Advancements in Predictive Maintenance: A Bibliometric Review of Diagnostic Models Using Machine Learning Techniques

Abstract

This bibliometric review investigates the advancements in machine learning techniques for predictive maintenance, focusing on the use of Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) for fault detection in wheelset axle bearings. Using data from Scopus and Web of Science, the review analyses key trends, influential publications, and significant contributions to the field from 2000 to 2024. The findings highlight the performance of ANNs in handling large datasets and modelling complex, non-linear relationships, as well as the high accuracy of SVMs in fault classification tasks, particularly with small-to-medium-sized datasets. However, the study also identifies several limitations, including the dependency on high-quality data, significant computational resource requirements, limited model adaptability, interpretability challenges, and practical implementation complexities. This review provides valuable insights for researchers and engineers, guiding the selection of appropriate diagnostic models and highlighting opportunities for future research. Addressing the identified limitations is crucial for the broader adoption and effectiveness of machine learning-based predictive maintenance strategies across various industrial contexts.

1. Introduction

In recent years, the advancements in artificial intelligence (AI) and machine learning (ML) have revolutionized various industrial applications, particularly in the realm of predictive maintenance through automation [1]. Predictive maintenance uses these technologies to anticipate and mitigate potential failures in machinery, thereby enhancing operational efficiency and reducing maintenance costs. One critical component within railway systems that benefits significantly from predictive maintenance is the wheelset axle bearing, whose failure can lead to catastrophic outcomes, including derailments.

Machine learning, a subset of AI, provides powerful tools for developing predictive models capable of identifying faults in axle bearings through the analysis of acoustic signal data. Among the number of existing ML techniques, Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) stand out due to their robust performance in classification and regression tasks. ANNs, with their ability to model complex, non-linear relationships, are well-suited for handling large datasets, while SVMs handles data requiring high accuracy and precision, especially with small-to-medium-sized datasets [2].

Several predictive models are employed in the lifecycle of predictive maintenance, each serving distinct purposes. Some models focus on detecting anomaly sounds, while others predict potential faults and failures [3]. For instance, sound detection models aim to recognize whether recorded sounds indicate normal or irregular operations. Neural networks have been used in such applications due to their ability to model complex patterns and relationships within large datasets [4]. These networks consist of layers of interconnected nodes or neurons that learn to perform tasks by considering examples without task-specific programming [5]. For sequential data like time series, which are common in equipment condition monitoring, neural networks offer significant advantages [6].

Similarly, SVMs are well-regarded for their effectiveness in classification and regression tasks, particularly when dealing with small-to-medium-sized datasets [7]. SVMs are capable of modelling complex nonlinear relationships and are especially useful in fault classification tasks in predictive maintenance, where accuracy and precision are paramount [8]. These algorithms create hyperplanes in multidimensional spaces to separate different classes, making them highly valuable for accurately classifying faults in predictive maintenance scenarios [9].

This study aims to develop and compare the performance of ANN- and SVM-based diagnostic models for identifying faults in wheelset axle bearings using acoustic signal data. The comparison will provide insights into the relative strengths and limitations of each approach, guiding the selection of the most effective model for real-world application in the South African railway context. Predictive maintenance aims to anticipate equipment failures and schedule necessary repairs or replacements before issues become critical. By integrating these advanced diagnostic models, railway operators can achieve timely maintenance interventions, thereby enhancing safety, reliability, and operational efficiency.

The significance of this research lies not only in its potential to improve the maintenance strategies of railway systems but also in its broader contribution to the field of AI-driven predictive maintenance. The findings from this comparative analysis will offer valuable guidance to engineers and researchers in selecting appropriate diagnostic models, ensuring data-driven, precise maintenance actions that reduce unplanned downtime and associated costs.

