
ML into existing CBM systems requires a deep Chatterjee & 2021 Focused on the use of AI for
understanding of the technology's capabilities and Dethlefs the operations and
limitations. maintenance of wind
turbines
ML algorithms offer several advantages over conventional
statistical and reliability models for CBM applications. They Elbouchikhi, Zia, 2021 Focused on signal
can automatically learn complex patterns and relationships Benbouzid, & El processing and ML for
from large, high-dimensional datasets without the need for Hani intelligent grid condition
explicit programming or model specification. This data- monitoring
driven approach allows ML models to capture intricate Jourdan, Longard, 2021 Focused on datasets for
dependencies and non-linear interactions that may be Biegel, & intelligent maintenance and
difficult to represent using traditional parametric models. Metternich corresponding use cases
Additionally, ML algorithms can adapt and improve their
predictions as new data becomes available, enabling Leukel, González, 2021 Only focused on the failure
continuous learning and refinement of the models. This & Riekert prediction aspect of CBM
adaptability is particularly valuable in dynamic industrial Nacchia, Fruggiero, 2021 Primarily focused on
environments where operating conditions and failure modes Lambiase, & identifying trends and gaps
may evolve over time. Specific ML techniques such as deep Bruton as opposed to in-depth
learning (DL) have demonstrated remarkable performance in understandings of specific
tasks like image recognition and signal processing, making application
them well-suited for analyzing sensor data and extracting
relevant features for CBM. Siang, Ahamd, & 2021 Only focused on the
Abidin anomaly detection aspect of
Table 1 shows a breakdown of existing review articles on CBM and only concentrate
CBM and tangentially related topics. Our analysis identified on tiny ML methods.
a need for a review article focusing on a broad overview of
the use of ML in all aspects of CBM. Current review articles Afridi, Ahmad, & 2022 Only focused on AI-based
largely focus on the use of ML in CBM for specific use cases, Hassan PdM of renewable energy
such as CBM of rolling element bearings, or individual systems
aspects of CBM, such as anomaly detection, or they only
Alsumaidaee et al. 2022 Only focused on the fault
focus on one type of ML applied to CBM tasks such as
detection aspect of CBM
reinforcement learning.
specifically a medium-
Table 1 Analysis of existing literature voltage switchgear.
Authors Year Limitations Ciaburro 2022 Only focused on the fault
detection aspect of CBM
Ellefsen, Æsøy, 2019 Focused on DL methods for
Ushakov, & Zhang prognostics and health Drakaki, Karnavas, 2022 Only focused on the use of
management (PHM) of Tziafettas, ML and DL for PdM of
autonomous ships Linardos, & induction motors.
Tzionas
Çınar et al. 2020 The focus is on using ML-
based PdM for sustainable Fernandes, 2022 Only focused on the fault
smart manufacturing Corchado, & diagnosis and prognosis
Marreiros aspect of CBM
Namuduri, 2020 Focused on DL methods for
Narayanan, PdM of electrochemical Ferreira & 2022 Only focused on the
Davuluru, Burton, sensors Gonçalves remaining useful life (RUL)
& Bhansali prediction aspect of CBM
Singh, Azamfar, Li, 2020 Focused on ML-based Nor, Kassim, 2022 Only focused on ML-based
& Lee PHM of rolling element Minhat, & Ya’acob PdM techniques for a
bearings nuclear reactor cooling
system.

2
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

Sanzana, Maul, 2022 Only focused on DL such as data handling and feature engineering, practical
Wong, Abdulrazic, applications in facility applications, and integration challenges. By addressing these
& Yip management and topics, the article seeks to spark a discussion on potential
maintenance for heating, future advancements in the field of maintenance engineering,
ventilation, and air focusing on improving efficiency, reliability, and the overall
conditioning effectiveness of maintenance strategies.
Serradilla, Zugasti, 2022 Only focused on DL models
2. REVIEW METHODOLOGY
Rodriguez, & for PdM
Zurutuza The first stage of our review process was to understand the
overall trends in the literature surrounding publications
Sharma, Mittal, & 2022 The focus on the regarding ML-based CBM systems. To acquire this data, we
Soni interpretability of ML queried an online research database, Dimensions, consisting
methods in CBM of over 140 million journal and conference articles.
Campos Olivares, 2023 This paper gives a general Dimensions supports Boolean queries in its search
Carrasco Muñoz, overview of ML for PdM functionality, allowing researchers to refine and narrow the
Mazzoleni, but is lacking in detail in search results.
Ferramosca, & terms of real-world use Our query to the Dimensions database was “condition-based
Luque Sendra cases and future directions. maintenance AND machine learning,” meaning that we were
Chen, Fu, Zheng, 2023 Only focused on the role of only searching for articles that contained both keywords. We
Tao, & Liu ML in digital twins for PdM refined the query further by limiting the search only to return
articles that included both of the keywords in the title and
Kumar, Khalid, & 2023 Only focused on PHM of abstract of the article, as well as imposing the limitation that
Kim rotating machinery of only articles published between 2003 and 2023 should be
industrial robots considered. This led to 1,705 articles that matched our criteria
Ogunfowora & 2023 Only focused on being returned.
Najjaran reinforcement and deep Once we had the initial results, we manually evaluated the
reinforcement learning articles that were returned to ensure relevance to our review
solutions for maintenance focus. After manual review, we removed 68 articles from
Payette & Abdul- 2023 This paper does not go in- consideration. These articles were found not to concern ML-
Nour depth into specific ML based CBM systems, and typically were included in the
applications within CBM results because they used the words “condition” and
“maintenance” frequently in non-industrial contexts, such as
Polverino et al. 2023 This study only reviews 50 patient/human maintenance in articles relevant to the medical
papers field, and plant maintenance in articles related to agriculture.
This removal of articles from the search results led to the final
Saurav, Avesh, 2023 Only focused on ML-based
count of articles being considered 1,637.
Sharma, & Hossain PdM of Indian railways
Once these 1,637 articles were identified, we sought to
Surucu, Gadsden, 2023 Only focused on the
determine the number of publications and citations in this
& Yawney condition monitoring aspect area per year over the 20 years between 2003 and 2023. One
of CBM of the reasons that Dimensions was chosen as the database to
Tama, Vania, Lee, 2023 Only focused on DL query is because it has built-in tools to visualize trends based
& Lim applications for fault on specific metrics. Figure 1 depicts the number of articles
diagnosis of rotating published over these 20 years, and Figure 2 depicts the
machinery using vibration number of times these articles were cited over the same
signals. period.
Gupta et al. 2024 Only focused on ML-based These two figures show that research on ML-based CBM
PdM for electric vehicle systems increased gradually from 2003 to 2016, then rapidly
power electronics increased post-2016, demonstrating increased interest in this
topic in recent years. Several key factors contributed to this
increase in research interest around this time. One of the most
This review article aims to provide a comprehensive critical contributing factors has been the significant
overview of the current landscape of ML-based CBM improvement and cost reduction of Internet of Things (IoT)
systems, covering the processes that enable ML integration, devices and sensors. This has led to a rapid uptick in Internet-

3
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

connected devices, with studies showing that the number of highly effective at pattern recognition, making them
such devices in use approximately doubles every five years, particularly suitable for CBM tasks.
with estimates showing that in 2020, there were 50 billion
As ML models have improved, so has their adoption rate in
IoT devices in use, up from 25 billion in 2015 (Singh, 2023).
various sectors across Industry 4.0 (Jan et al., 2023). As more
In the context of CBM research, this improvement in the
organizations adopt ML-based approaches to CBM and show
capabilities and availability of IoT devices has made it much
increased efficiency, the rate of funding from both
easier and cost-effective to collect vast amounts of real-time
governmental and private sectors for this type of research has
data from machinery and equipment, which is crucial in CBM
increased, which is evident by Figure 3, which depicts a trend
strategies.
line of 317 grants awarded per year between 2003-2023 to
encourage research in the area of ML applications for CBM.
400 This data was obtained using the same search query to
350 Dimensions described earlier in this discussion. In addition
300 to increased funding, collaboration between industry and
250 academia has also risen in recent years, with studies showing
that articles in the area of AI published since 2012 that are a
200
collaborative effort between industry and academia receive
150
more citations and online interest than articles published by
100 industrial or academic researchers individually (Färber &
50 Tampakis, 2023). One last consideration that needs to be
0 addressed to understand the growth in ML-based CBM
research is the wide availability of open-source ML
frameworks and tools for CBM purposes, which has allowed
researchers and practitioners in the field of maintenance
engineering to implement and experiment with advanced ML
Figure 1. Articles published per year with the keywords models more easily than before (Zhao et al. 2020).
“condition-based maintenance” AND “machine learning” in
the title and abstract 60
50
6000
40
5000
30
4000
20
3000
10
2000
0
1000
0

Figure 3. Grants awarded per year sponsoring research into

ML-based CBM
Once we analyzed these trends in ML-based CBM research,
Figure 2. The citation count per year of papers with the
we next defined the scope of our review. We identified
keywords “condition-based maintenance” AND “machine
several areas we wanted to address: a discussion of data
learning” in the title and abstract
handling and feature engineering techniques that facilitate the
This vast amount of real-time data would be useless without use of ML for CBM, an introduction to ML and the
the ability to process it, which leads to a second major techniques that have found applications in CBM, specific
contributing factor to the rise of ML-based CBM systems: real-world use cases where ML was integrated into CBM
increased computational power. In recent years, there have systems, the challenges and limitations of integrating ML
been significant advancements in graphical processing units into CBM systems, and future trends and directions that could
(GPUs) and cloud computing technologies. This be explored within the area of ML-based CBM. Once this
technological advancement has enabled the rapid processing scope was defined, we switched the academic database we
of large datasets of condition information from machinery queried from Dimensions to Google Scholar due to the
and equipment. It has enabled the integration of complex ML increased size of the Google Scholar database and its more
models, particularly DL models (Jauro et al., 2020), that are expansive search capabilities. We then queried the database

