Vehicle Maintenance Demand Prediction: A Survey

Abstract

The implementation of predictive maintenance (PdM) represents a pivotal strategy for enhancing operational efficiency and reducing costs within the automotive industry. While extensive research has yielded sophisticated machine learning and deep learning models capable of accurately predicting failures in specific vehicle components—including suspensions and batteries—these approaches predominantly operate in isolation. A principal challenge remains the transition from predicting failures in single components to managing the health of the vehicle as an integrated system. This review systematically examines recent advances in forecasting methodologies for vehicle service parts and predictive maintenance technologies, while critically evaluating the strengths and limitations of diverse approaches. The analysis highlights crucial unresolved technical and strategic challenges that must be addressed to realize fully integrated predictive maintenance systems for next-generation vehicles, with particular emphasis on the transition from isolated component monitoring to holistic Vehicle Health Management (VHM).

1. Introduction

Predictive maintenance (PdM) represents a paradigm shift in the automotive industry, leveraging data from in-vehicle sensors to forecast failures, thereby optimizing inventory management and minimizing operational downtime. This paradigm shift from conventional reactive and preventive maintenance approaches is predominantly driven by advancements in machine learning (ML) and deep learning (DL) methodologies. As evidenced by the publication trend depicted in Figure 1, research interest in PdM has surged over the past decade, a trend that accelerated markedly following the widespread adoption of deep learning methodologies around 2015. Subsequent research has demonstrated the considerable efficacy of models such as Support Vector Machines (SVMs), Long Short-Term Memory networks (LSTMs), and Convolutional Neural Networks (CNNs) in accurately predicting faults in specific components, including the suspension, brakes, and batteries.The application of AI in heavy-duty trucks for predictive maintenance and other tasks has also been extensively reviewed, highlighting the broad impact of these technologies across different vehicle types [1].

Despite these advances, a significant gap remains between isolated component-specific prognostics and the overarching goal of comprehensive Vehicle Health Management (VHM). Given that a vehicle operates as a complex interconnected system, a holistic assessment of its overall condition is imperative for effective PdM.

This review critically examines recent advancements in forecasting methodologies for vehicle service parts and predictive maintenance technologies. It assesses the strengths of diverse approaches, ranging from statistical ML to digital twins, with a specific focus on the critical challenge of transitioning from component-level to vehicle-level prognostics. Furthermore, this paper synthesizes the current research landscape to delineate the unresolved technical and strategic challenges that must be addressed to realize fully integrated and efficient vehicle maintenance systems.

2. Maintenance Costs

According to a report by the U.S. Department of Energy [2], the implementation of predictive maintenance can yield savings of approximately 8% to 12% compared to planned preventive maintenance. This approach not only extends asset lifespan and reduces unplanned downtime but also lowers costs associated with spare parts inventory and labor. Furthermore, it enhances operational safety, improves plant reliability, and optimizes equipment performance, leading to direct energy conservation. However, the implementation of PdM requires considerable initial capital investment for diagnostic equipment acquisition and configuration, in addition to expenditures for specialized personnel training. Despite these upfront costs, empirical studies and industrial surveys indicate that the long-term benefits of predictive maintenance substantially outweigh the initial investments. For instance, organizations implementing PdM programs have been reported to eliminate 70–75% of asset failures, reduce overall maintenance costs by 25–30%, and increase production output by 20–25%. The return on investment (ROI) for such initiatives averages tenfold, thereby substantiating PdM as a strategically viable and economically sound investment.

Figure 2 illustrates the distribution of reference types analyzed in this survey. The prevalence of journal articles, followed by conference papers, underscores the robust academic foundation upon which predictive maintenance research is built.

3. Predictive Maintenance in Vehicle Systems

As outlined by Sajid et al. [3], predictive maintenance (PdM) strategies can be systematically classified into several distinct methodological paradigms, as illustrated in Figure 3.

Physical Model Approach: This methodology employs physics-based or mathematical models to assess component degradation. Its predictive accuracy is intrinsically dependent on the fidelity of the underlying model, necessitating the frequent use of statistical techniques for model validation [4].

Knowledge-Based Approach: Representative techniques in this domain include expert systems and fuzzy logic [5].

Data-Driven Approach: Capitalizing on advances in computational power and the availability of large-scale datasets, this paradigm is primarily subdivided into statistical models [6], stochastic models [7], and machine learning models.

Digital Twin Approach: This advanced methodology involves the integration of real-time data and analytical models to create a dynamic bidirectional link between a physical asset and its high-fidelity digital representation [8].

3.1. Physics-Based Models

Physics-based models constitute a foundational methodology for predictive maintenance, relying on detailed mathematical representations of physical degradation processes. These models incorporate parameters that capture the thermal, mechanical, chemical, and electrical properties of system components. The primary advantage of this methodology lies in its ability to provide accurate descriptions of system states, based on the fundamental physical comprehension of failure mechanisms. However, model accuracy is critically dependent on domain-specific expertise and the robustness of the mathematical formulations. Conversely, this approach is often hampered by high complexity, significant implementation costs, and the system-specific nature of the developed models, which constrains their generalizability and scalability.

The study presented in [4] introduces a streamlined physics-based modeling framework for the RHEvo system, a production-line roller hemming apparatus comprising spring-loaded components, rollers, and bearing mechanisms. Following model development, neural networks are employed to assess real-time component conditions. A detailed analysis of spring behavior identifies a significant correlation between material aging and a reduction in the elastic coefficient due to fatigue-induced degradation. This correlation facilitates subsequent quantification of the remaining useful life (RUL) through stochastic modeling. The proposed methodology demonstrates versatility for application in various spring-dependent systems, such as vehicle suspension systems, firearm recoil mechanisms, and combustion engine valve assemblies, where analogous wear patterns are observed.

Complementing this research, ref. [9] proposes an integrated predictive maintenance framework that incorporates vibration analysis for fault identification, condition assessment, and service life prognosis. The developed solution emphasizes two critical aspects: (1) the necessity of a comprehensive understanding of system dynamics for robust algorithm development, and (2) the importance of failure mechanism analysis for reliable performance prediction under variable operational conditions. Furthermore, a novel decision-support module is introduced to enable optimal technique selection and monitoring parameter optimization. Empirical validation across diverse industrial case studies confirms the framework’s practical adaptability and implementation potential.