1.1. Objectives

The use of machine learning in predictive maintenance has emerged as a prominent area of research. The integration of machine learning can significantly improve the effectiveness and efficiency of predictive maintenance strategies. This bibliometric analysis seeks to explore available literature on the use of machine learning in in predictive maintenance by looking at advancements in the field and suggesting emerging areas of research. This will be achieved through the following objectives:

• To identify and analyse the key research trends and patterns in the field of diagnostic models for predictive maintenance.
• To evaluate the contributions and impact of different machine learning techniques, particularly ANN and SVM, in the context of fault detection in axle bearings.
• To compare the effectiveness of ANN and SVM models based on bibliometric data, highlighting their strengths and limitations in predictive maintenance applications.
• To provide insights and recommendations for future research directions by identifying gaps and emerging themes in the literature.

By achieving these objectives, this bibliometric analysis aims to contribute to a deeper understanding of machine learning in predictive maintenance.

1.2. Overview

The rest of this review paper proceeds as follows: Section 2 presents the materials and methods embraced in this study, emphasizing the keywords and bibliometric review process. The review results are presented in detail in Section 3, along with the summaries of the most influential studies in literature. Section 4 then presents the thematic analysis, highlighting the relevant themes as well as the key developments under each theme. We conclude the paper in Section 5, presenting the main observations, the limitations, contributions of the work, as well as the direction for future studies.

2. Materials and Methods

This bibliometric review aims to systematically examine the literature on diagnostic models for predictive maintenance, with a particular focus on machine learning techniques such as Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs). A comprehensive search was conducted on Scopus and Web of Science. The search terms included “diagnostic model”, “machine learning”, “neural networks”, “support vector machine”, “predictive maintenance”, and “condition-based maintenance”. Table 1 summarizes the relationships between the keywords of the study as searched.

Table 1. Keywords.

Keyword	Synonyms
Predictive maintenance	condition-based maintenance
Diagnostic models	machine learning, neural networks, support vector machine

The search was limited to peer-reviewed journal articles, conference papers, conference reviews, and review articles published between 2000 and 2024, as these have been reviewed, ensuring the quality and merit of the articles. The study is not limited to any subject areas. The source type, source title, and other filters were also left unfiltered. The search yielded 1288 papers on Web of Science and 3746 papers on Scopus. This study used Biblioshiny to merge the data and remove duplicates, in which 1060 duplicates were removed. Microsoft Excel and Biblioshiny were used to analyse the data. The bibliometric analysis process is outlined in Figure 1.

Figure 1. Bibliometric review process.

3. Results

The design science theoretical framework provides a structured approach to understanding the development and application of machine learning (ML) techniques for diagnostic models. This framework emphasizes the creation and evaluation of artifacts designed to solve identified problems, thereby contributing to both practical and theoretical knowledge [10]. This review uses the design science framework to examine the landscape of diagnostic models utilizing ML techniques, highlighting key trends, influential authors, and significant publications.

3.1. Main Publication Information

This section provides an overview of the main publication information related to diagnostic models using ML techniques. The analysis includes the annual scientific production, distribution of publications by country, the twenty most relevant authors, and the twenty most influential publications in the field. Table 2 presents the main publication information. The contribution of 10,942 authors has resulted in 3923 publications to 2638 journals and conferences between 2000 and 2004, with an annual growth rate of 18.44%.

Table 2. Main publication information.

Description	Results
Main Information
Timespan	2000:2024
Sources (journals, books, etc.)	1651
Documents	3923
Annual growth rate %	18.44
Authors
Authors	10,942
Authors of single-authored documents	129
Document Types
Articles	597
Review articles	36
Conference paper articles	1897
Conference paper reviews	108

3.2. Annual Scientific Production

The annual scientific production of diagnostic models using ML techniques has shown a significant upward trend over the past decade. This increase reflects the growing interest and advancements in the application of ML for diagnostic purposes across various fields, including healthcare, engineering, and finance. Figure 2 illustrates the annual growth rate of publications within this field. Before 2017, there were minimal publications, with very slow growth. From 2017, there was noticeable growth of publications which were due to the advancements and increased interest in the application of machine learning techniques for diagnostics. In 2023, there was a peak in publications, and this was probably due to the improvements in computational power, data availability, and machine learning algorithms.

Figure 2. Annual growth rate (this review was completed mid-year 2024, explaining the drop in publications for 2024).

3.3. Most Influential Publications

The twenty most influential publications were determined based on citation metrics and their contribution (see Table 3). Key papers include foundational works on neural networks, Support Vector Machines, and ensemble methods, which have provided the theoretical and practical foundation for modern diagnostic models. These publications are summarized in Table 4.