4
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

using keywords relevant to each of the areas we wanted to Technologies and tools
address, as well as imposing the limitation that the
publication date of the articles must fall between 2018 and Sensor, industrial sensors, Internet of
2024 to identify the 71 articles that make up our reference Sensors and IoT Things (IoT), Power IoT, IoT, edge
list. A bar chart showing the distribution in terms of computing, wireless transmission
publication year of the articles in the reference list is shown Neural networks, convolutional
in Figure 4. neural networks, recurrent neural
Computational
From the reviewed literature, we found that a total of 221 network, LSTM autoencoder,
models and
unique keywords classified the articles. We analyzed these generative adversarial networks,
algorithms
keywords and categorized them into ten distinct categories random forest, support vector
to declare their relevance to ML-based CBM systems. This regression, deep neural networks
categorization is shown in Table 2. Feature extraction, feature selection,
Data processing feature importance, principal
35
and analysis components analysis, data
30 techniques preparation, data imbalance,
25 oversampling
20 Data Blockchain, digital twin, knowledge
management and base, application server, private
15
platforms LPWAN, LoRaWAN gateway
10
Applications and systems
5
0 Industrial system, smart factory,
Industrial and
2018 2019 2020 2021 2022 2023 2024 industrial automation, industrial
manufacturing
manufacturing, industrial robots
Figure 4. Number of publications per year in the references Rotating machines, pumps, hydraulic
Table 2 Categorization of unique keywords from the Vehicles and systems, electrical machines,
literature machinery bearings, ball screw drives, aircraft,
trucks, buses
Category Keywords
Energy and Upstream oil & gas, energy supply,
Fundamental concepts utilities wind turbine
Artificial intelligence, machine Infrastructure Railway infrastructure, high-speed
learning, deep learning, supervised and railway (HSR), switches and
Machine learning transportation crossings, vehicles
learning, unsupervised learning,
reinforcement learning
Maintenance tasks and objectives
Predictive maintenance, preventative
maintenance, prescriptive analytics, Anomaly detection, fault detection,
predictive analytics, condition-based Detection and defect inspection, fault diagnosis,
Maintenance diagnosis motor fault detection, failure
maintenance, condition-based
types prediction, degradation
maintenance (CBM), predictive
health maintenance, proactive Reliability, reliability assessment,
maintenance Reliability and
remaining useful life, health
assessment
Maintenance, maintenance indicator, lifetime prediction
Maintenance management, optimal maintenance Prediction methods, prediction,
strategies planning, smart plant maintenance Prediction and
multi-step multivariate time series
system forecasting
forecasting
Statistics, Bayesian inference, Techniques and methodologies
Statistical
statistical methods, data science, data
methods and Analysis and
analytics, factor analysis, data-driven Regression, Bayesian regression,
analysis optimization
decision-making particle swarm optimization, gradient

5
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

boosting, hyperparameter In addition to the implementation of AI, there is also an

optimization emphasis on harnessing the rapid growth of sensor data and
industrial IoT (IIoT) technologies to enable real-time
Transfer learning, self-supervised monitoring, data collection, and interconnected intelligent
Learning and
learning, continual learning, domain systems for CBM. There is also a drive to develop robust data
adaptation
adaptation management pipelines to handle large, complex datasets and
Inspection and Image restoration, image processing, address challenges like noise, missing values, and class
quality control defect classification imbalance through data preparation techniques and digital
twins. The literature exhibits methodological diversity, with
Theoretical and methodological foundations researchers combining traditional statistical methods and
optimization algorithms with cutting-edge ML architectures
Statistical and Canonical link function, Hurst
like generative adversarial networks (GANs) and
mathematical exponent, T2 chart, Weibull failure
reinforcement learning models. Applications range from
models rate function
static machinery to dynamic assets like rotating equipment,
Wavelet spectrum analysis, vehicles, aircraft, and wind turbines.
Analytical compressed sensing, multivariate Along with improving prediction accuracy and lead times,
techniques signal processing, principal researchers are increasingly focusing on broader goals like
component analysis reliability assessment, life cycle management, and
Decision support, Bayesian decision maximizing asset utilization while minimizing downtime and
Decision-making theory, partially observable Markov maintenance costs over the entire operational lifecycle. The
and planning decision process, planning under emphasis on RUL prediction and lifecycle perspectives
uncertainty indicates a holistic approach to maintenance optimization
rather than solely focusing on fault detection. Additionally,
Emerging trends and considerations there is a growing interest in enhancing model interpretability
through explainable AI (XAI), knowledge transfer via
Generative adversarial network
AI and advanced transfer learning, and improving data efficiency. Ethical
(GAN), variational autoencoder,
models considerations like transparency and trustworthiness of AI
black box, knowledge reasoning
systems for high-stakes maintenance decisions are also
Standards and Industry 4.0, CPPS, OPERAND, gaining attention.
frameworks micro-world, experimental platform
ML-based CBM research is a rapidly evolving field marked
Transparency, explainable artificial by significant technological innovations, cross-disciplinary
Ethical and convergence, and a strong drive toward intelligent, predictive
intelligence, decision support
societal impacts maintenance systems. As the capabilities of ML models and
systems
IIoT technologies continue to advance, there is immense
potential for optimizing asset performance, operational
Based on these keywords, it is clear that the literature on ML efficiency, and sustainability across industries. The balanced
for CBM reflects an increasingly multidisciplinary and coverage of theory and applications suggests a solid
collaborative approach through the integration of advanced grounding complemented by a solutions-oriented mindset.
analytics, information technology, and data science within The discussion and analysis presented in this section serve as
traditional engineering domains. There is a clear shift from a general overview of the current literature. The following
reactive maintenance strategies to more proactive and sections will expand on this discussion and analysis.
optimized approaches focused on predictive and prescriptive
maintenance. This trend is driven by the need to enhance 3. DATA HANDLING & FEATURE ENGINEERING
asset reliability, reduce downtime, and minimize CBM systems are largely powered by smart devices
maintenance costs across various industries, and is facilitated strategically placed to monitor the condition of a machine.
by the widespread adoption of sophisticated AI models, These devices generate a continuous data stream, capturing
especially neural network (NN) based architectures like deep parameters such as temperature, pressure, vibration, and
neural networks (DNNs), convolutional neural networks more. Because CBM relies so heavily on accurate, real-time
(CNNs), recurrent neural networks (RNNs), and information about the condition of a machine, selecting the
autoencoders. These models offer robust pattern recognition correct data collection methods is necessary to ensure that the
and anomaly detection capabilities crucial for accurate failure data being received from the machine accurately reflects its
prediction, health monitoring, and RUL estimation of operational condition. This section details various methods
industrial assets across sectors like manufacturing, energy, and characteristics of data collection devices. It highlights
transportation, and others. various methods of handling this data and its features to

6
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

ensure proper integration of ML approaches into the CBM fundamental aspect of this process is data imputation, which
process. Figure 5 visualizes multiple aspects of the data involves filling in incomplete data points to maintain the
handling and feature engineering phase of an ML-based overall integrity of the dataset. Various techniques, including
CBM system. mean imputation (Martins, Fonesca, Farinha, Reis, &
Cardoso, 2022), interpolation (Roux, Fang, & Barros, 2022),
or ML-based imputation methods (Ward, Jenab, & Ortega-
Moody, 2023), can be applied to ensure a comprehensive and
representative dataset. Another important step is outlier
detection, aimed at identifying and managing data points that
deviate significantly from the norm. This is necessary to
prevent extreme values from negatively influencing the
predictions of the ML model. Common methods for outlier
detection include statistical approaches such as z-score (Yin,
Liu, Huang, & Pan, 2022) and interquartile range
(Aqueveque, Radrigan, Pastene, Morales, & Guerra, 2021),
as well as ML-based anomaly detection methods (Lourenço
et al., 2023). These preprocessing and cleaning measures
collectively contribute to refining the raw data, enhancing its
quality, and ultimately facilitating the strong performance of
ML models in the CBM process.
Feature selection involves techniques that aid in identifying
the most informative features from the data while discarding
irrelevant or redundant ones. Widely employed methods for
feature selection include mutual information (Hamaide &
Glineaur, 2021), recursive feature elimination (Nemat Saberi,
Belahcen, Sobra, & Vaimann, 2022), and the extraction of
feature importance scores from tree-based models (Allah
Figure 5. Potential components of the data handling and Bukhsh, Saeed, Stipanovic, & Doree, 2019). In instances
feature engineering phase of an ML-based CBM system where temporal patterns characterize equipment data, time-
series features assume paramount importance. The
Sensors are one of the most commonly used devices to enable engineering of time-specific features, such as statistical
CBM, measuring physical characteristics of equipment moments (Fong, 2022), trend analysis (Ngoma, Mativenga,
performance, and often send data in the form of time-series & Pretorius, 2020), and spectral analysis (Bae, Mun, Chang,
measurements. This temporal nature makes sensor data & Vidakovic, 2019), can provide valuable insights into the
particularly useful for performing CBM, as it facilitates the behavior of the equipment over time. Additionally, domain
identification of patterns over time that could be indicative of knowledge and expertise are increasingly important when
potential issues. Sensor data is often augmented through IoT incorporating ML into CBM, as the incorporation of domain-
devices, which can contribute additional real-time specific can lead to the creation of features that capture
information such as the location, environmental conditions, critical aspects of the equipment's health.
and operational parameters of a machine.
Inadequacies in data quality or scarcity of data points can 4. PRINCIPLES OF ML
impede the performance of ML models within the CBM
framework. Addressing data quality is imperative, involving 4.1. Categories of ML Algorithms
measures to rectify issues related to calibration, sensor drift, ML algorithms can be classified into several different
and measurement errors. The regular calibration and categories (Sarker, 2021), each with its own relevance to
maintenance of sensors are crucial to maintaining data CBM. These categories are described below and depicted
accuracy, ensuring that the information used to train ML visually in Figure 6. The next section includes an expanded
models in a CBM system is reliable. When teaching these discussion of each category and specific algorithms
models, a sufficient data volume is essential to enable these employed for use in CBM systems.
models to generalize effectively and produce reliable
predictions across diverse scenarios. 1. Supervised Learning: In the context of CBM, supervised
learning algorithms use labeled training data to learn the
The initial data obtained from sensors and IoT devices may relationship between the input features (e.g., sensor
exhibit characteristics such as noise, incompleteness, or the readings) and the output variable (e.g., equipment
presence of outliers. Addressing these issues is crucial to
preparing the data for effective utilization in ML models. One