While physics-based models provide high accuracy through a detailed physical understanding of degradation processes, their implementation faces significant limitations. The requirement for extensive domain expertise, high development costs, and system-specific nature restrict their scalability and adaptability across different vehicle platforms. Furthermore, these models often struggle to capture complex interactions between multiple components in modern vehicle systems. These limitations necessitate the exploration of alternative approaches that can leverage accumulated expert knowledge in a more flexible and transferable manner.

3.2. Knowledge-Based Models

Knowledge-based modeling approaches leverage domain expertise by codifying and automating specialist knowledge within computational systems. These models typically employ expert systems, which integrate a knowledge base of field-specific information with an inference engine to emulate human decision-making processes. The two predominant implementation methodologies are (1) rule-based systems, which are valued for their interpretability and simplicity but are limited in managing complex conditional relationships or extensive rule sets, and (2) fuzzy logic systems, which emulate human reasoning through intuitive graded representations of system states. The efficacy of knowledge-based systems, much like their physics-based counterparts, is fundamentally contingent upon the accuracy and completeness of the underlying knowledge representation and is characterized by strong domain specificity.

Contemporary research demonstrates a growing emphasis on hybrid approaches that integrate knowledge-based and data-driven methodologies. This trend is exemplified by [5], who developed an innovative fault diagnosis system for electric vehicles. To address the computational constraints inherent in vehicular embedded systems, the authors devised a hybrid architecture that integrates a neural network with a fuzzy logic classifier. Their methodology involves (1) collecting operational data from three distinct electric vehicle platforms, (2) training neural networks to identify fault–feature correlations, and (3) deploying a fuzzy logic system for real-time condition assessment. This optimized solution achieves a diagnostic accuracy of 88% while complying with the stringent computational limitations of onboard systems, thereby providing effective capabilities for anomaly detection and failure prevention.

As illustrated in Figure 4, which provides a comprehensive overview of the technical methods employed in predictive maintenance research, deep learning approaches predominate in the recent literature. This distribution reflects the evolving landscape of PdM methodologies and their application frequencies across different automotive subsystems.

In a related domain, recent work by Choi and Lee [10] provides a comprehensive review of fault diagnosis and tolerance techniques for automotive motor control systems, including inverters, permanent magnet synchronous motors, and sensors. Their statistical survey of common fault modes and failure rates offers valuable insights for reliability assessment methodologies in automotive applications. The authors discuss various fault diagnosis and tolerant control methods, emphasizing the critical importance of accurate and timely detection to prevent fault propagation and catastrophic system failures. This knowledge-based approach demonstrates how systematic analysis of failure patterns can effectively inform predictive maintenance strategies for critical vehicle subsystems.

In summary, knowledge-based approaches successfully bridge the gap between expert knowledge and computational systems, offering interpretable solutions for fault diagnosis. However, their effectiveness remains constrained by the quality and completeness of the underlying knowledge representation, and they often struggle with complex nonlinear relationships that are difficult to codify through explicit rules. The dependency on human expertise and limited adaptability to novel fault scenarios highlight the need for more autonomous learning methods that can extract patterns directly from operational data, thereby motivating the emergence of sophisticated data-driven approaches.

3.3. Data-Driven Methods

Figure 4 provides a comprehensive overview of the technical methods employed in predictive maintenance research, demonstrating the predominance of deep learning approaches in contemporary literature. This distribution reflects the evolving landscape of PdM methodologies and their relative application frequencies across various automotive subsystems.

3.3.1. Statistical and Stochastic Approaches

Statistical and stochastic modeling techniques provide robust frameworks for analyzing complex systems characterized by uncertain temporal evolution. These methods are particularly valuable for predicting key reliability metrics—including survival probability and mean time to failure—in critical mechanical systems such as electric vehicle batteries and gear trains. Recent developments highlight their increasing effectiveness in fault detection and system health monitoring applications.

(1) Battery Fault Diagnosis Applications

Recent investigations demonstrate innovative implementations of statistical methods in battery monitoring:

In [11], the authors developed a battery fault diagnosis system that integrates neural networks with a statistical analysis of voltage variations. Their methodology employs a Gaussian distribution-based 3σ multilevel screening process, which demonstrates high predictive accuracy for battery cell failures through large-sample statistical calibration.

Similarly, the research presented in [12] introduces a monitoring framework for series-connected lithium-ion batteries that utilizes mean squared error (MSE) analysis for voltage state assessment. This approach applies Z-score parameters to identify abnormal voltage conditions, thereby facilitating early fault detection.

Advancing this field further, Zhang et al. [13] proposed a novel battery fault prediction method combining Gradient Boosting Decision Tree (GBDT) with Isolation Forest and Boxplot techniques. Their GBDT–iForest–Boxplot approach achieves accurate voltage prediction six minutes in advance and provides effective fault diagnosis by comprehensively considering battery characteristics and driving behavior. This methodology demonstrates significant innovation in addressing traditional limitations through multi-state and multidimensional analysis, offering robust support for electric vehicle safety warning systems.In addition to fault diagnosis, accurate state of charge (SOC) estimation is crucial for battery management. [14] proposed a comprehensive framework for SOC estimation using various machine and deep learning techniques, which is essential for safe and efficient operation of electric vehicles.

(2) Mechanical System Monitoring

Advanced statistical techniques exhibit considerable potential in mechanical component analysis: For instance, Ashok Raj et al. [6] proposed a vibration-based gear fault detection methodology utilizing fourth-order statistical moments and kurtosis indices. Their technique applies empirical mode decomposition (EMD) to extract intrinsic mode functions (IMFs) from vibration signals, enabling precise fault severity estimation in both temporal and spectral domains.

Complementary research by Garay et al. [15] modeled degradation processes using stochastic functions that account for parameter variability, while Shen et al. [9] developed a two-stage Wiener process for degradation analysis in lithium-ion batteries.

(3) Multivariate Statistical Techniques

Modern industrial systems benefit substantially from advanced multivariate analytical approaches. Among these methodologies, Principal Component Analysis (PCA) has established itself as a powerful multivariate technique with extensive industrial applications, including process monitoring, fault detection/diagnosis, and sensor validation across multiple sectors. A key advantage of PCA is its capacity for dimensionality reduction during exploratory data analysis, which significantly decreases the computational complexity while preserving critical information. Researchers in [16] demonstrated this capability through the development of a PCA-based fault amplification method designed to trace fault propagation pathways in industrial systems. While this methodology proves effective for systems with limited variables, its computational demands become prohibitive as system complexity and variable dimensionality increase, presenting significant scalability challenges for large-scale applications.