Table 3. Most influential publications.

Rank	Author	Year	Publication	Total
1	Ali, JB	2015 [11]	Mechanical Systems and Signal Processing	449
2	Tian, Z	2011 [12]	Renewable Energy	330
3	Xu, Y	2019 [13]	IEEE Access	316
4	Tian, Z	2012 [14]	Journal of Intelligent Manufacturing	305
5	Zhang, B	2019 [15]	Computers in Industry	276
6	Chen, X	2022 [16]	Neural Computing and Applications	264
7	Tran, V	2012 [17]	Mechanical Systems and Signal Processing Proceedings of the ISA/IEEE 2005 Sensors for Industry Conference, Sicon’05	228
8	Mahamad, A	2010 [18]	Computers and Mathematics with Applications	222
9	Xiao, S	2017 [19]	31st AAAI Conference on Artificial Intelligence, AAAI 2017	190
10	Wu, S	2007 [20]	IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and Humans	187
11	Unal, M	2014 [21]	Measurement: Journal of the International measurement Confederation	186
12	Zhu, J	2020 [22]	IEEE Sensors Journal	183
13	Essien, A	2020 [23]	IEEE Transactions on Industrial Informatics	175
14	Kanawaday, A	2017 [24]	Proceedings of the IEEE International Conference on Software Engineering and Services Sciences, ICSESS	175
15	Paolanti, M	2018 [25]	2018 14th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications, MESA 2018	168
16	Amruthnath, N	2018 [26]	2018 5th International Conference on Industrial Engineering and applications, ICIEA 2018	168
17	Sipos, R	2014 [27]	Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining	160
18	Hinchi, A	2018 [28]	Procedia Computer Science	159
19	Huda, A	2013 [29]	Applied Thermal Engineering	158
20	Cheng, F	2018 [30]	IEEE Transactions on Sustainable energy	131

Table 4. Summary of the most influential publications.