7
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

failure). Once trained, supervised learning models can Implementing DL into supervised learning tasks for CBM
predict future failures or estimate the RUL of equipment. can lead to augmented capacity to predict equipment failure
with greater accuracy, as DL approaches are better equipped
to deal with complex, nonlinear relationships in large-scale
maintenance data. For unsupervised tasks, DL can enhance
the accuracy of anomaly and pattern detection by effectively
learning the normal operational baselines of the equipment,
even without the presence of labeled data. When applied to
reinforcement learning tasks, DL can improve the decision-
making process, enabling the development of more
sophisticated and adaptive maintenance strategies that can
dynamically respond to the equipment’s condition and
Figure 6. Categories of ML algorithms and their operational demands.
characteristics
4.2. ML Techniques used for CBM
2. Unsupervised Learning: These algorithms, including
clustering and dimensionality reduction approaches, Employing a variety of ML techniques significantly enhances
detect anomalous behavior or new patterns in equipment CBM's efficacy. These techniques transform raw data into
sensor data that could indicate a potential failure. actionable insights, facilitating proactive maintenance
3. Reinforcement Learning: Although less common in ML- decisions. This section delves into the different ML
based CBM systems than their supervised and techniques applied in CBM and their contributions to the
unsupervised counterparts, reinforcement learning maintenance process. Figure 7 shows several examples of
algorithms can develop maintenance policies where a ML techniques for each category of algorithm.
model learns to make decisions by performing actions Supervised models are prominent in applying ML in the
and assessing the results. CBM process, especially in cases that require the prediction
In recent years, there has been a rapid rise in the adoption of and classification of equipment conditions. Regression
DL approaches within each category, largely driven by analysis is employed to forecast continuous outcomes, such
advancements in NN architectures, layers, objectives, and as the time until failure of the degradation rate of
optimization techniques (Schneider & Vlachos, 2023). DL is components. Techniques such as linear regression (LR)
facilitated by large NNs capable of making accurate data- (BahooToroody, De Carlo, Paltrinieri, Tucci, & Van Gelder,
driven decisions (Kelleher, 2019). Because DL methods are 2020), support vector regression (SVR) (Hong, Xu, & Zhang,
particularly suited to contexts where there is a large amount 2019), and ensemble methods such as random forests (RF)
of complex data (Kelleher, 2019), it should be no surprise that (Fredriksson, 2022) leverage historical data to anticipate
DL has found many applications within the context of CBM. future failure points. On the other hand, classification models
categorize equipment conditions into distinct classes, such as

Figure 7. Examples of ML algorithms used for CBM

8
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

‘normal operation,’ ‘needs inspection,’ or ‘immediate Autoencoders are another powerful tool that can enhance
maintenance required.’ Algorithms such as decision trees anomaly detection and dimensionality reduction tasks.
(DT) (Allah Bukhsh et al., 2019), naïve Bayes (NB) Powered by DL, autoencoders can aid in boosting the
(Maheswari & Umamaheswari, 2020), k-nearest neighbors effectiveness of anomaly detection by learning a compressed
(KNN) (Rathore & Harsha, 2022), and NNs (Berghout, representation of normal operating conditions and
Benbouzid, Muyeen, Bentrcia, & Mouss, 2021) process subsequently reconstructing the input (Ahmad, Styp-
sensor data to identify and signal impending failures based on Rekowski, Nedelkoski, & Kao, 2020). This allows
the predefined classes. autoencoders to pinpoint deviations in the reconstruction,
identifying potential issues and contributing to the early
Specific DL models can be trained using a supervised
identification of abnormalities (Ahmad et al., 2020). When
approach, where extensive, labeled datasets are used to teach
used for dimensionality reduction, autoencoders can encode
these models to recognize intricate patterns and relationships
data into a lower-dimensional space that retains the most
within the data. Examples of these models include DNNs and
significant features necessary for representing the original
CNNs. DNNs extend the capabilities of supervised
dataset (Ahmad et al., 2020).
approaches by providing more nuanced pattern recognition in
complex datasets, often found in the type of sensor data that Further unsupervised DL techniques that have been found to
CBM applications rely on. CNNs, on the other hand, are be useful in CBM include recurrent neural networks (RNNs)
primarily designed for processing and interpreting image and long-short term memory (LSTM) networks. RNNs have
data, such as thermal images (Huerta Herraiz, Pliego a unique recurrent structure capable of preserving previous
Marugán, & García Márquez, 2020) or visual inspections information and passing it into the current calculating
(Doğru, Bouarfa, Arizar, Aydoğan, 2020). CNNs prove process, enabling it to perceive the association of time
effective in detecting signs of wear or damage, offering a sequence data at different time intervals (Cheng, Wang, Wu,
comprehensive means of visual analysis within CBM Zhu, & Lee, 2022). While powerful, RNNs do suffer from a
frameworks. drawback called the vanishing gradient problem. This
problem occurs when the length of a sequence increases; the
Where supervised models aid in the prediction and
gradient magnitude typically decreases alongside it, slowing
classification capabilities of ML-based CBM models,
the training process of the RNN. LSTMs have emerged as a
unsupervised approaches are crucial for uncovering patterns
solution to this problem by introducing various gates to the
and anomalies within sensor data. Clustering techniques,
RNN architecture: the input, output, and forget gates. These
such as k-means (Wang, Liu, Wei, Chen, & Xu, 2020) or
gates enable a “short-term memory” for RNNs (Wang, Bu, &
hierarchical clustering (Zhu & Zhou, 2023) facilitate the
He, 2020). RNNs and LSTMs are particularly adept at
grouping of similar data points. The clusters aid in identifying
handling sequential data like time-series sensor readings, as
patterns associated with various operational modes and
they excel in identifying temporal patterns that serve as
detecting outlier conditions that may signal potential
indicators of equipment health status, enabling a nuanced
malfunctions.
understanding of dynamic operational conditions. This
Advances in DL techniques, such as GANs, have improved makes RNNs and LSTMs handy tools for modern ML-based
the robustness of models in detecting outlier conditions (Lu, CBM systems.
Du, Qian, He, & Wang, 2022). A GAN is an ML framework
Both supervised and unsupervised approaches to
that involves using two competing NNs in a zero-sum game,
incorporating ML into CBM are well-studied. These
where a gain in one NN represents a loss in the other. The use
techniques integrate well into the maintenance process and
of GANs to aid anomaly detection involves the generation of
have become essential components in any ML-based CBM
new synthetic data containing anomalies in an existing
system. However, these techniques represent just a tiny
dataset, which can improve anomaly detection accuracy by
portion of the full body of literature on this topic. More
addressing the common class imbalance problem in anomaly
advanced applications of ML in CBM revolve around
detection, where there are often many more data points
integrating reinforcement learning algorithms into the
representing normal operating conditions compared to
maintenance process.
anomalous ones (Lu et al., 2022).
Reinforcement learning techniques such as Q-learning
Additionally, principal component analysis (PCA) is
(Tanimoto, 2021) and policy gradients (Cheng, Liu, Li, & Li,
commonly employed for dimensionality reduction in CBM
2023) can optimize maintenance schedules and resource
applications (Quatrini, Constantino, Di Gravio, & Patriarca,
allocation by allowing ML models to learn the most effective
2020). PCA enables the grouping of sensor data into principal
actions under specific conditions. Essentially, the models are
components, capturing the most significant variations. This
trained to make decisions that lead to optimal timing for
dimensionality reduction enhances the efficiency of
maintenance interventions, thereby minimizing downtime
subsequent analysis by other ML models, offering a more
and associated costs. The application of reinforcement
streamlined and insightful understanding of equipment
learning in this context enhances the adaptability of
conditions.

9
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

maintenance strategies, allowing for a more dynamic and An example of data preparation and feature selection being
responsive approach to preserving equipment health and used in the real-world ML-based CBM system is in the
functionality. Deep reinforcement learning (Zhang & Si, development of a hybrid data preparation method to predict
2020) incorporates NNs into reinforcement learning failures in aircraft equipment (Celikmih, Inan, & Uguz,
frameworks to aid decision-making in complex 2020). The authors who developed this system employed the
environments, improving the efficacy of maintenance ReliefF feature selection method and a modified K-means
schedules and resource allocation while minimizing algorithm to identify the most compelling features and
downtime and associated maintenance costs. eliminate noisy or inconsistent data. Through this meticulous
process of converting raw sensor data into a refined format
ML models have transformed CBM from a reactive to a
suitable for ML analysis, the authors ensured their subsequent
predictive and even prescriptive practice. The ability to
models were fed high-quality information tailored for precise
accurately forecast potential failures and prescribe
predictions. Once the data had been adequately prepared,
maintenance actions can significantly reduce unplanned
they evaluated the model using the multi-layer perceptron
downtimes and maintenance costs, improving equipment
(MLP), SVR, and LR ML algorithms, achieving good results
availability and longevity. The selection of an appropriate
that demonstrated the effectiveness of an ML-based system
model depends on the nature of the data, the specific
in predicting equipment failure.
maintenance task, and the desired outcome. In addition,
deploying these models requires a careful balance between
5.2. Detection
predictive performance and computational efficiency,
especially when real-time analysis is required. The following The ability to quickly and accurately detect potential issues
section focuses on real-world examples where ML before they worsen and lead to failures is a pivotal component
algorithms were implemented into CBM systems. of the CBM strategy. The detection process is critical to CBM
systems as it lays the foundation for all maintenance actions
5. REAL-WORLD APPLICATIONS that come after it, dictating the efficiency and effectiveness
of the overall maintenance process. Incorporating ML
Effective deployment of CBM strategies in industry relies on
algorithms into a CBM system can lead to easier
systems capable of considering several canonical tasks in the
identification of subtle anomalies or trends that may not be
maintenance lifecycle of a machine. These tasks include the
apparent when using traditional methods.
detection of anomalies or potential failures, the diagnosis of
faults to identify their nature and cause, the classification of An example of this is the use of ML to detect concept drifts
different types of faults to streamline maintenance processes, in continuous data streams, specifically in the context of
the analysis of root causes to prevent future occurrences, and predictive maintenance (Zenisek, Holzinger, & Affenzeller,
the optimization of maintenance schedules to ensure 2019). The authors of this study aimed to proactively identify
maximum operational efficiency with minimal downtime. wear and tear in industrial machinery to prevent breakdowns.
ML techniques, with their ability to analyze and interpret vast They evaluated the LR, RF, and symbolic regression (SR)
amounts of data, have significantly contributed to the ability algorithms. They found that the models built with the RF
of CBM-based systems to address these canonical regressor performed better during the training phase. Still, the
maintenance tasks. This section dives into real-world SR models outperformed the other two during the testing
applications of ML in CBM, demonstrating how ML phase, indicating an increased capacity to generalize to
techniques can be used for each of these tasks within a CBM unknown data. Despite these differences, each of these three
framework. models was deemed very accurate for the predictive
maintenance tasks.
5.1. Data Preparation and Feature Handling
Another study proposed a data-driven predictive maintenance
The data preparation and feature selection stages of system for manufacturing production lines (Ayvaz & Alpay,
incorporating ML into CBM are critical steps that 2021). The data for this system was generated from IoT
significantly impact the effectiveness of predictive models. sensors in real-time, and ML algorithms were applied to this
The raw input data is transformed into a usable format in the data to detect potential failures before they occurred.
data preparation stage. In the feature selection stage, the most Following an evaluation of the efficacy of different ML
relevant data attributes are identified. These phases, models, it was determined that the RF and eXtreme gradient
involving data cleaning, normalization, and strategic feature boosting (XGBoost) algorithms outperformed all of the other
selection to capture underlying patterns indicative of models, and it was these two algorithms that were
equipment health, are essential in bridging the gap between implemented into the production line.
raw data and actionable insights, enabling the development
The ability of ML algorithms to automate event-oriented
of robust predictive maintenance strategies that are accurate
maintenance systems through unstructured, textual, and
and reliable.
unsupervised data has also been studied (Decker, Leite,
Minarini, Tisbeni, & Bonacorsi, 2022). The specific goal of