In a related development, Phan et al. [17] developed an innovative forecasting framework that combines Ensemble Empirical Mode Decomposition (EEMD) with Dynamic Mode Decomposition (DMD) to address complex temporal patterns in parts ordering. Their method decomposes demand signals into intrinsic mode functions (IMFs) and models dynamical transitions, achieving a 30% reduction in forecasting errors compared to conventional techniques when validated on 8,605 service parts. Despite demonstrating excellence in capturing seasonal and trend patterns, the approach remains challenged by erratic demand spikes and unprecedented shift events—a limitation prevalent across most data-driven methods in this domain.

(4) Hybrid Approaches

In complex systems, faults often propagate across interconnected variables, significantly complicating the fault identification process. Granger causality (GC) analysis has emerged as a valuable technique for uncovering these causal relationships between system variables, enabling more accurate root cause analysis. The method’s computational efficiency and interpretable results have facilitated successful applications across diverse domains, including process industries [18] and energy systems [19,20].

Bhat et al. [21] demonstrated GC’s effectiveness in identifying dependency changes between sensor data streams, showing superior noise and latency tolerance compared to Pearson correlation. However, their analysis revealed the method’s propensity for false positives, suggesting it functions most effectively as a preliminary screening tool that requires subsequent validation through structural modeling.

Addressing computational efficiency concerns, Qiu et al. [22] enhanced traditional GC analysis through innovative stochastic and parallel processing optimizations, significantly improving both the detection accuracy and computational efficiency for multivariate time series data.

The vehicle sector has particularly benefited from GC applications. Kordes et al. [23] successfully implemented GC modeling for in-vehicle networks (IVNs), analyzing CAN bus signals between electronic control units (ECUs). Their approach proved effective in both simulated environments and real-world data scenarios, demonstrating particular promise for detecting mechanical wear-related faults that impact vehicle safety systems.

Building on these developments, researchers have identified synergistic benefits when combining GC with other analytical methods. As noted in [16], integrating Principal Component Analysis (PCA) with GC algorithms creates a more robust process monitoring framework. This hybrid approach leverages PCA’s dimensionality reduction capabilities while maintaining GC’s causal relationship detection strengths.

Collectively, these advancements in causal analysis techniques continue to enhance our ability to detect and diagnose faults in increasingly complex industrial systems, particularly where variables exhibit strong interdependencies. The ongoing refinement of these methods promises to deliver even more reliable predictive maintenance solutions across various engineering domains.

3.3.2. Machine Learning Algorithms

(1) Traditional Machine Learning Algorithms

Traditional machine learning algorithms have found widespread application in predictive maintenance domains. The most widely utilized traditional algorithms in predictive maintenance include the followin:

• Linear Regression (LR)
• Gaussian Process Regression (GPR)
• Artificial Neural Networks (ANN)
• Decision Trees (DT)
• Support Vector Machines (SVM)
• k-Nearest Neighbors (k-NN).

Linear Regression

Linear regression models establish linear relationships between input features and target variables, providing interpretable solutions for predictive maintenance tasks. Dehning et al. [24] employed multiple linear regression to analyze and quantify the factors influencing energy intensity in vehicle manufacturing plants. Their model enables the strategic planning and forecasting of future energy requirements, thereby providing valuable insights for automotive manufacturers and related stakeholders.

In mechanical component analysis, Kong et al. [25] developed multiple linear regression (MLR) models to estimate the fatigue life of vehicle coil springs by analyzing vertical vibrations and the natural frequencies of suspension systems. In their formulation, the dependent variable is the spring’s fatigue life $f\left(X_i\right)$, and the independent variables include weighted acceleration (which indicates vertical vibrations) and the natural frequencies of the suspension system.

A case study [26] implemented a predictive maintenance (PdM) strategy for electromechanical systems. Sensor data were collected from lithium-ion batteries (voltage, current, temperature) and BLDC motors (vibration, current, voltage), with each sample labeled with its remaining useful life (RUL) or a fault type (e.g., healthy, cracked propeller, shaft eccentricity, ESC failure). Following data preprocessing, regression models were employed to predict battery RUL, and classification models were utilized to identify motor faults. Model selection was guided by performance metrics, including the accuracy and F1-score. The deployed system facilitated real-time monitoring, which enabled early fault detection and generated maintenance alerts, thereby enhancing the overall system reliability and lifespan.

The study in [27] addresses fault prediction using both regression and classification techniques. The authors evaluate two machine learning approaches: an autoregressive model for forecasting vehicle failure rates based on historical data and a method that aggregates failure predictions for individual vehicles based on their usage patterns.

Gaussian Process Regression

Gaussian Process Regression (GPR) is a flexible nonparametric Bayesian method widely applied to regression and classification tasks. It is an effective tool for constructing predictive distributions by incorporating prior knowledge through carefully chosen kernel functions. For example, Aye et al. [28] developed a GPR model to estimate the remaining useful life (RUL) of slow-speed bearings, demonstrating enhanced accuracy and reduced prediction errors.

The authors of [29] employ Gaussian Process Regression (GPR) to model the capacity regeneration phenomenon in lithium-ion battery degradation, with a focus on vehicle battery health management. In this context, GPR leverages inputs such as the battery’s initial state of charge (SOC), rest time, and current state of health (SOH) to forecast performance fluctuations caused by capacity regeneration during charge–discharge cycles. These predictions enable automakers and maintenance systems to proactively schedule battery replacements or servicing, thereby preventing unexpected failures and enhancing the reliability and safety of electric vehicles.

However, conventional methods, including linear regression and GPR, may occasionally produce unreliable predictions. Tosun et al. [30] compared linear regression (LR) and artificial neural networks (ANNs) for predicting the performance and emission characteristics of a diesel engine fueled with alcohol–biodiesel blends. Their results indicated that while LR struggled to accurately predict key parameters, the ANN model achieved significantly superior performance. This limitation of traditional regression methods motivates the exploration of more sophisticated neural network approaches.

Artificial Neural Network

An artificial neural network (ANN) comprises numerous interconnected processing units, commonly termed artificial neurons. These units emulate the behavior of biological neurons, hence the name “neural network.” ANNs are organized in layers, with input units receiving data, which the network then processes to generate an output through hidden layers.

Drawing inspiration from human cognitive processes, ANNs utilize a training algorithm called backpropagation to progressively enhance their performance. During the training phase, the network learns to identify underlying patterns within the data. This supervised learning process involves comparing the network’s output to the expected or target output. The error between these values is minimized via backpropagation, an algorithm that adjusts the interconnection weights by propagating backward from the output layer to the input layer. This iterative process continues until the error converges to a minimum.