Rank	Summary
1	The article presents an approach for predicting the remaining useful life (RUL) of rolling element bearings (REBs) using a combination of the Weibull distribution and a Simplified Fuzzy Adaptive Resonance Theory Map (SFAM) neural network. The method involves using the Weibull distribution to fit measurements during the training phase to smooth out fluctuations in the time domain, which is followed by training the SFAM neural network based on these fitted measurements. The SFAM network is then tested with real-time measurements without fitting, and a smoothing algorithm is applied to improve RUL predictions. Experimental results demonstrate that this approach can reliably predict the RUL of REBs based on vibration signals, offering a promising tool for Prognostics and Health Management (PHM) in various mechanical systems [11].
2	This study proposes a condition-based maintenance (CBM) strategy for optimizing the maintenance of wind power generation systems. The CBM approach uses condition monitoring data from wind turbine components to make informed maintenance decisions that minimize operational and maintenance costs. Unlike traditional methods that handle turbine components separately, this approach considers the economic dependencies among multiple turbines and their components within a wind farm. The proposed policy is defined by two failure probability threshold values at the wind turbine level, allowing for the calculation of failure probabilities and optimal maintenance actions. A simulation method is developed to evaluate the cost-effectiveness of the CBM policy, demonstrating its advantage over constant-interval maintenance policies through numerical examples and a comparative study. The results indicate that the CBM approach can significantly reduce maintenance costs, improving the reliability and efficiency of wind power systems [12].
3	This article proposes a method of fault diagnosis in smart manufacturing. The proposed method integrates digital twins and deep transfer learning (DTL) to better the fault diagnosis process. The methodology involves two phases, a development phase which creates a high-fidelity virtual model of the physical system and used to train a deep neural network (DNN) for fault diagnosis using simulated data and a maintenance phase which adapts the pre-trained model to the physical system using DTL to account for real-time monitoring and predictive maintenance. This approach addresses common issues in traditional fault diagnosis methods, such as insufficient training data and differing data distributions, by using the virtual model to simulate various fault conditions and transferring learned knowledge to the physical domain. A case study in a car body-side production line is used to demonstrate the effectiveness of the proposed method, showing improvements in diagnosis accuracy and efficiency compared to traditional DNN-based methods [13].
4	The article develops an Artificial Neural Network (ANN) method to accurately predict the remaining useful life (RUL) of equipment using condition monitoring data. The proposed ANN model utilizes age and multiple condition monitoring measurements from both current and previous inspection points as inputs, and outputs the life percentage of the equipment. To improve prediction accuracy, the method involves fitting the measurement series with a generalized Weibull distribution function to reduce noise and implementing a validation mechanism during the training process to avoid overfitting. The approach is validated using real-world vibration monitoring data from pump bearings, demonstrating improved prediction performance compared to existing methods. The study underscores the ANN’s ability to provide reliable RUL predictions, facilitating effective condition-based maintenance and optimizing maintenance schedules [14].
5	This study proposes a method for assessing the performance degradation of bearings using Long Short-Term Memory (LSTM) Recurrent Neural Networks. It introduces a new indicator called “waveform entropy” (WFE), which captures the degradation state of bearings. A simulation model based on vibration response mechanisms is used to validate the proposed indicator. The LSTM network, optimized using particle swarm optimization, processes the WFE and other conventional indicators to predict bearing states and remaining useful life (RUL). Experimental results demonstrate that the LSTM model can accurately identify bearing degradation stages and provide reliable RUL predictions, offering a robust approach for condition-based maintenance and predictive maintenance in rotating machinery [15].
6	The article introduces a new deep learning model for predicting the remaining useful life (RUL) of aeroengines. This model uses the Transformer architecture, enhanced with a position-sensitive self-attention (PSA) mechanism to better capture local context, and a gated hierarchical long short-term memory (GHLSTM) network to model long-term dependencies at multiple time scales. The PSA mechanism addresses the traditional Transformer’s limitation in local context insensitivity, while the GHLSTM network improves the model’s temporal feature extraction capabilities. Experiments conducted on the C-MAPSS dataset demonstrate that this novel approach outperforms existing RUL prediction methods, achieving high accuracy and robustness in various operational conditions and fault modes. This advancement in RUL prediction contributes to more effective maintenance strategies and improved reliability in aerospace applications [16].
7	This article presents a three-stage method to assess machine performance degradation and predict the remaining useful life (RUL). The first stage involves using an autoregressive moving average (ARMA) model based on normal operating conditions to identify system behaviour and calculate degradation indices from residual errors. In the second stage, Cox’s Proportional Hazard Model (PHM) is employed to estimate the survival function of the system. The final stage uses Support Vector Machine (SVM) regression, combined with time-series forecasting techniques, to predict the RUL. The method is validated using data from the condition monitoring of a low methane compressor, demonstrating its effectiveness as a reliable tool for machine prognostics and maintenance decision-making [17].
8	This study investigates the use of Artificial Neural Networks (ANNs) to predict the remaining useful life (RUL) of rotating machinery, specifically bearings. The proposed method utilizes a Feedforward Neural Network (FFNN) with the Levenberg–Marquardt training algorithm. It inputs time, root mean square (RMS), and kurtosis values from both current and previous inspections, which are fitted using Weibull hazard rates to minimize noise and improve prediction accuracy. The output is the normalized life percentage of the bearing. The method is validated using vibration signal data from the Center for Intelligent Maintenance Systems at the University of Cincinnati, showing that the ANN model performs well in predicting the RUL of bearings, thereby improving condition-based maintenance strategies [18].
9	This article proposes an approach to modelling point processes using Recurrent Neural Networks (RNNs). The method involves two RNNs: one aligned with time series data to model the spontaneous background, and another aligned with asynchronous events to capture long-range dependencies. This composite neural network approach allows for end-to-end training and avoids the limitations of predefined parametric forms traditionally used in point process modelling. The model is applied to a predictive maintenance scenario using a large dataset of ATM event logs from a global bank. The results demonstrate that this RNN-based model significantly improves the accuracy of event type and timestamp predictions, showing its effectiveness and utility in real-world applications [19].