10
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

this study was to detect periods of anomalous activity based Another method of performing fault classification is through
on content and information extracted from log events. For an unsupervised clustering approach. In one study, a k-means
this purpose, the authors evaluated the one-class SVM, cluster-based fault identification model was constructed,
isolation forest (IF), and local outlier factor (LOF) ML which was made up of three components: a k-means cluster
algorithms, finding that IF provided the best fault detection analysis component, a fault mode – fault cluster centroid
accuracy. knowledge base component, and a fault identification
component (Wang et al., 2020b). It was found that the
5.3. Fault Diagnosis accuracy of this model when classifying surge, rubbing, and
misalignment faults for rotating machinery was 94%, 100%,
Simply detecting the presence of potential issues is not and 80%, respectively. Hierarchical clustering can break a
enough for an effective CBM system. Once potential
conventional classification problem into many sub-problems
problems are identified, the next step lies in diagnosing the
arranged in a hierarchy (Adams et al., 2019). One study found
various issues. Diagnosing faults requires a detailed analysis
that the proposed hierarchical classification method reduced
of potential fault types and their potential causes. This step is
resource consumption in such a system compared to a more
crucial for maintenance strategies like CBM that address the
traditional classification approach (Adams et al., 2019).
root of a problem instead of only its symptoms. Through the
use of ML, it becomes possible to not only identify but also Deep unsupervised methods such as RNNs have also been
understand the complex patterns indicating faulty conditions. employed in the fault classification. One study analyzed
The advanced analytical capabilities enabled by ML allow several different fault classification models: SVM, RF,
maintenance teams to make informed decisions, ensuring eXtreme gradient boosting (XGBoost), RNN, LSTM, and
timely and appropriate interventions. gated recurrent unit (GRU) (Huang, Chen, & Huang, 2019).
This study aimed to determine how to improve classification
An example of ML being used to enable fault diagnosis in a accuracy through dimensionality reduction best. Commonly
CBM system is the use of an ensemble ML technique based
used methods for dimensionality reduction, such as
on the RF, support vector machine (SVM), and MLP
autoencoders and variational autoencoders, did not
algorithms using LR as a metamodel to diagnose states of a
effectively improve classification accuracy and, in some
rotating machine to determine if the machine was operating cases, reduced it. However, when the variational autoencoder
normally, or whether it was experiencing faulty conditions was enhanced to be based on an RNN, the classification
(Jenab, Ward, Isaza, Ortega-Moody, & Anaya, 2024). The
accuracy of all models was significantly improved.
results of this study demonstrated the effectiveness of hybrid
approaches for determining specific maintenance needs
5.5. Root Cause Analysis
based on the machine's condition.
Once faults have been detected, diagnosed, and classified, the
5.4. Fault Classification next logical step in the CBM process is the performance of
root cause analysis. This is the process of identifying the
After abnormal conditions have been detected and diagnosed, underlying reasons for identified faults. Through this
the next phase in the CBM process lies in classifying faults.
analysis, maintenance teams move past the superficial
These classes are predefined based on the characteristics and
symptoms of a fault to address foundational issues that can
underlying causes of known faults. Accurately classifying
lead to compromised reliability and performance in
faults is crucial to streamlining maintenance procedures,
equipment. Accurate root cause analysis is essential for
facilitating targeted interventions, and enhancing decision- implementing measures to prevent the recurrence of issues.
making processes. The use of ML algorithms in CBM can ML algorithms can enhance root cause analysis through deep
enable the automated classification of different types of
pattern and correlation analysis in vast datasets.
faults, which allows a more sophisticated understanding of
equipment behavior and maintenance requirements. The RF and artificial neural network (ANN) models have
been used to conduct root cause analysis on a compressor
Since fault classification is a classification problem, using (Steurtewagen & Van den Poel, 2019). The RF model was
supervised ML algorithms is a natural choice. For example, used to classify compressor behavior into regular versus
the DT classifier has been used to classify pump failures in
erratic operation on sensor data. In contrast, the ANN model
the oil and gas industry (Aliyu, Mokhtar, & Hussin, 2022).
was used to predict whether the compressor was operating
This model attempted to classify operational condition data
within specifications. Root cause analysis ranked key
points into three classes: regular, broken, and recovering. The
variables contributing to compressor failures based on their
authors of this study found that their model achieved 91.94% Gini importance using the RF model. In this study, the
accuracy in the testing phase and 74.4% in the testing phase. authors demonstrate that insights from ML models, when
Another study utilized a CNN to improve the classification
combined with expert knowledge, can hypothesize the root
accuracy within an innovative plant maintenance system
causes of high vibrations in a compressor and suggest specific
through blob detection processing (Shin, Jo, Cha, & Lee,
maintenance actions to address these issues.
2020).

11
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

5.6. Optimization of Maintenance Schedules challenges is data compatibility. It's crucial to ensure that the
data from sensors and monitoring devices is in a format that
Once all prior tasks have been completed, the final step of
ML models can process. This often requires data
ML-based CBM centers on transforming synthesized insights
preprocessing (Masmoudi, Jaoua, Jaoua, & Yacout, 2021),
from ML analyses into actionable, strategic maintenance
conversion into compatible formats (Zhang, He, Yan, Jiang,
plans. The optimization of maintenance schedules is not
& Zhu, 2022), and addressing any data quality issues
merely about timing; it is about precision – ensuring
(Timocin, 2020).
maintenance activities are conducted at the optimal time to
prevent future failures, enhance equipment longevity, and
maintain operational efficiency. With the solid predictive
capabilities of ML algorithms, maintenance teams can
forecast equipment health deterioration and schedule
interventions proactively. By aligning maintenance activities
with the actual condition of equipment, organizations can
significantly reduce downtime, cut costs, and elevate the
reliability of operations.
Advanced statistical and ML methods for multi-step
multivariate time series forecasting in predictive maintenance
have been evaluated (Tessoni & Amoretti, 2022). In terms of
statistical methods, the authors considered vector
autoregression (VAR), vector moving average (VMA), and
vector autoregression moving average (VARMA), a novel
approach coined as Theta. In terms of ML models, they
evaluated variants of RNNs, namely the Elman RNN
(ERNN), LSTM, and gated recurrent unit (GRU). The Figure 8. Challenges and limitations of ML-based CBM
authors evaluated these models on eight different predictive systems
maintenance datasets, with ERNN outperforming advanced In terms of the types of data quality issues that may be
statistical methods in two of the datasets and LSTM and GRU encountered in ML-based CBM systems, maintenance data,
outperforming the statistical techniques for two. The success particularly historical records, often face problems like
of these RNN variants in outperforming statistical methods in incompleteness, noise, or sensor errors, leading to data gaps
multiple datasets highlights the potential of such methods to and variability that can impede the training and accuracy of
aid in scheduling maintenance activities. ML models (Nunes, Santos, & Rocha, 2023). Additionally,
CBM datasets typically exhibit class imbalance, where
A two-stage dynamic scheduling framework for aircraft fleet instances of failure are much fewer than normal operation
maintenance under a CBM strategy has been proposed cases, potentially leading to biased models that favor the
(Tseremoglou & Santos, 2024). In the first stage, the authors majority class (Sridhar & Sanagavarapu, 2021). Moving past
address uncertainty in predicting component health by these potential issues with data quality and availability,
planning the optimal maintenance policy based on the belief another aspect of ML-based CBM systems that can prove
state-space of component health, formulated as a partially challenging for organizations seeking to implement such
observable Markov decision process (POMDP) solved using systems is the need for real-time processing. CBM systems
the partially observable Monte Carlo planning (POMCP) frequently require immediate data analysis to promptly detect
algorithm. The second stage integrates this maintenance and predict equipment issues, which can be particularly
policy with the scheduling of preventive and corrective tasks strenuous in environments with limited resources. This
using a deep Q-network (DQN) model that continuously necessity calls for ML models that process data with minimal
adjusts the maintenance schedule based on new task latency, enabling rapid maintenance decisions and alert
information and resource availability constraints. Testing on generation (Tran, Doan, Vu, & Vu, 2023).
a case study from a large airline showed the model could
schedule 96.4% of monitored components on time while Alert generation is one of the primary features of an ML-
achieving a 46.2% maintenance cost reduction compared to a powered CBM system (Sinha, Pandaw, & Das, 2023). ML
corrective maintenance approach. models in these systems can generate alerts and notifications
upon detecting any anomalies or predicting potential failures.
6. CHALLENGES AND LIMITATIONS OF ML INTEGRATION Such alerts are crucial as they facilitate timely interventions,
IN CBM SYSTEMS
preventing equipment failures before they occur.
Additionally, these systems can enhance maintenance
This section delves into the challenges and considerations efficiency through condition-based scheduling (Tseremoglou
that arise when attempting to seamlessly incorporate ML into & Santos, 2024) and adaptive maintenance. By analyzing the
CBM systems, as depicted in Figure 8. One of the primary current state of the equipment, these systems can optimize