The authors of [31] employed artificial neural networks (ANNs) to optimize the side-structure design of a battery enclosure for electric vehicles (EVs). By analyzing the relationship between the geometric parameters of a honeycomb structure and its performance metrics and by integrating the NSGA-II algorithm, their methodology identifies an optimal design solution. The final design achieved a 23.9% reduction in weight, a 38.6% reduction in cost, and a 3% improvement in crashworthiness. This approach not only enhances the design efficiency of battery protection systems but also offers a methodology that can be extended to other automotive engineering applications, including crash safety design and chassis lightweighting. Consequently, it effectively reduces development costs associated with traditional trial-and-error methods, presenting an intelligent approach to structural optimization in vehicles.

In [32], the SFTI-LVAE technical framework is employed for health prediction and maintenance alerts in diesel engine lubrication systems. Lubricant parameters—including viscosity, temperature, and wear particle concentration—are collected from multiple sensors to construct a time-series model. Deep learning techniques are then utilized to predict and evaluate the system’s health status. This approach facilitates early fault detection (e.g., lubricant failure) and continuous degradation monitoring, thereby supporting predictive maintenance strategies for industrial equipment.

Support Vector Machine

Support Vector Machines (SVMs) constitute a significant category of supervised learning algorithms. SVMs operate by mapping input data into a high-dimensional feature space and identifying the optimal hyperplane that separates the data into distinct classes. The primary objective of an SVM is to maximize the margin, defined as the distance between the closest data points from each class to the hyperplane. The optimal hyperplane is situated at the midpoint of this margin. The data points closest to the decision boundary, known as the separating hyperplane, are termed support vectors because they define the margin. These points are equidistant from the optimal hyperplane, thereby ensuring maximal separation between the classes.

Support Vector Machines (SVMs) are extensively applied in maintenance applications, particularly for fault classification tasks. For example, Jeong et al. [33] developed a fault detection algorithm to identify sensor malfunctions in vehicle suspension systems, employing an SVM for classification. This machine learning approach significantly reduces the complexity involved in designing fault diagnosis systems while maintaining high accuracy. A limitation of this method, however, is that its performance is contingent on the composition of the training dataset. Their study utilized simulation-generated data; consequently, real-world implementation would require the collection of experimental data from actual vehicles operating under diverse conditions.

Similarly, Biddle et al. [34] utilized an SVM to diagnose faults in autonomous vehicle sensor systems. They introduced a predictive algorithm capable of detecting sensor degradation and estimating the time to failure. Their experimental results demonstrated the promise of this approach, achieving a prediction accuracy of 75.35% with a straightforward implementation.

In heavy vehicle applications, Raveendran et al. [35] developed a Gaussian kernel SVM (G-SVM) for air brake system fault diagnosis in Heavy Commercial Road Vehicles (HCRVs) using wheel speed data. Their approach achieved 96.54% classification accuracy with high precision (94.75%) and recall (99.15%) across 1937 test cases. The method leverages existing ABS sensor data without requiring additional hardware, making it particularly suitable for predictive maintenance applications. The standard deviation of prediction accuracy was only 1.57%, demonstrating the model’s robustness across varying operating conditions.

The authors of [36] combined a Support Vector Machine (SVM) with the Recursive Feature Elimination (RFE) algorithm (SVM-RFE) to optimize feature selection for predicting the State of Health (SOH) of lithium-ion batteries in electric vehicles (EVs). This solution is particularly suitable for Battery Management Systems (BMSs) in EVs, as it dynamically identifies key battery aging characteristics—such as the capacity decay rate and internal resistance changes—to provide data-driven support for predictive maintenance strategies.

The research presented in [37] applied ensemble learning, utilizing a Relevance Vector Machine (RVM) as a weak learner, to predict equipment health trends under uncertainty. This model effectively converted point estimates into continuous probabilistic forecasts, enhancing the robustness of the predictions.

k-Nearest Neighbors

The k-nearest neighbors (k-NN) algorithm represents another machine learning technique, primarily employed for pattern recognition and fault classification in predictive maintenance frameworks. For a given test instance, the algorithm identifies the k closest training instances in the feature space to facilitate classification or regression.

A case study presented in [38] demonstrates the implementation of a cyber–physical system for electric motor maintenance. This system continuously monitors vibration patterns to assess severity levels, thereby enabling predictive maintenance. For vibration classification, the researchers employed the k-NN algorithm. In a related work, Vasavi et al. [39] developed a real-time vehicle health monitoring system leveraging edge computing. Their system processes data from both onboard and external sensors. The results indicate that a hybrid approach, which combines an artificial neural network (ANN) with the k-NN algorithm, achieves superior accuracy compared to using either method in isolation.

The work in [40] proposes a fault prediction method for medium-voltage switchgears. K-type thermocouples were installed at critical locations within the switchgear to continuously collect temperature data, and the K-means++ clustering algorithm was employed for automatic data classification. This approach, which leverages thermocouple sensors and unsupervised learning, facilitates real-time monitoring of the equipment’s health status. Consequently, this method provides a cost-effective and scalable predictive maintenance solution for power equipment, making it particularly suitable for resource-constrained industrial environments.

Decision Tree

Among the various machine learning techniques utilized for fault classification, Decision Tree (DT) algorithms represent a versatile supervised learning approach applicable to both discrete and continuous variable prediction, tasks commonly referred to as classification and regression, respectively. Structurally, DTs operate as hierarchical models wherein internal nodes represent dataset features, branches encode decision rules, and leaf nodes denote final classifications. This architecture comprises two primary node types: (1) decision nodes, which evaluate features and branch into subsequent subtrees, and (2) terminal leaf nodes, which yield final classification outcomes without further subdivisions [41]. The model originates from a root node and progressively branches into a multi-level decision structure through the iterative application of feature-based rules.

Prominent DT construction algorithms include the following:

• Classification and Regression Trees (CART) [42]
• Iterative Dichotomiser 3 (ID3) [43]
• C4.5 (an extension of ID3) [44]
• For comprehensive algorithmic comparisons, readers are referred to [45,46].

A representative implementation is presented in [47], where the C4.5 algorithm was applied to classify five distinct fault modes in axle box bearings, thereby demonstrating the efficacy of DTs in translating complex vibration patterns into actionable diagnostic insights. This approach facilitates the systematic categorization of bearing defects through rule-based feature interrogation, highlighting the interpretability advantage of DTs in maintenance decision-support systems.