10	This paper develops an integrated neural network-based decision support system aimed at optimizing predictive maintenance for rotational equipment. It includes three main components: a vibration-based degradation database created through condition monitoring, an Artificial Neural Network (ANN) model to estimate the life percentile and failure times of roller bearings, and a cost matrix with a probabilistic replacement model to optimize expected maintenance costs. The ANN uses vibration data to predict machine life percentages, which inform maintenance decisions to balance costs between predictive and corrective actions. The proposed system is validated through experiments, showing its effectiveness in various industrial applications by reducing unexpected downtimes and maintenance costs [20].
11	This article presents a study focused on diagnosing faults in rolling bearings, through a hybrid approach combining envelope analysis, Fast Fourier Transform (FFT), and Hilbert Transform for feature extraction from vibration signals. These features are then used to train an Artificial Neural Network (ANN), with the architecture optimized through a genetic algorithm (GA) for accurate fault classification. The study demonstrates that the proposed GA-optimized ANN model significantly enhances fault detection accuracy, achieving up to 98% success in distinguishing fault-free and faulty bearings, as well as identifying fault locations [21].
12	This study presents a different approach to diagnosing bearing faults in machinery using deep transfer learning. The method uses a Convolutional Neural Network (CNN) and introduces domain adaptation to handle varying operational conditions, which often cause differences in data distribution between the source and target domains. Through a combination of multiple Gaussian kernels to measure domain discrepancies, the model improves its ability to transfer learned features across different working conditions without requiring extensive labelled data in the target domain. The study validates this approach with real data, demonstrating that the proposed method outperforms traditional CNN and other transfer learning methods in terms of classification accuracy and adaptability [22].
13	This study proposes an advanced deep learning architecture designed for machine speed prediction in smart manufacturing environments. This model uses a Convolutional Long Short-Term Memory (ConvLSTM) neural network combined with autoencoders to improve time-series forecasting capabilities for machine performance data. Through convolutional layers for feature extraction and LSTM networks for sequential learning, the model addresses the challenges of predicting machine speed across varying conditions. The study demonstrates the effectiveness of this hybrid architecture through empirical evaluations using real-world data, where the proposed model outperforms traditional deep learning approaches, contributing to optimizing smart manufacturing processes, particularly in predictive maintenance and production scheduling [23].
14	This study explores how machine learning techniques, particularly ARIMA forecasting, can be applied to predict machine failures and improve productivity in industrial settings using IoT sensor data. The study focuses on a slitting machine, where sensor data on parameters like tension, pressure, width, and diameter are collected in real-time. The data are processed using various machine learning models, including neural networks and Support Vector Machines (SVMs), to identify patterns and predict quality failures. The integration of ARIMA forecasting allows for proactive maintenance, helping to prevent downtime and product defects. The results indicate that deep neural networks perform the best for classifying machine performance, offering a valuable approach for industrial prognostics and efficiency improvements [24].
15	This article outlines a strategy for implementing predictive maintenance in industrial environments using a Random Forest classifier. The study was conducted on a cutting machine in a real industrial setting, where the predictive models successfully estimated the spindle rotor status and machine performance. The approach achieves 95% accuracy, demonstrating its effectiveness in reducing downtime, improving maintenance planning, and optimizing operational efficiency within the framework of Industry 4.0 [25].
16	This article explores the application of unsupervised learning algorithms in detecting faults in industrial machinery for predictive maintenance. The authors focus on vibration data collected from an exhaust fan, testing several algorithms including Principal Component Analysis (PCA), K-Means, Fuzzy C-Means, hierarchical clustering, and Gaussian Mixture Models (GMMs). The study benchmarks these methods based on their accuracy and ability to detect faults early. PCA and T2 statistics provided earlier fault detection compared to clustering methods, while clustering proved more effective at identifying various fault levels. The research concludes that T2 statistics are ideal for detecting faults early, but clustering techniques offer a more meaningful understanding of different fault stages, making them suitable for continuous monitoring of machine health [26].
17	This study proposes a multiple-instance learning (MIL) technique to extract predictive features from equipment event logs, which are typically not designed for failure prediction but contain rich operational data. By applying machine learning models to these logs, the method allows early detection of failures, enabling proactive maintenance scheduling. The study uses real-world datasets from large fleets of medical equipment, with experimental results demonstrating the effectiveness of the approach. The proposed methodology was successfully deployed by a major medical device provider, highlighting its practical utility in improving maintenance efficiency and reducing unexpected downtime [27].
18	This article introduces a deep learning model for estimating the remaining useful life (RUL) of rolling element bearings, which are prone to failure in rotating machinery. The proposed method combines Convolutional Neural Networks (CNNs) to extract features from raw sensor data and Long Short-Term Memory (LSTM) networks to model the temporal aspects of the degradation process. This end-to-end model is trained directly on high-frequency vibration data without the need for manual feature extraction. Experimental results on data from FEMTO-ST Institute show that the model is effective at predicting RUL, outperforming traditional methods [28].
19	This study discusses the use of infrared thermography as a non-invasive, reliable method for detecting thermal defects in electrical equipment. The technique enables early identification of potential failures by capturing thermal images of equipment during operation without interrupting processes. The study explores the use of multilayer perceptron networks (MLP) and discriminant analysis classifiers for classifying thermal conditions into ‘defect’ or ‘no defect’ categories based on temperature data. Infrared thermography is highlighted as an effective tool for both preventive and predictive maintenance, reducing downtime, improving equipment lifespan, and lowering maintenance costs. The discriminant analysis classifier achieved a classification accuracy of 82.4%, demonstrating its effectiveness in improving predictive defect diagnosis [29].
20	This article focuses on developing a novel method for predicting the remaining useful life (RUL) of wind turbine gearboxes based on current signal analysis. The authors propose a hybrid approach combining adaptive neuro-fuzzy inference systems (ANFIS) and particle filtering (PF) to model the degradation of gearboxes. This method utilizes the gearbox’s generator stator current signals to extract fault features and predict the gearbox’s RUL. By utilizing ANFIS to learn the state transition function and applying PF for recursive updates, the model offers accurate predictions of the gearbox’s fault progression and RUL [30].