12
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

maintenance schedules and dynamically adjust based on the predictive accuracy (Maschler, Vietz, Jazdi, & Weyrich, 2-
actual health and needs of the equipment, ensuring that 2020). Moreover, implementing CBM can be resource-
maintenance is conducted only when necessary. This intensive, posing a challenge for smaller organizations or
approach minimizes equipment downtime and increases industries with budget constraints. The lack of skilled
longevity while reducing maintenance costs and increasing personnel with expertise in the domain and ML technology
maintenance effectiveness. exacerbates the situation. Complying with industry-specific
regulations and standards presents further challenges
To support these ML models, the system's infrastructure must
(Ramuhalli, Huning, Guler Yigitoglu, & Saxena, 2023), as
be robust and scalable, capable of handling the computational
does ensuring security and privacy in the face of potential
demands of sophisticated ML algorithms. This requirement
cyberattacks. Additionally, adopting commercial CBM
extends beyond hardware components, encompassing
solutions may lead to vendor lock-in, creating dependency on
software platforms, to ensure efficient operation under
specific vendors for software, hardware, or data services.
various conditions. The data storage and management system
may be the most critical component within this Interoperability also poses a challenge for integrating ML
hardware/software infrastructure. For any CBM system, the into CBM. Before 2018, there were no coherent standards to
ability to store and manage large volumes of sensor data is promote the intra- and inter-enterprise interoperability
paramount. This capability is often accomplished through required for modern CBM systems (Kaur, Selway,
database systems that are capable of handling time-series Grossmann, Stumptner, & Johnston, 2018). In two papers
data. released in 2018, Karamjit Kaur and colleagues from the
University of South Australia, MIMOSA, and the PdMA
Since ML models, especially in sectors with large and
Corporation outlined the Open Industrial Interoperability
complex data sources, demand significant computational
Ecosystem (OIIE) architecture.
resources, the infrastructure may need to include cloud
computing solutions or potent on-premises hardware to meet The OIIE effort aims to promote open standards and
these demands. Additionally, data visualization and reporting protocols that will improve industrial system interoperability.
tools are indispensable in CBM systems. These tools provide Its main goal is to provide a unified, integrated environment
clear insights into the health and maintenance performance of where various industrial systems, such as enterprise resource
equipment and empower maintenance personnel with the planning systems, manufacturing machinery, and supply
necessary data interpretation capabilities to make informed chain management tools, may effectively communicate and
and effective decisions. work together. Promoting open architecture for greater
system integration flexibility, data sharing capabilities for
One of the major hurdles with implementing ML models is
vital functions like CBM and real-time monitoring, and
the interpretability of the models themselves, especially in
standardization initiatives to guarantee system compatibility
complex DL networks, which often operate as black-box
are all essential to the OIIE's purpose. To further the growth
models, making it challenging to derive meaningful insights
of the ecosystem, the architecture also strongly emphasizes
from their predictions (Hussain, 2019). One method of
cooperative efforts between technology suppliers, business
resolving this issue is designing user-friendly interfaces for
leaders, and end users. In line with the objectives of Industry
maintenance personnel. These human-machine interfaces
4.0, the OIIE aims to improve decision-making in industrial
should enable easy interaction with ML models and facilitate
operations, promote innovation, and increase efficiency
the interpretation of results. These interfaces are critical to the
through various initiatives, such as automation, data sharing,
successful integration of ML into CBM systems (Quispe G.,
and the use of IoT devices in industrial settings.
Rajabiyazdi, & Jamieson, 2020).
The OIIE is a promising and effective series of guidelines for
Integrating ML models into CBM systems involves advanced
encouraging interoperability between industrial systems to
software development and deployment strategies. A crucial
facilitate the incorporation of CBM processes into existing
step in this process is model deployment, where the ML
systems. However, there still remains a gap in the literature
models must be integrated compatibly with the existing CBM
for standards that center on interoperability between CBM
software architecture. This integration could encompass
systems and ML algorithms. This gap deserves further
various techniques like containerization, utilizing
exploration.
microservices, or direct integration with existing software
modules.
7. FUTURE TRENDS AND DIRECTIONS
Ethical considerations are increasingly important, especially
The landscape of ML-based CBM is continuously evolving,
concerning potential biases in data and models that might
driven by technological advancements in ML and related
lead to unfair maintenance decisions. Addressing and
fields, some of which are shown in Figure 9. A significant
mitigating these biases is crucial yet challenging (Bacelar,
development for the future of ML-based CBM is the rise of
2021). The dynamic nature of equipment conditions requires
XAI, which aims to make complex ML models more
continuous model monitoring and updating to maintain
transparent and understandable, thus increasing the

13
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

trustworthiness of CBM systems (Krishnamurthy, Nezafati, Li, 2021). Additionally, the field of meta-learning, which
Stayton, & Singh, 2020). The integration of edge computing focuses on developing algorithms that can quickly adapt to
is set to play a pivotal role in facilitating real-time data new tasks with minimal retraining, could facilitate the
processing closer to the data sources, thereby enhancing the development of CBM models that can rapidly generalize to
efficiency and scalability of CBM systems (Liu, Hu, Jia, & new asset types of operating conditions, reducing the need for
Tao, 2021). Furthermore, developing and using novel extensive data collection and retraining for each new
ensemble techniques is expected to continue, leveraging the application (Yang, Wang, & Luo, 2024).
strengths of different algorithms for improved predictive
Self-supervised learning techniques, which enable models to
accuracy and robustness in CBM (Jenab et al., 2024).
learn rich representations from unlabeled data, could
potentially alleviate the data labeling bottleneck often
encountered in CBM applications. Using the vast amounts of
unlabeled sensor data available, self-supervised models could
learn meaningful feature representations that can be
transferred to downstream CBM tasks, reducing the reliance
on extensive labeled datasets (Chen, Ma, Xu, Jin, & Zhou,
2024). The process of domain adaption, which aims to
transfer knowledge from one domain to another, could be
instrumental in developing CBM models that can seamlessly
adapt to new asset types or operating environments,
leveraging knowledge gained from related domains (Nejjar,
Figure 8. Future trends and directions for ML-based CBM Geissmann, Zhao, Taal, & Fink, 2024).
systems
New sensor technologies are expected to generate high- 8. CONCLUSION
fidelity data (Levinski et al., 2023) and enable more precise
monitoring through multi-sensor fusion (van Staden & Boute, As time has passed, knowledge has increased, and technology
2021). Another exciting development is the shift towards AI- has evolved and improved, the use of ML has become pivotal
driven prognostics (Zschech, Heinrich, Bink, & Neufeld, to modern CBM systems. The application of ML technology
2021) and prescriptive maintenance (Ansari, Glawar, & has led to more effective decision-making processes, which
Nemeth, 2019), where AI algorithms will not only predict has had a significant positive effect on operational reliability
failures but also suggest optimal maintenance actions. This and cost-effectiveness. Despite these advancements, properly
progression is complemented by the trend towards integrating ML techniques into existing CBM systems is
continuous learning (Maschler et al., 2020) and adaptive challenging. Issues including poor data quality,
systems (Xiong, Zhou, Ma, Zhang, & Lin, 2023), where incompatibility between systems, and the intensive demands
models dynamically update themselves in response to real- of real-time data processing present significant hurdles to
time data, maintaining effectiveness despite changing overcome.
equipment conditions. Furthermore, ML models' complexity and black-box nature
Blockchain technology is anticipated to play a crucial role in can sometimes act as a barrier to integration, specifically in
enhancing data security and traceability in CBM, ensuring the terms of interpretability and usability. Adequately addressing
integrity and transparency of maintenance records (Tran et these challenges is essential for the continued efficacy of ML-
al., 2022). Additionally, the increasing collaboration between based approaches to CBM. These challenges aside, the future
humans and AI models, mainly through user-friendly potential of ML in CBM is excellent. The rise of XAI is
interfaces in CBM systems, is expected to foster a synergistic making ML models more transparent and understandable,
approach to maintenance decision-making. Moreover, with edge computing is enhancing the processing capabilities at
the expanding role of AI in CBM, ethical considerations will the data collection site, and the development of hybrid ML
gain prominence, emphasizing responsible AI practices to models promises to deliver more robust and efficient
ensure fairness and accountability. maintenance solutions.
Several cutting-edge ideas and techniques from the broader As industries continue to evolve, it is becoming increasingly
ML community hold significant promise for enhancing CBM important for organizations to embrace technological
systems in the future. One such concept is federated learning, innovation balanced and sustainably. This includes being
which enables collaborative training of ML models across mindful of the ethical implications and challenges
multiple decentralized data sources without sharing raw data. accompanying such advancements. ML has significant
This approach could be precious in CBM scenarios where potential to bring about more reliable, efficient, and
data privacy and security are critical concerns, allowing optimized maintenance processes in CBM systems. As such,
models to learn from diverse data sources while preserving organizations and industries must continue to focus on
data ownership and confidentiality (Zhang, Li, Ma, Luo, &

14
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

innovation, collaboration, and continuous improvement to Access, 9, 17365-17381.