Advancing beyond single decision trees, Arun Balaji and Sugumaran [48] employed a random forest (RF) classifier for suspension system fault detection, achieving accuracies of 95.88%, 94.88%, and 92.01% for no-load, half-load, and full-load conditions, respectively. Their approach utilized vibration signals and statistical feature extraction with J48 decision tree algorithm for feature selection. The RF classifier outperformed other machine learning algorithms including SVM, KNN, and Naive Bayes, demonstrating the effectiveness of ensemble tree-based methods for complex automotive system diagnostics. The majority voting principle of random forests effectively eliminated overfitting problems and reduced bias error, making it suitable for real-world suspension system monitoring.

In a comprehensive survey, Shah [49] systematically explored machine learning algorithms for predictive maintenance in autonomous vehicles, emphasizing the critical importance of system reliability and safety. The study evaluated regression techniques, classification methods, ensemble techniques, clustering approaches, and deep learning models for system maintenance assessment. The experimental results demonstrated that predictive maintenance significantly improves system design and mitigates risk threats in autonomous vehicles. The research highlighted the particular importance of battery health monitoring, with stacked autoencoders and random forest regressors showing promising results for State of Health (SOH) prediction.

To facilitate a systematic comparison of the six aforementioned algorithms,

Table 1. Comparison of machine learning algorithms.

Algorithm	Representative Paper	Performance Metrics	Advantages	Disadvantages
Linear Regression	Kong et al. [25]	RMSE: 0.12 (fatigue life prediction)	Computationally efficient, interpretable, suitable for linear relationships	Cannot capture nonlinearity, relies on statistical assumptions
Gaussian Process Regression	Aye et al. [28]	RMSE: 0.08, R2: 0.94 (bearing RUL prediction)	Provides confidence intervals, no predefined function required	High computational cost, sensitive to kernel choice
Artificial Neural Network	Chen et al. [32]	Accuracy: 92.3%, F1-score: 0.91 (lubrication system health prediction)	Automatic feature learning, handles complex patterns	Black-box model, high training cost, hyperparameter dependent
Support Vector Machine	Raveendran et al. [35]	Accuracy: 96.54%, precision: 94.75%, recall: 99.15% (air brake system diagnosis)	Good generalization, works well with high-dimensional features	Memory-intensive, sensitive to kernel selection
k-Nearest Neighbors	Rubio et al. [38]	Accuracy: 89.7% (vibration classification)	No training phase, adapts to dynamic data updates	Computational inefficiency with large data, noise-sensitive
Decision Tree	Arun Balaji and Sugumaran [48]	Accuracy: 95.88% (no-load), 94.88% (half-load), 92.01% (full-load) (suspension system fault detection)	Highly interpretable, handles mixed data types	Prone to overfitting, unstable with small data changes

evaluates their key characteristics across four critical dimensions: model complexity, performance, advantages, and disadvantages.

The temporal evolution of predictive maintenance methodologies is depicted in Figure 5, which illustrates the growing dominance of deep learning approaches since 2015, while traditional machine learning methods continue to maintain relevance for specific applications. This trend underscores the complementary nature of different technical approaches in addressing the multifaceted challenges of vehicle maintenance prediction.

(2) Deep Learning Approaches

As illustrated in Figure 6, deep learning methods have demonstrated progressive evolution in predictive maintenance applications, with Long Short-Term Memory (LSTM) networks representing early adoptions and recent innovations encompassing vision transformers and large language models. This chronological progression reflects the continuous advancement of neural network architectures specifically tailored for automotive maintenance applications.

This section analyzes the advanced deep learning (DL) techniques commonly applied in predictive maintenance within the automotive industry. DL, as a specialized subset of machine learning, utilizes multi-layered artificial neural networks (ANNs) to model complex nonlinear relationships between input and output variables. Owing to their high parameterization, DL models generally necessitate substantial volumes of data to achieve high accuracy and ensure effective generalization. These methods have been widely adopted across various automotive domains, including autonomous driving, vehicle development, and manufacturing [50].

Several prominent DL techniques in predictive maintenance include the following:

• Long Short-Term Memory (LSTM) [51]
• Autoencoder (AE)
• Convolutional Neural Network (CNN)
• Recurrent Neural Network (RNN)
• Generative Adversarial Network (GAN)
• Large Language Model (LLM)
• Deep Belief Networks (DBN)
• Deep Reinforcement Learning (DRL)
• Vision Transformer Applications (VTA).

For comprehensive reviews of these methodologies, readers are directed to [52,53]. Notable applications of DL in vehicle predictive maintenance are detailed below.

Long Short-Term Memory (LSTM)

The study by [54] employs Long Short-Term Memory (LSTM) models to predict vehicle maintenance needs by analyzing real-time vehicle network data, such as brake temperature, engine speed, and charging cycles. The experimental results indicate that this approach outperforms conventional methods in anomaly detection and maintenance demand prediction, facilitating proactive maintenance optimization and mitigating the risk of unexpected failures. Furthermore, ref. [55] introduced an LSTM-based system to estimate the remaining fatigue life of vehicle suspension components.

External environmental conditions, including weather, traffic, and terrain, can significantly impact vehicle reliability. A recent work [56] incorporated these factors—collected via GPS and telematics—into time-between-failure (TBF) modeling. A merged-LSTM (M-LSTM) architecture was proposed to integrate real-time and historical maintenance data, which enhanced the accuracy of TBF predictions.

The study in [57] utilizes pre-trained deep learning networks, such as VGG16, ResNet-50, AlexNet, and GoogLeNet, to achieve high-precision fault classification in vehicle suspension systems by analyzing vibration signals. These signals are collected by accelerometers and converted into radar plots. The experiments demonstrated that the VGG16 model, after hyperparameter optimization (e.g., batch size, learning rate, and train–test ratio), attains a classification accuracy of up to 98.30%, effectively identifying seven common faults—including shock absorber wear and ball joint damage—in addition to normal conditions. This end-to-end approach for feature extraction and classification supersedes traditional complex signal processing techniques (e.g., FFT), thereby significantly improving the fault detection efficiency and providing a reliable solution for proactive maintenance prediction.