The majority of the reviewed works primarily rely on sensor data such as from vibration sensors, thermal sensors, and/or acoustic signals. However, the study by Sipos et al. [27] focuses on analysing historical log data to predict maintenance needs, distinguishing it from the other studies. Equipment and/or log data provide a rich history of machine behaviour, but they often lack granularity compared to continuous sensor data. Combining log-based analysis with real-time IoT sensor data, as seen in Kanawaday and Sane [24], could provide a more comprehensive view of machine health. This hybrid approach could enable both immediate fault detection and long-term trend analysis, resulting in improved predictive capabilities.

Based on the works reviewed in this review, there are research gaps that exist that can inform future studies, with one being adaptability, which refers to the ability of models to generalize on unseen data, in this paper, referring to generalizing across varying operating conditions and machinery types [31]. While many studies, including those reviewed, have developed effective models for specific use cases—such as using Convolutional Neural Networks, LSTM networks, and neuro-fuzzy systems—these models often struggle when applied in different environments or machinery with varied operational parameters [23,30]. For instance, the reviewed works on wind turbine gearbox fault prediction and rolling bearing fault detection used tailored models that may not perform equally well under different load conditions or in environments with fluctuating operational characteristics [28,30]. However, there is progress being made on this gap through some research work, such as those focusing on transfer learning and domain adaptation. These demonstrate how predictive maintenance models can be tailored to different operational environments or changing working conditions, such as in the case of bearing fault diagnosis under variable loads and speeds. For instance, the paper on deep transfer learning for bearing fault diagnosis [22] highlights successful adaptability across different working conditions without requiring extensive labelled data, showcasing a model’s robustness in diverse industrial contexts. Similarly, current-based monitoring techniques in wind turbines, as discussed in the RUL prediction of wind turbine gearboxes [30], provide cost-effective, adaptable solutions that bypass the need for additional sensors, emphasizing the practicality of adaptability in reducing deployment complexity and costs. These studies demonstrate that while there is progress, more research is needed to improve the transferability of these models across different industries and machinery types.