utilize these benefits fully. https://doi.org/10.1109/ACCESS.2021.3051583
Ayvaz, S., & Alpay, K. (2021). Predictive maintenance
REFERENCES system for production lines in manufacturing: A machine
learning approach using IOT data in real-time. Expert
Adams, S., Meekins, R., Beling, P. A., Farinholt, K., Brown,
Systems with Applications, 173, 114598.
N., Polter, S., & Dong, Q. (2019) Hierarchical fault
https://doi.org/10.1016/j.eswa.2021.114598
classification for resource constrained systems.
Bacelar, M. (2021). Monitoring bias and fairness in machine
Mechanical Systems and Signal Processing, 134, learning models: A review. ScienceOpen Preprints.
1062666. https://doi.org/10.1016/j.ymssp.2019.106266 Bae, S. J., Mun, B. M., Chang, W., & Vidakovic, B. (2019).
Adryan, F. A. & Sastra, K. W. (2021). Predictive
Condition monitoring of a steam turbine generator using
maintenance for aircraft engine using machine learning:
wavelet spectrum based control chart. Reliability
Trends and challenges. International Journal of Aviation
Engineering & System Safety, 184, 13-20.
Science and Engineering, 3(1), pp. 37-44.
https://doi.org/10.1016/j.ress.2017.09.025
https://doi.org/10.47355/AVIA.V3I1.45 BahooToroody, A., De Carlo, F., Paltrinieri, N., Tucci, M., &
Afridi, Y. S., Ahmad, K., & Hassan, L. (2021). Artificial Van Gelder, P. H. A. J. M. (2020) Bayesian regression
intelligence based prognostic maintenance of renewable
based condition monitoring approach for effective
energy systems: A review of techniques, challenges, and
reliability prediction of random processes in autonomous
future research directions. International Journal of
energy supply operation. Reliability Engineering &
Energy Research, 46(15), pp. 21619-21642.
System Safety, 201, 106966.
https://doi.org/10.1002/er.7100 https://doi.org/10.1016/j.ress.2020.106966
Ahmad, S., Styp-Rekowski, K., Nedelkoski, S., & Kao, O. Berghout, T., Benbouzid, M., Muyeen, S. M., Bentrcia, T., &
(2020). Autoencoder-based condition monitoring and
Mouss, L.-H. (2021). Auto-nahl: A neural network
anomaly detection method for rotating machines. 2020
approach for condition-based maintenance of complex
IEEE International Conference on Big Data (Big Data).
industrial systems. IEEE Access, 9, 152829–152840.
https://doi.org/10.1109/bigdata50022.2020.9378015 https://doi.org/10.1109/access.2021.3127084
Ahmed, M. S., Al Bloushi, M. A., & Ali, A. (2022). Case Campos Olivares, D., Carrasco Muñoz, A., Mazzoleni, M.,
study: Application of wireless condition based
Ferramosca, A., & Luque Sendra, Amalia. (2023).
monitoring by applying machine learning models. In
Screening of machine learning techniques on predictive
ADIPEC. OnePetro.
maintenance: A scoping review. Dyna, 99, pp. 159-165.
Aliyu, R., Mokhtar, A. A., & Hussin, H. (2022) Classification
https://doi.org/10.6036/10950
of pump failure using a decision tree technique. Carvalho, T. P., Soares, F. A., Vita, R., Francisco, R. da,
Proceedings of the International Conference on Basto, J. P., & Alcalá, S. G. (2019). A systematic
Renewable Energy and E-mobility (ICREEM 2022),
literature review of machine learning methods applied to
319-336. https://doi.org/10.1007/978-981-99-5946-
predictive maintenance. Computers & Industrial
4_26
Engineering, 137, 106024.
Allah Bukhsh, Z., Saeed, A., Stipanovic, I., & Doree, A. G.
https://doi.org/10.1016/j.cie.2019.106024
(2019). Predictive maintenance using tree-based Celikmih, K., Inan, O., & Uguz, H. (2020). Failure prediction
classification techniques: A case of railway switches. of aircraft equipment using machine learning with a
Transportation Research Part C: Emerging
hybrid data preparation method. Scientific
Technologies, 101, 35–54.
Programming, 2020, 1–10.
https://doi.org/10.1016/j.trc.2019.02.001
https://doi.org/10.1155/2020/8616039
Alsumaidaee, Y. A. M. et al. (2022). Review of medium- Chatterjee, J. & Dethlefs, N. (2021). Scientometric review of
voltage switchgear fault detection in a condition-based artificial intelligence for operations & maintenance of
monitoring system by using deep learning. Energies,
wind turbines: The past, present and future. Renewable
15(18), 6762. https://doi.org/10.3390/en15186762
and Sustainable Energy Reviews, 144, 111051.
Ansari, F., Glawar, R., & Nemeth, T. (2019). Prima: A
https://doi.org/10.1016/j.rser.2021.111051
prescriptive maintenance model for cyber-physical
Chen, C., Fu, H., Zheng, Y., Tao, F., & Liu, Ying. (2023).
production systems. International Journal of Computer The advance of digital twin for predictive maintenance:
Integrated Manufacturing, 32(4–5), 482–503. The role and function of machine learning. Journal of
https://doi.org/10.1080/0951192x.2019.1571236
Manufacturing Systems, 71, pp. 581-594.
Aqueveque, P., Radrigan, L., Pastene, F., Morales, A. S., &
https://doi.org/10.1016/j.jmsy.2023.10.010
Guerra, E. (2021). Data-driven condition monitoring of
Chen, P., Ma, Z., Xu, C., Jin, Y., & Zhou, C. (2024). Self-
mining mobile machinery in non-stationary operations
supervised transfer learning for remote wear evaluation
using wireless accelerometer sensor modules. IEEE in machine tool elements with imaging transmission

15
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

attenuation. IEEE Internet of Things Journal. Fernandes, J., Reis, J., Melão, N., Teixeira, L., & Amorim,
https://doi.org/10.1109/JIOT.2024.3382878 M. (2021). The role of Industry 4.0 and BPMN in the
Cheng, J., Liu, Y., Li, W., & Li, T. (2023). Deep arise of condition-based and predictive maintenance: A
reinforcement learning for cost-optimal condition-based case study in the automotive industry. Applied Sciences,
maintenance policy of offshore wind turbine 11(8), 3438. https://doi.org/10.3390/app11083438
components. Ocean Engineering, 283, 115062. Ferreira, C. & Gonçalves, G. (2022). Remaining useful life
https://doi.org/10.1016/j.oceaneng.2023.115062 prediction and challenges: A literature review on the use
Cheng, Y., Wang, C., Wu, J., Zhu, H., & Lee, C. K. M. of machine learning methods. Journal of Manufacturing
(2022). Multi-dimensional recurrent neural network for Systems, 63, pp. 550-562.
remaining useful life prediction under variable operating https://doi.org/10.1016/j.jmsy.2022.05.010
conditions and multiple fault modes. Applied Soft Fong, S. (2022). Unsupervised methods for condition-based
Computing, 118, 108507. maintenance in non-stationary operating conditions.
https://doi.org/10.1016/j.asoc.2022.108507 Doctoral dissertation. University of Waterloo, Waterloo,
Ciaburro, G. (2022). Machine fault detection methods based Ontario, Canada.
on machine learning algorithms: A review. https://uwspace.uwaterloo.ca/handle/10012/18191
Mathematical Biosciences and Engineering, 19(11), pp. Fredriksson, G. (2022). Prognostics for condition-based
11453-11490. https://doi.org/10.3934/mbe.2022534 maintenance of electrical control units using on-board
Çınar, Z. M. et al. (2020). Machine learning in predictive sensors and machine learning. Master’s thesis.
maintenance towards sustainable smart manufacturing in Linköping University, Linköping, Sweden.
Industry 4.0. Sustainability, 12(19), 8211. https://www.diva-
https://doi.org/10.3390/su12198211 portal.org/smash/record.jsf?pid=diva2%3A1650950&d
Decker, L., Leite, D., Minarini, F., Tisbeni, S. R., & swid=7028
Bonacorsi, D. (2022). Unsupervised learning and online Gupta, P. et al. (2024). A review on machine learning
anomaly detection. International Journal of Embedded enhanced predictive maintenance for electric vehicle
and Real-Time Communication Systems, 13(1), 1–16. power electronics: A pathway to improved reliability and
https://doi.org/10.4018/ijertcs.302112 longevity. 3rd International Conference on Applied
Doğru, A., Bouarfa, S., Arizar, R., & Aydoğan, R. (2020). Research and Engineering (ICARAE2023).
Using convolutional neural networks to automate https://doi.org/10.1051/e3sconf/202450503017
aircraft maintenance visual inspection. Aerospace, Hamaide, V. & Glineur, F. (2021). Unsupervised minimum
7(12), 171. https://doi.org/10.3390/aerospace7120171 redundancy maximum relevance feature selection for
Drakaki, M., Karnavas, Y. L., Tziafettas, I. A., Linardos, V., predictive maintenance: Application to a rotating
& Tzionas, P. (2022). Machine learning and deep machine. International Journal of Prognostics and
learning based methods towards Industry 4.0 predictive Health Management, 12(2).
maintenance in induction motors: State of the art survey. https://doi.org/10.36001/ijphm.2021.v12i2.2955
Journal of Industrial Engineering & Management, Hong, X., Xu, Z., & Zhang, Z. (2019). Abnormal condition
15(1), pp. 31-57. https://doi.org/10.3926/jiem.3597 monitoring and diagnosis for coal mills based on support
Elbouchikhi, E., Zia, M. F., Benbouzid, M., & El Hani, S. vector regression. IEEE Access, 7, 170488–170499.
(2021). Overview of signal processing and machine https://doi.org/10.1109/access.2019.2955249
learning for smart grid condition monitoring. Huang, Y., Chen, C. -H., & Huang, C. -J. (2019). Motor fault
Electronics, 10(21), 2725. detection and feature extraction using RNN-based
https://doi.org/10.3390/electronics10212725 variational autoencoder. IEEE Access, 7, 139086-
Ellefsen, A. L., Æsøy, V., Ushakov, S., & Zhang, H. (2019). 139096.
A comprehensive survey of prognostics and health https://doi.org/10.1109/ACCESS.2019.2940769
management based on deep learning for autonomous Huerta Herraiz, Á., Pliego Marugán, A., & García Márquez,
ships. IEEE Transactions on Reliability, 68(2), pp. 720- F. P. (2020). Photovoltaic plant condition monitoring
740. https://doi.org/10.1109/TR.2019.2907402 using thermal images analysis by convolutional neural
Färber, M., & Tampakis, L. (2023). Analyzing the impact of network-based structure. Renewable Energy, 153, 334–
companies on AI research based on publications. 348. https://doi.org/10.1016/j.renene.2020.01.148
Scientometrics, 129(1), 31–63. Hussain, J. (2019). Deep learning black box problem
https://doi.org/10.1007/s11192-023-04867-3 (Dissertation). Retrieved from
Fernandes, M., Corchado, J. M., & Marreiros, G. (2022). https://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-
Machine learning techniques applied to mechanical fault 393479
diagnosis and fault prognosis in the context of real Jan, Z., Ahamed, F., Mayer, W., Patel, N., Grossmann, G.,
industrial manufacturing use-cases: A systematic Stumptner, M., & Kuusk, A. (2023). Artificial
literature review. Applied Intelligence, 52, pp. 14246- intelligence for industry 4.0: Systematic review of
14280. https://doi.org/10.1007/s10489-022-03344-3 applications, challenges, and opportunities. Expert

16
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

Systems with Applications, 216, 119456. Control Conference (IAEAC).