A novel deep symptoms-based model (Deep-SBM) proposed in [58] demonstrated superior performance in accuracy, precision, and F-score compared to existing models for efficient fault prediction. A study by [59] demonstrated its efficacy through the development of two distinct models—Long Short-Term Memory (LSTM) and Random Survival Forest (RSF)—to predict the lifespan of heavy-duty lead-acid batteries, which are critical components for engine ignition and climate control systems. The RSF model was shown to excel in estimating battery degradation trajectories, thereby underscoring the potential of RF-based methods for predictive maintenance. This methodology can be directly applied to vehicle repair demand forecasting by training RF models on multivariate sensor data, such as voltage, temperature, and diagnostic trouble codes, to predict failures in critical components like batteries and alternators. By leveraging historical failure patterns, RF facilitates proactive maintenance scheduling, which can significantly reduce vehicle downtime and optimize spare parts inventory management for service centers. This effectively bridges the gap between advanced predictive analytics and practical vehicle maintenance requirements.

The study in [60] proposes a deep learning model that integrates LSTM networks, an attention mechanism, and a gated unit for the comprehensive prediction of vehicle maintenance demands. The model maps mileage data into the same feature space as maintenance items through representation learning, captures key temporal information using LSTM and attention mechanisms, and fuses mileage with maintenance project data via a gated unit to predict future needs.

Expanding on this, ref. [61] introduces a collaborative learning model that integrates a co-occurrence matrix (for static correlation awareness) and an attention-weighted Gated Recurrent Unit (GRU) (for dynamic correlation awareness). This approach mines statistical co-occurrence relationships between maintenance items and temporal dependencies among maintenance records to achieve a holistic prediction of overall vehicle maintenance demands.

Similarly, ref. [62] proposes a collaborative multi-view time series modeling approach that amalgamates GRU, LSTM, and multiple attention mechanisms. This method models the temporal dependencies among historical maintenance items and weights information from key time points to comprehensively predict a vehicle’s overall maintenance requirements.

Advancing temporal modeling capabilities, Zhou et al. [63] proposed a novel Reinforced Memory GRU (RMGRU) network for bearing remaining useful life (RUL) prediction. Their approach combined an unsupervised health indicator based on a Gaussian Mixture Model (GMM) and Kullback–Leibler divergence with the RMGRU architecture that reuses state information from previous moments. The method demonstrated superior performance compared to GAHAU, GDAU, MMALSTM, JANET, GRU, and LSTM on IEEE PHM 2012 bearing datasets, showing strong potential for bearing RUL prediction in mechanical equipment. In the domain of electric drive system prognostics, recent research has addressed the challenge of predicting the remaining useful life (RUL) for critical components. One notable study [64] focuses on electric drive bearings in new energy vehicles, where accurate RUL prediction throughout the entire service life cycle is essential for optimizing maintenance strategies and reducing operational costs. The methodology employs a piecewise linear degradation model that effectively handles the long time spans and abrupt degradation patterns characteristic of full life bearing data.

Autoencoder (AE)

The Autoencoder (AE) is an unsupervised neural network model primarily utilized for data dimensionality reduction, feature extraction, and anomaly detection [65]. Research in this field frequently integrates AE with other machine learning techniques to improve the predictive accuracy [66]. For instance, a hybrid model integrating an AE with a Long Short-Term Memory (LSTM) network was applied to mechanical failure prediction in time-series data [67]; here, the AE was employed to extract salient features from raw sensor readings, which were subsequently analyzed by the LSTM to capture temporal dependencies. In another study, a Convolutional Autoencoder (CAE) was combined with an LSTM to predict the remaining useful life (RUL) of electric valves [68]. This hybrid architecture leveraged the CAE’s proficiency in processing spatial patterns within sensor data, achieving superior accuracy compared to traditional machine learning methods.

Pushing the boundaries of hybrid architectures, Chinta et al. [69] developed an advanced predictive maintenance system using Serial Cascaded Deep Learning (SCDL) with Henry Gas Solubility Search and Rescue Optimization (HGSSRO). Their hybrid architecture combines autoencoders for optimal feature selection, LSTM for temporal feature extraction, and DNN for final prediction. The optimized system demonstrated 4.27–9.02% higher sensitivity compared to traditional models like LS-SVM, LSTM-HBA, and ensemble methods, effectively addressing both early fault warnings and severity differentiation in automotive applications.

Addressing critical sensor reliability challenges, Min et al. [70] proposed a sophisticated fault diagnosis framework for autonomous vehicles with sensor self-diagnosis capabilities. Their approach combined a residual consistency checking algorithm based on sensor redundancy with a Denoising Shrinkage Autoencoder (DSAE) for anomaly detection. The DSAE incorporated a shrinkage block with soft thresholding for feature representation enhancement, achieving superior performance in terms of the AUC_ROC and F₁-score compared to conventional machine learning anomaly detectors. This framework effectively addresses the critical challenge of sensor reliability in autonomous driving systems, where faulty sensor readings can lead to catastrophic consequences.

The methodologies based on AEs can be effectively applied to forecast vehicle repair demands by detecting early indicators of component degradation within vehicle sensor data, such as engine vibrations, temperature fluctuations, or pressure changes. For example, an AE can be trained to identify anomalous patterns in multivariate telemetry streams, flagging potential failures—such as worn-out bearings or impending transmission issues—prior to their occurrence. When integrated with sequence models like LSTM, this approach facilitates not only the prediction of imminent failures but also the estimation of the remaining useful life of components. Consequently, this enables proactive maintenance scheduling and the optimization of spare parts inventory management, thereby reducing the unplanned downtime and enhancing the overall repair efficiency within vehicle service networks.

Convolutional Neural Network (CNN)

The Convolutional Neural Network (CNN) represents a deep learning architecture recognized for its effectiveness in processing data with grid-like structures, such as images, time series, and sensor signals. A study employing a CNN-based classifier [71] demonstrated its capability by achieving a 99.84% accuracy rate in identifying faulty signals from multisensor data, thereby facilitating real-time system health monitoring. When applied to real-time vehicle sensor data—including engine vibration, temperature, and pressure—CNN models can enable the early detection of potential faults and the prediction of maintenance requirements. This capability significantly reduces the risk of unexpected breakdowns and allows for the optimization of maintenance scheduling.

Demonstrating practical implementation, Shahid et al. [72] implemented a real-time CNN-based diagnostic method for multi-cylinder diesel engines, achieving over 99% accuracy in detecting cylinder misfires and engine load conditions. Their approach utilized crank angle degree (CAD) signals representing complete combustion cycles and employed a single convolutional layer for efficient feature extraction. The method demonstrated a low computation complexity and prediction time, making it suitable for real-time engine monitoring applications. The CNN architecture effectively combined feature extraction and classification capabilities in a single learner, outperforming traditional machine learning classifiers and previous ANN-based approaches.