The issue of explainability, which refers to the ability to interpret how machine learning models make decision, is another crucial research gap that has been inadequately addressed [32]. Machine learning models, particularly deep learning techniques such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks used in fault prognosis, often function as “black boxes”, limiting their explainability [28]. The study on unsupervised learning for fault detection and infrared thermography [29] also demonstrates that while these models show high accuracy, they often lack interpretability, which could prevent their widespread application. This highlights the need for future research to focus on integrating explainable AI (XAI) techniques into predictive maintenance frameworks, providing clearer insights into how predictions are generated.

Table 5 shows the comparison between the different neural networks, SVM, and other methods, showing the advantages and disadvantages of each method.

Table 5. Comparison table between the advantages and disadvantages of each method.

Model/Method	Advantages	Disadvantages
Genetic Algorithm Optimized Neural Network	Optimized for high accuracy, avoids local minima	Time-consuming, requires large data sets
Feedforward Neural Network (FFNN)	Simple architecture, good for basic tasks	Limited for complex time-series or sequence data
Artificial Neural Network (ANN)	Versatile, handles both classification and regression	Prone to overfitting, requires tuning
Deep Transfer Learning Neural Network	Adaptable across domains, reduces retraining	Depends on source domain data availability
Convolutional LSTM Neural Network	Combines feature extraction and sequence prediction	High computational cost, prone to overfitting
Support Vector Machine (SVM)	Effective for classification with limited data	Not adaptable to time-series data, sensitive to noise
Unsupervised Learning (k-means, hierarchical clustering)	No need for labelled data, good for early detection	Lower accuracy, prone to noise
IoT Sensor Data with Machine Learning	Real-time monitoring with IoT, adaptable	Dependent on sensor data quality, security issues
Log-based Predictive Maintenance	Adaptable to various industries, data-rich environments	Requires large datasets, hard to generalize
Explainable AI (XAI)	Transparency and interpretability in decision-making	Trade-off between explainability and performance

4. Thematic Analysis

Figure 3 shows a thematic map related to predictive maintenance. The map categorizes themes into four quadrants: Niche Themes (low relevance, high development), Motor Themes (high relevance, high development), Basic Themes (high relevance, low development), and Emerging or Declining Themes (low relevance, low development).

Figure 3. Thematic map showing relevance and development of themes.

The map highlights the importance and focus on predictive maintenance, using machine learning, and overall maintenance practices as core themes driving current research and development. Niche themes like condition-based maintenance and neural networks show high development but may need to integrate more into mainstream relevance. Declining themes like forecasting are considered less central to current trends in maintenance practices. Table 6 presents the themes derived by Biblioshiny of the 3923 papers. There are three separate themes with 12 occurrences for each, which is a cut number from the 250 code occurrences from the combined themes.

Table 6. Thematic analyses.

Theme	Code Occurrence	Description
Predictive Maintenance	pred maintenance (1195) machine learning (705) maintenance (818) machine-learning (533) learning systems (527) learning algorithm (354) anomaly detection (206) internet of things (232) predictive analytics (170) decision making (165) data analytics (125) artificial intelligence (149) Industry 4.0 (151)	This theme is of a proactive maintenance strategy that involves using data analysis tools and techniques to detect anomalies and predict equipment failures before they occur. This theme relies on real-time data collected from sensors and advanced algorithms, often incorporating machine learning and artificial intelligence to analyse the data. Organizations aims to schedule maintenance activities only when necessary, thereby minimizing downtime, reducing maintenance costs, and extending the lifespan of equipment. By predicting failures and addressing them proactively, organizations can improve the reliability and efficiency of their operations.
Condition-based maintenance (CBM)	condition-based maintenance (584) condition monitoring (412) neural networks (375) fault detection (348) support vector machines (197) failure analysis (156) preventive maintenance (127) classification (of information) (127) condition (107) vibration analysis (96) health (94) feature extraction (92)	This theme is of a maintenance strategy that monitors the actual condition of an asset to decide what maintenance needs to be performed. This approach is based on the real-time assessment of equipment condition using various sensors and diagnostic technologies. CBM aims to perform maintenance only when certain indicators show signs of decreasing performance or upcoming failure. By focusing on the actual condition rather than a predetermined schedule, CBM helps in preventing unexpected equipment failures, optimizing maintenance schedules, and reducing unnecessary maintenance activities, thereby saving costs and improving operational efficiency.
Forecasting	forecasting (454) deep learning (289) remaining useful lives (215) long short-term memory (161) convolutional neural networks (128) neural networks (114) convolution (105) convolutional neural network (97) deep neural networks (95) time series (90)	This theme involves predicting future maintenance needs and equipment performance based on historical data and trend analysis. This technique uses various statistical and machine learning methods to analyse past data and project future conditions. Forecasting helps in planning and scheduling maintenance activities in advance, ensuring that resources are available and potential issues are addressed before they become critical. By anticipating maintenance needs, organizations can improve the reliability of their equipment, reduce downtime, and manage maintenance budgets more effectively.