https://doi.org/10.1016/j.eswa.2022.119456 https://doi.org/10.1109/iaeac50856.2021.9390623
Jauro, F., Chiroma, H., Gital, A. Y., Almutairi, M., Lourenço, A., Ferraz, C., Meira, J., Marreiros, G., Bolón-
Abdulhamid, S. M., & Abawajy, J. H. (2020). Deep Canedo, V., & Alonso-Betanzos, A. (2023). Automated
Learning Architectures in Emerging Cloud Computing green machine learning for condition-based
Architectures: Recent Development, challenges and next maintenance. Proceedings of the European Symposium
research trend. Applied Soft Computing, 96, 106582. on Artificial Neural Networks, Computational
https://doi.org/10.1016/j.asoc.2020.106582 Intelligence and Machine Learning, October 4-6,
Jenab, K., Ward, T., Isaza, C., Ortega-Moody, J., & Anaya, Bruges, Belgium.
K. (2024). Intelligence based condition monitoring https://doi.org/10.14428/esann/2023.ES2023-85
model. In International Congress and Workshop on Lu, H., Du, M., Qian, K., He, X., & Wang, K. (2022). GAN-
Industrial AI (pp. 639-650). Cham: Springer Nature based data augmentation strategy for sensor anomaly
Switzerland. https://doi.org/10.1007/978-3-031-39619- detection in industrial robots. IEEE Sensors Journal,
9_47 22(18), 17464-17474.
Jourdan, N., Longard, L., Biegel, T., & Metternich, J. (2021). https://doi.org/10.1109/JSEN.2021.3069452
Machine learning for intelligent maintenance and quality Maheswari, R. U., & Umamaheswari, R. (2020). Wind
control: A review of existing datasets and corresponding turbine drivetrain expert fault detection system:
use cases. Proceedings of the Conference on Production Multivariate empirical mode decomposition based multi-
Systems and Logistics : CPSL 2021. sensor fusion with Bayesian learning classification.
https://doi.org/10.15488/11280 Intelligent Automation And Soft Computing, 26(3), 479-
Kaur, K., Selway, M., Grossmann, G., Stumptner, M., & 488. https://doi.org/10.32604/iasc.2020.013924
Johnston, A. (2018). Towards an open-standards based Martins, A. B., Fonesca, I., Farinha, J., Reis, J., & Cardoso,
framework for achieving condition-based predictive A. J. M. (2022). Prediction maintenance based of
maintenance. Proceedings of the 8th International vibration analysis and deep learning – A case study of a
Conference on the Internet of Things. drying press supported on a hidden Markov model.
https://doi.org/10.1145/3277593.3277608 SSRN. http://doi.org/10.2139/ssrn.4194601
Kaur, K., Selway, M., Grossmann, G., Stumptner, M., Maschler, B., Vietz, H., Jazdi, N., & Weyrich, M. (2020).
Johnston, A., & Wong, P. (2018). Towards an Open Continual learning of fault prediction for turbofan
Standards-based Architecture for Condition-based engines using deep learning with elastic weight
Predictive Maintenance and IIoT. Semantic Web consolidation. 2020 25th IEEE International Conference
Journal, 46, 704-719. on Emerging Technologies and Factory Automation
Kelleher, J. D. (2019). Deep learning. MIT press. (ETFA).
Krishnamurthy, V., Nezafati, K., Stayton, E., & Singh, V. https://doi.org/10.1109/ETFA46521.2020.9211903
(2020). Explainable AI framework for imaging-based Masmoudi, O., Jaoua, M., Jaoua, A., & Yacout, S. (2021).
predictive maintenance for automotive applications and Data preparation in machine learning for condition-
beyond. Data-Enabled Discovery and Applications, based maintenance. Journal of Computer Science, 17(6),
4(1). https://doi.org/10.1007/s41688-020-00042-2 525–538. https://doi.org/10.3844/jcssp.2021.525.538
Kumar, P., Khalid, S., & Kim, H. S. (2023). Prognostics and Nacchia, M., Fruggiero, F., Lambiase, A., & Bruton, K.
health management of rotating machinery of industrial (2021). A systematic mapping of the advancing use of
robot with deep learning applications – A review. machine learning techniques for predictive maintenance
Mathematics, 11(13), 3008. in the manufacturing sector. Applied Sciences, 11(6),
https://doi.org/10.3390/math11133008 2546. https://doi.org/10.3390/app11062546
Leukel, J., González, J., & Riekert, M. (2021). Adoption of Namuduri, S., Narayanan, B. N., Davuluru V. S. P., Burton,
machine learning technology for failure prediction in L., & Bhansali, S. (2020). Deep learning methods for
industrial maintenance: A systematic review. Journal of sensor based predictive maintenance and future
Manufacturing Systems, 61, pp. 87-96. perspectives for electrochemical sensors. Journal of The
https://doi.org/10.1016/j.jmsy.2021.08.012 Electrochemical Society, 167, 037552.
Levinski, O., Verhagen, W. J., Muscarello, V., Scott, M. J., https://doi.org/10.1149/1945-7111/ab67a8
Fayek, H. M., & Marzocca, P. (2023). An innovative Nejjar, I., Geissmann, F., Zhao, M., Taal, C., & Fink, O.
high-fidelity approach to structural health monitoring. In (2024). Domain adaptation via alignment of operation
AIAIC 2023: 20th Australian International Aerospace profile for remaining useful lifetime prediction.
Congress (pp. 247-252). Engineers Australia. Reliability Engineering & System Safety, 242, 109718.
Liu, Z., Hu, C., Jia, J., & Tao, F. (2021). Design of equipment https://doi.org/10.1016/j.ress.2023.109718
condition maintenance knowledge base in power IOT Nemat Saberi, A., Belahcen, A., Sobra, J., & Vaimann, T.
based on edge computing. 2021 IEEE 5th Advanced (2022). Lightgbm-based fault diagnosis of rotating
Information Technology, Electronic and Automation machinery under changing working conditions using

17
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

modified recursive feature elimination. IEEE Access, 10, POMDP model using interpolated value function. IFAC-
81910–81925. PapersOnLine, 55(16), 128-135.
https://doi.org/10.1109/access.2022.3195939 https://doi.org/10.1016/j.ifacol.2022.09.012
Ngoma, W. J., Mativenga, P. T., & Pretorius, J. (2020) Sanzana, M. R., Maul, T., Wong, J. Y., Abdulrazic, M. O. M.,
Enabling condition based maintenance in a precious & Yip, C. -C. (2022). Application of deep learning in
metal processing plant. Procedia CIRP. 91, 893-898. facility management and maintenance for heating,
https://doi.org/10.1016/j.procir.2020.04.138 ventilation, and air conditioning. Automation in
Nor, A. A. M., Kassim, M., Minhat, M. S., & Ya’acob, N. Construction, 141, 104445.
(2022). A review on predictive maintenance technique https://doi.org/10.1016/j.autcon.2022.104445
for nuclear reactor cooling system using machine Sarker, I. H. (2021). Machine learning: Algorithms, real-
learning and augmented reality. International Journal of world applications and Research Directions. SN
Electrical and Computer Engineering (IJECE), 12(6), Computer Science, 2(3). https://doi.org/10.1007/s42979-
pp. 6602-6613. https://10.11591/ijece.v12i6.pp6602- 021-00592-x
6613 Saurav, K., Avesh, M., Sharma, R. C., & Hossain, I. (2023).
Nunes, P., Santos, J., & Rocha, E. (2023). Challenges in Deploying machine learning algorithms for predictive
predictive maintenance – A review. CIRP Journal of maintenance of high-value assets of Indian railways.
Manufacturing Science and Technology, 40, 53-67. Transportation Energy and Dynamics, pp. 401-426.
https://doi.org/10.1016/j.cirpj.2022.11.004 https://doi.org/10.1007/978-981-99-2150-8_17
Ogunfowora, O. & Najjaran, H. (2023). Reinforcement and Schneider, J. & Vlachos, M. (2023). A survey of deep
deep reinforcement learning-based solutions for machine learning: From activations to transformers.
maintenance planning, scheduling policies, and arXiv:2302.00722.
optimization. Journal of Manufacturing Systems, 70, pp. https://doi.org/10.48550/arXiv.2302.00722
244-263. https://doi.org/10.1016/j.jmsy.2023.07.014 Serradilla, O., Zugasti, E., Rodriguez, J., & Zurutuza, U.
Payette, M. & Abdul-Nour, G. (2023) Machine learning (2022). Deep learning models for predictive
applications for reliability engineering: A review. maintenance: A survey, comparison, challenges and
Sustainability, 15(7), 6270. prospects. Applied Intelligence, 52, pp. 10934-10964.
https://doi.org/10.3390/su15076270 https://doi.org/10.1007/s10489-021-03004-y
Polverino, L. et al. (2023). Machine learning for prognostics Sharma, J., Mittal, M. L., & Soni, G. (2022). Condition-based
and health management of industrial mechanical systems maintenance using machine learning and role of
and equipment: A systematic literature review. interpretability: A review. International Journal of
International Journal of Engineering Business System Assurance Engineering and Management.
Management, 15. https://doi.org/10.1007/s13198-022-01843-7
https://doi.org/10.1177/18479790231186848 Shin, M. -K. Woo, J. J., Cha, H. M., & Lee, S. -H. (2020). A
Quatrini, E., Constantino, F., Di Gravio, G., & Patriarca, R. study on the condition based maintenance evaluation
(2020). Condition-based maintenance – An extensive system of smart plant device using convolutional neural
literature review. Machines, 8(2), 31. network. Journal of Mechanical Science and
https://doi.org/10.3390/machines8020031 Technology, 34, 2507-2514.
Quispe G, D. A., Rajabiyazdi, F., & Jamieson, G. A. (2020). https://doi.org/10.1007/s12206-020-0526-4
A machine learning-based Micro-World platform for Siang, Y. Y., Ahamd, M. R., & Abidin, M. S. Z. (2021).
condition-based maintenance. 2020 IEEE International Anomaly detection based on tiny machine learning: A
Conference on Systems, Man, and Cybernetics (SMC). review. Open International Journal of Informatics, 9(2),
https://doi.org/10.1109/smc42975.2020.9283448 pp. 67-78. https://doi.org/10.11113/oiji2021.9nSpecial
Ramuhalli, P., Huning, A., Guler Yigitoglu, A., & Saxena, A. Issue 2.148
(2023). Status Report on Regulatory Criteria Applicable Singh, D. (2023). Internet of things. Factories of the Future:
to the Use of Artificial Intelligence (AI) and Machine Technological Advancements in the Manufacturing
Learning (ML) (No. ORNL/SPR-2023/3072). Oak Industry, 195–227.
Ridge National Laboratory (ORNL), Oak Ridge, TN https://doi.org/10.1002/9781119865216.ch9
(United States). Singh, J., Azamfar, M., Li, F., & Lee, J. (2020). A systematic
Rathore, M. S., & Harsha, S. P. (2022). Prognostic analysis review of machine learning algorithms for prognostics
of high-speed cylindrical roller bearing using Weibull and health management of rolling element bearings:
distribution and k-nearest neighbor. Journal of Fundamentals, concepts and applications. Measurement
Nondestructive Evaluation, Diagnostics and Prognostics Science and Technology, 32, 012001.
of Engineering Systems, 5(1), 011005. https://doi.org/10.1088/1361-6501/ab8df9
https://doi.org/10.1115/1.4051314 Sinha, A., Pandaw, A. S., & Das, D. (2023). An intelligent
Roux, M., Fang, Y. -P., & Barros, A. (2022). Maintenance fault detection framework for HVAC systems with alert
planning under imperfect monitoring: An efficient