Generative Adversarial Network (GAN)

The TimeGAN approach, introduced in the study by [73], offers a novel methodology for generating synthetic time-series data, which can be utilized to create high-risk driving scenarios for autonomous vehicle testing. When adapted to the predictive maintenance sector, its robust data generation capability can significantly enhance the robustness of fault prediction models. This is particularly advantageous for addressing the common challenge of limited availability of real-world fault data. Future research directions could explore the development of associative models that establish concrete relationships between the synthetically generated data and specific maintenance labels, such as component replacement records.

Large Language Models (LLMs)

The study in [74] introduces a framework that leverages Large Language Models (LLMs) to automatically optimize feature extraction parameters from vehicle sensor data, thereby eliminating the inefficiencies associated with manual tuning. A deep learning network, specifically a Multi-Head Attention Gated Recurrent Unit (MHA-GRU), subsequently analyzes these optimized features to accurately predict potential failures in critical components, such as engines and transmissions. The proposed system demonstrates superior fault detection accuracy compared to conventional methods while maintaining operational efficiency under real-time constraints. By integrating LLM-based parameter optimization with advanced neural network diagnostics, this approach facilitates more reliable and timely predictions for vehicle maintenance, which contributes to a reduction in unexpected breakdowns and associated repair costs.

Demonstrating the expanding capabilities of foundation models, Huang and Chen [75] developed an innovative EV motor fault diagnosis framework using a fine-tuned Qwen2.5-7B model. Their approach integrated current and vibration signals through time–frequency analysis and employed LoRA and QLoRA for lightweight fine-tuning. The model achieved impressive performance with 96.5% accuracy on single datasets, 91.2% cross-condition robustness, and 82.7% small-sample performance, significantly outperforming traditional CNN and LSTM models. This research demonstrates the substantial potential of large language models in industrial intelligent applications for electric vehicle motor health monitoring.

Deep Belief Networks (DBN)

A Deep Belief Network (DBN) is a probabilistic generative model constructed by stacking multiple Restricted Boltzmann Machines (RBMs) [76]. The study by [77] explores the application of DBNs in the automotive sector, specifically for the design of solar-powered on-board intelligent charging stations. In this context, a DBN is employed as an intelligent controller to manage the integration of solar panels and full-bridge converters, thereby optimizing energy management and the charging process. The capability of DBNs to extract complex nonlinear features from input data facilitates the intelligent control of the charging system. Through unsupervised pre-training via stacked RBMs, the DBN can extract salient features from datasets pertaining to solar power generation and battery storage, which enables the prediction and optimization of charging behavior.

Deep Reinforcement Learning (DRL)

Reinforcement Learning (RL) is a branch of machine learning in which an intelligent agent learns to make decisions through continuous interaction with a dynamic environment. The study in [78] employs a suite of Deep Reinforcement Learning (DRL) algorithms to train autonomous vehicles, encompassing both value-based methods—such as Deep Q-Network (DQN), Double DQN (DDQN), and Dueling DQN—and policy-based methods, including Advantage Actor–Critic (A2C) and Proximal Policy Optimization (PPO). The key technical contributions of this work include (1) the development of an end-to-end autonomous driving framework utilizing visual inputs for lane keeping and navigation; (2) the validation of Simulation-to-Reality (Sim2Real) transfer techniques in bridging the gap between simulated and real-world performance; and (3) the successful deployment and verification of the system in real-world environments. By enhancing simulation data and integrating it with real-world observations, this approach improves the predictive accuracy and robustness of the model, particularly in scenarios characterized by data scarcity or dynamic environmental changes.

The study by [79] proposes a multi-agent DRL framework comprising two independent agents: a Twin Delayed Deep Deterministic Policy Gradient (TD3) agent for continuous action space control (e.g., acceleration adjustment) and a DQN agent for discrete action space control (e.g., steering angle adjustment). When applied to autonomous adaptive cruise control, this framework facilitates adaptive following to enhance safety and comfort, optimizes energy efficiency to reduce consumption in electric vehicles, and alleviates the computational burden through its distributed design, making it suitable for onboard deployment. Furthermore, DRL excels in learning optimal strategies in complex environments without reliance on precise mathematical models. This capability can be extended to predictive maintenance by leveraging multi-source data—such as historical maintenance records, mileage, operational environment, and component status—to train agents within a DRL framework. These agents learn underlying data patterns to accurately forecast future vehicle maintenance needs, thereby enabling proactive planning and preparation. Collectively, these advancements underscore the expanding role of deep learning in enhancing predictive maintenance systems within the automotive sector.

Vision Transformer Applications (VTA)

Exploring modern computer vision architectures, Arun et al. [80] introduced a novel suspension system fault diagnosis approach using Vision Transformer (ViT) with spectrogram images derived from vibration signals. Their method achieved 98.12% accuracy in classifying various suspension faults including bushings, ball joints, struts, and tires. The ViT model’s self-attention mechanism effectively captured long-range dependencies in vibration data, outperforming traditional signal processing methods. While demonstrating strong performance, the approach requires high-end computational resources for real-time implementation, highlighting the trade-off between accuracy and practical deployment constraints.

To facilitate a systematic comparison of the aforementioned deep learning algorithms,

Table 2. Comparison of deep learning algorithms.

Algorithm	Representative Paper	Performance Metrics	Advantages	Disadvantages
Long Short-Term Memory (LSTM)	Chen et al. [60]	Accuracy: 98.30% (fault classification)	Captures temporal dependencies, handles sequential data well	Requires large datasets, computationally intensive training
Autoencoder (AE)	Min et al. [70]	AUC_ROC: 0.978, F1-score: 0.952 (sensor fault diagnosis)	Unsupervised feature learning, effective for anomaly detection	May miss subtle temporal patterns, requires careful architecture design
Convolutional Neural Network (CNN)	Shahid et al. [72]	Accuracy: >99% (engine misfire detection)	Excellent for spatial feature extraction, real-time capability	Limited temporal modeling, requires structured input data
Generative Adversarial Network (GAN)	Yoon et al. [73]	Not specified (TimeGAN for synthetic data generation)	Generates realistic synthetic data, addresses data scarcity	No direct performance metrics for maintenance prediction provided
Large Language Models (LLMs)	Huang and Chen [75]	Accuracy: 96.5% (single dataset), 91.2% (cross-condition)	Automatic feature optimization, strong generalization	High computational requirements, large model size
Deep Belief Networks (DBN)	Wang et al. [77]	Not specified (charging station optimization)	Probabilistic modeling, unsupervised pre-training	Complex training process, limited automotive maintenance applications
Deep Reinforcement Learning (DRL)	Li et al. [78]	Not specified (autonomous driving performance)	Learns optimal strategies in complex environments	Requires extensive training, simulation-to-real gap challenges
Vision Transformer (ViT)	Arun et al. [80]	Accuracy: 98.12% (suspension fault diagnosis)	Captures long-range dependencies, high accuracy	High computational resources required for real-time use

evaluates their key characteristics across four critical dimensions: model complexity, performance, advantages, and disadvantages.