5. Conclusions

This bibliometric review has explored the advancements in machine learning techniques for predictive maintenance, with a particular focus on diagnostic models using Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs). The analysis outlines the progress made in applying these techniques to predict and mitigate potential failures in machinery. The review emphasizes the performance of ANNs in handling large datasets and modelling complex, non-linear relationships, making them suitable for detecting anomalies in acoustic signal data. SVMs, on the other hand, have shown high accuracy and precision in fault classification tasks, especially with small-to-medium-sized datasets, owing to their ability to create effective hyperplanes in multidimensional spaces.

5.1. Limitations

Despite these advancements, the review identified several limitations that must be addressed to improve the practical application of these models. These include the following:

• Data quality: Many models, such as those by Ali et al. [11] and Mahamad et al. [18], heavily depend on the quality and quantity of condition monitoring data like vibration signals. This dependency can be a significant drawback if such data are not readily available or is noisy, affecting the model’s accuracy and generalizability.
• Significant computational resources: Approaches like those proposed by Tian et al. [14] and Zhang et al. [15] often require significant computational resources for training and inference, posing a challenge for implementation in environments with constrained resources.
• High computational power: The complexity of models such as the transformer-based approach by Chen [16] and the digital-twin-assisted method by Xu et al. [13] demand high computational power and huge amounts of data, which might not always be accessible.
• Adaptability of these models to different machinery and operational conditions: For instance, the methodologies developed by Tran et al. [17] and Tian et al. [12] might not generalize well across different types of equipment without customization to the data and parameter tuning.
• Interpretability and underlying assumptions: Models such as those by Xiao et al. [19] and Wu et al. [20] also face challenges related to model interpretability and the assumptions underlying their predictive maintenance strategies. The proportional hazard model combined with SVM by Tran et al. [17] assumes proportionality, which may not always hold, while the cost matrix approach by Wu et al. [20] assumes constant maintenance costs, which might not reflect the dynamic nature of real-world scenarios.

These limitations highlight the need for further research to improve the robustness, interpretability, and applicability of predictive maintenance models in diverse industrial contexts.

There are also limitations to this study which includes access to data only being through Scopus and web of science which might have excluded other papers which are not available through these databases. The review may not capture the full breadth of research in predictive maintenance, potentially overlooking significant contributions from other academic and industrial sources. This limitation suggests the need for a more comprehensive literature search that includes additional databases and grey literature to provide a holistic view of the field.

5.2. Contributions

The work makes three major contributions to the body of knowledge, as follows:

• Providing insights on which model, between the two, is more accurate, reliable, and computationally efficient;
• Analysing patterns that exists within the field, therefore identifying research trends in predictive maintenance;
• Providing an understanding of the strengths and limitations of ANN and SVM models, through which stakeholders can make informed decisions about selecting and implementing the most appropriate diagnostic models for their specific needs.

5.3. Future Work

While machine learning-based predictive maintenance models are promising for improving operational efficiency and safety in industrial applications, addressing their current limitations through continued research and development is essential. This will facilitate the broader adoption of these advanced techniques, ultimately leading to more reliable and cost-effective maintenance strategies across various industries. The directions for future work can include the following:

• Development of more efficient algorithms that require less computational power without compromising accuracy;
• Applying transfer learning techniques to adapt models trained on one type of machinery to another, therefore allowing adaptability of developed models.