18
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

generation. SN Computer Science, 4(5). Wang, Q., Bu, S., & He, Z. (2020). Achieving predictive and
https://doi.org/10.1007/s42979-023-02107-2 proactive maintenance for high-speed railway power
Sridhar, S., & Sanagavarapu, S. (2021). Handling data equipment with LSTM-RNN. IEEE Transactions on
imbalance in predictive maintenance for machines using Industrial Informatics, 16(10), 6509-6517.
SMOTE-based oversampling. 2021 13th International https://doi.org/10.1109/TII.2020.2966033
Conference on Computational Intelligence and Wang, Q., Liu, J., Wei, B., Chen, W., & Xu, S. (2020).
Communication Networks (CICN). Investigating the construction, training, and verification
https://doi.org/10.1109/cicn51697.2021.9574668 methods of K-means clustering fault recognition model
Surucu, O., Gadsden, S. A., & Yawney, J. (2023). Condition for rotating machinery. IEEE Access, 8, 196515–
monitoring using machine learning: A review of theory, 196528. https://doi.org/10.1109/access.2020.3028146
applications, and recent advances. Expert Systems with Ward, T., Jenab, K., & Ortega-Moody, J. (2024). Machine
Applications, 221, 119738. learning models for condition-based maintenance with
https://doi.org/10.1016/j.eswa.2023.119738 regular truncated signals. Decision Science Letters,
Tanimoto, A. (2021). Combinatorial Q-learning for 13(1), 197-210. https://doi.org/10.5267/j.dsl.2023.9.006
condition-based infrastructure maintenance. IEEE Xiong, J., Zhou, J., Ma, Y., Zhang, F., & Lin, C. (2023).
Access, 9, 46788–46799. Adaptive deep learning-based remaining useful life
https://doi.org/10.1109/access.2021.3059244 prediction framework for systems with multiple failure
Tama, B. A., Vania, M., Lee, S., & Lim, S. (2023). Recent patterns. Reliability Engineering & System Safety, 235,
advances in the application of deep learning for fault 109244. https://doi.org/10.1016/j.ress.2023.109244
diagnosis of rotating machinery using vibration signals. Yang, J., Wang, X., and Luo, Z. (2024) Few-shot remaining
Artificial Intelligence Review, 56, pp. 4667-4709. useful life prediction based on meta-learning with deep
https://doi.org/10.1007/s10462-022-10293-3 sparse kernel network. Information Sciences, 653,
Tessoni, V., & Amoretti, M. (2022). Advanced statistical and 119795. https://doi.org/10.1016/j.ins.2023.119795
machine learning methods for multi-step multivariate Yin, X., Liu, Q., Huang, X., & Pan, Y. (2022). Perception
time series forecasting in Predictive Maintenance. model of surrounding rock geological conditions based
Procedia Computer Science, 200, 748–757. on TBM operational big data and combined
https://doi.org/10.1016/j.procs.2022.01.273 unsupervised-supervised learning. Tunnelling and
Teixeira, H. N., Lopes, I., & Braga, A. C. (2020). Condition- Underground Space Technology, 120, 104285.
based maintenance implementation: A literature review. https://doi.org/10.1016/j.tust.2021.104285
Procedia Manufacturing, 51, 228-235. Zenisek, J., Holzinger, F., & Affenzeller, M. (2019). Machine
https://doi.org/10.1016/j.promfg.2020.10.033 learning based concept drift detection for predictive
Timocin, T. (2020). Data quality in the interface of industrial maintenance. Computers & Industrial Engineering, 137,
manufacturing and machine learning (Dissertation). 106031. https://doi.org/10.1016/j.cie.2019.106031
Retrieved from Zhang, H., He, X., Yan, W., Jiang, Z., & Zhu, S. (2022). A
https://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva- machine learning-based approach for product
419983 maintenance prediction with reliability information
Tran, M.-Q., Doan, H.-P., Vu, V. Q., & Vu, L. T. (2023). conversion. Autonomous Intelligent Systems, 2(1).
Machine learning and IOT-based approach for Tool https://doi.org/10.1007/s43684-022-00033-3
Condition Monitoring: A review and future prospects. Zhang, N., & Si, W. (2020). Deep reinforcement learning for
Measurement, 207, 112351. condition-based maintenance planning of multi-
https://doi.org/10.1016/j.measurement.2022.112351 component systems under dependent competing risks.
Tran, V. H., Lenssens, B., Kassab, A., Laks, A., Rivière, E., Reliability Engineering & System Safety, 203, 107094.
Rosinosky, G., & Sadre, R. (2022). Machine‐as‐a‐ https://doi.org/10.1016/j.ress.2020.107094
service: Blockchain ‐ based management and Zhang, W., Li, X., Ma, H., Luo, Z., & Li, X. (2020).
maintenance of industrial appliances. Engineering Federated learning for machinery fault diagnosis with
Reports, 5(7). https://doi.org/10.1002/eng2.12567 dynamic validation and self-supervision. Knowledge-
Tseremoglou, I. & Santos B.F. (2024). Condition-based Based Systems, 213, 106679,
maintenance scheduling of an aircraft fleet under partial https://doi.org/10.1016/j.knosys.2020.106679
observability: A deep reinforcement learning approach. Zhao, Z., Li, T., Wu, J., Sun, C., Wang, S., Yan, R., & Chen,
Reliability Engineering & System Safety, 241, 109582. X. (2020). Deep learning algorithms for rotating
https://doi.org/10.1016/j.ress.2023.109582 machinery intelligent diagnosis: An open source
van Staden, H. E., & Boute, R. N. (2021). The effect of multi- benchmark study. ISA Transactions, 107, 224–255.
sensor data on condition-based maintenance policies. https://doi.org/10.1016/j.isatra.2020.08.010
European Journal of Operational Research, 290(2), Zhu, M., & Zhou, X. (2023). Hierarchical-clustering-based
585–600. https://doi.org/10.1016/j.ejor.2020.08.035 joint optimization of spare part provision and
maintenance scheduling for serial-parallel multi-station

19
 INTERNATIONAL JOURNAL OF PROGNOSTICS AND HEALTH MANAGEMENT

manufacturing systems. International Journal of Jorge Ortega-Moody received his

Production Economics, 264, 108971. Ph.D. degree in Mechatronics, National
https://doi.org/10.1016/j.ijpe.2023.108971 Polytechnic Institute, Campus
Zschech, P., Heinrich, K., Bink, R., & Neufeld, J. S. (2019). Querétaro, M.Sc. degree in
Prognostic model development with missing labels: A Manufacturing Systems, Monterrey
condition-based maintenance approach using machine Institute of Technology and Higher
learning. Business & Information Systems Engineering, Education, Campus Queretaro, and BSc
61(3), 327–343. https://doi.org/10.1007/s12599-019- degree in Engineering, Monterrey
00596-1 Institute of Technology and Higher Education, Campus
Cuernavaca. Dr. Ortega-Moody's research interests include
intelligent control systems and heuristic algorithms applied
BIOGRAPHIES to industrial processes and automation. Another area that
interests him is the application of mechatronics in projects
Tyler Ward was born in Weston, West
that include sustainable and renewable energy processes and
Virginia in 1999. He received a B.S. the food processing industry. He would also like to explore
degree in Computer Science from the possibility of applying mechanical design and control to
Morehead State University (MSU) in
the same. Currently, he is on a new application of physics
2021, and an M.S. in Engineering &
engines that he calls Virtual Mechatronics which seeks to
Technology Management in 2023 from
virtualize industrial scenarios for training and education,
the same institution. He is currently
simulation, and industrial virtual commissioning
employed as a Research Associate by
applications.
MSU’s Virtual Reality Laboratory in the
Department of Engineering Sciences. His research interests Selva Staub has extensive experience
include computer vision and machine learning, and to date he in the private and public sectors. She
has published several papers in journals and conferences in started her career as an assistant
these areas. He is currently a member of several professional professor in her native country, Turkey.
organizations, including the Institute of Electrical and After relocating to the United States, Dr.
Electronics Engineers (IEEE) and the Association of Staub worked as a Project Manager for
Technology, Management, and Applied Engineering overseas operations at Consumers
(ATMAE). He served on the board of directors for the latter Energy (Jackson, Michigan). Also in Michigan, Dr. Staub
organization from 2022 to 2023. worked at Eastern Michigan University as Regulatory
Compliance Officer, and Vice President of Operations at
Kouroush Jenab received his BSc InterConnect of Ann Arbor, a software development
degree from the Industrial Eng. company. In 2004, the Staubs decided to move to North
Department at the Isfahan University of
Carolina. Dr. Staub has worked at the non-profit
Technology (1989), M.Sc. degree from
organization, East Carolina Community Development, as
the Industrial Eng. Department at Tehran
advisor to the President, and worked at Community Colleges
Polytechnic (1992), Ph.D. degree from
System, as program director and strategic planner as well as
Industrial Eng. Department at IUST, and instructor. After 17 years in USA, Dr. Staub joined Haliç
Ph.D. degree from Department of University as Assistant Professor. Recently, she joined
Mechanical Engineering at the
Bandırma Onyedi Eylül University and teaches both graduate
University of Ottawa, Canada (2005). He served as a
and undergraduate programs. She is currently on the board of
Technical Lead and Senior Manager in auto, telecom,
Continues Education department and Foreign Languages
aerospace, and high-tech industries for 20 years. He joined
program at Bandırma Onyedi Eylül University.
National Research Council Canada as a Research Officer
where he conducted several international projects in Rail
Industry. Currently, Dr. Jenab is an associate professor in the
Department of Engineering Sciences at Morehead State
University. Dr. Jenab is a senior member of IEEE, IIE. He
has published over 170 papers in international journals and
conferences based on his experiences in industries. Dr.
Jenab’s fields of research are Reliability, Quality,
Manufacturing Systems, Human Reliability & Error, and
Maintenance Engineering.

20