Data-driven methods, particularly machine learning and deep learning, have revolutionized predictive maintenance by enabling automatic pattern discovery from large datasets without explicit programming. These approaches demonstrate remarkable performance in fault detection and remaining useful life prediction for individual components. However, they face challenges including the need for extensive labeled data, limited interpretability of complex models, and difficulties in capturing the holistic interdependencies within complete vehicle systems. The black-box nature of many data-driven models and their inability to incorporate physical constraints motivate the development of integrated frameworks that combine data-driven insights with comprehensive system representations.

3.4. Digital Twin Technology

The research landscape of predictive maintenance, as visualized in Figure 7, emphasizes key technical domains including digital twins, deep learning, and various machine learning algorithms. This visualization underscores the interconnected nature of these methodologies in addressing vehicle maintenance challenges.

Digital Twin (DT) technology creates a comprehensive virtual representation of a physical asset, system, or process, integrating real-time operational data throughout its entire lifecycle. This technology has found applications across various industries, as reviewed in [81]. In contrast to static virtual prototypes, a digital twin evolves dynamically in parallel with its physical counterpart, continuously updating through streams of performance, maintenance, and health data [82]. (For comprehensive explorations of DT technologies and their cross-industry applications, see [83].)

Within the automotive sector, digital twins have emerged as a pivotal technology for enabling predictive maintenance strategies. For instance, the study presented in [84] deployed IoT sensors on brake systems to acquire real-time data on pressure, temperature, and rotational speed. These data were used to create a high-fidelity brake digital twin within a CAD environment. The model was continuously correlated with sensor streams to run simulations that predict pad wear, triggering maintenance alerts when predefined wear thresholds are exceeded. Conversely, the authors of [85] developed integrated 0-D, 1-D, and 3-D brake models. These models are calibrated using empirical braking data and are employed to simulate wear across diverse driving conditions. The simulations facilitate the creation of a proactive maintenance schedule based on the predicted remaining useful life (RUL) of the brake pads, with the models being systematically refined as new operational data are ingested.

A multidimensional Digital Twin (DT) model [86] enhances both production quality and maintenance protocols for constant velocity (CV) joints, which are critical transmission components essential for vehicle safety, maneuverability, and ride comfort. By accurately simulating a wide range of operational conditions, this methodology significantly reduces the costs associated with physical testing while simultaneously enabling advanced fault diagnosis and design optimization.

The study in [87] proposes a Predictive Maintenance (PdM) system for vehicles, aimed at identifying potential faults prior to their occurrence. The proposed PdM framework integrates three core technologies: Digital Twin, Machine Learning (ML), and Blockchain. The Digital Twin provides continuous real-time monitoring of vehicle operating conditions. Machine Learning facilitates data-driven predictive analytics for maintenance scheduling, while Blockchain technology is utilized to ensure data integrity and traceability across both the physical and virtual twin environments.

In [88], the application of Digital Twin technology for predicting vehicle maintenance demand is primarily demonstrated through its enabling of proactive maintenance and early fault warnings via real-time data acquisition and virtual simulation. As indicated in the study, modern vehicles are equipped with a multitude of sensors (e.g., monitoring engines, suspensions, etc.), which digital twins leverage to construct high-fidelity dynamic virtual models. By integrating artificial intelligence (AI) analytics, these models can predict potential failures in advance, thereby facilitating scheduled maintenance that minimizes unexpected breakdowns and reduces associated operational costs.

The efficacy of digital twin technology in predicting vehicle maintenance demand is further exemplified through sophisticated real-time data collection and analysis pipelines. A prominent example is Tesla, Inc., which generates a unique digital twin for each vehicle sold. Thousands of onboard sensors continuously stream real-time operational data to the company’s cloud-based simulation systems. By leveraging high-performance cloud computing and generative design algorithms, a wide array of objectives can be optimized. Depending on the specific component or use case, designers can utilize AI-driven simulations to optimize parts for extended lifespan, maintained structural integrity, reduced weight, specific thermal transfer thresholds, or minimized drag—all of which are critical factors for accurate maintenance demand prediction [89].

Digital twin (DT) technology is also gaining significant traction in the development and validation of autonomous vehicles. Prior to real-world deployment, the safety and motion planning algorithms of autonomous vehicles are rigorously validated within commercial and open-source digital simulation environments. These environments can model complex systems, such as electric propulsion drives, which are common across various vehicle brands and types [90,91]. Crucially, these digital twins interface with physical test benches, enabling virtual vehicles to operate in high-fidelity simulated worlds while simultaneously feeding performance data back to physical components. This bidirectional data exchange facilitates robust design validation and generates critical insights for predictive maintenance strategies.

The study presented in [92] proposes a hybrid methodology that integrates digital twins (DTs) with Neural Networks (NNs) to predict the safety factors (SFs) of gear transmission systems. This framework consists of a physics-based simulation layer that generates gear safety factor datasets in accordance with ISO standards, and a data-driven layer that employs neural networks trained on the simulated data to construct a lightweight predictive model. This model serves as a computationally efficient alternative to traditional simulation calculations. The proposed approach is not only applicable to gear systems but is also extensible to the health monitoring of other rotating machinery components, such as bearings and turbines, thereby offering substantial reference value for the intelligent operation and maintenance of industrial equipment.

The integration of high-fidelity simulation technologies with real-time data streams enables digital twins to effectively bridge the gap between predictive analytics and physical system management. This capability is fundamentally transforming maintenance paradigms across a wide range of industries.

Digital twin technology represents a paradigm shift by creating dynamic virtual replicas that evolve alongside physical assets, enabling real-time monitoring and predictive capabilities. While digital twins offer unprecedented opportunities for holistic vehicle health management, their implementation faces substantial challenges including high computational demands, complex integration requirements, and the need for continuous data synchronization between physical and virtual domains. The limitations in scalability and the gap between component-level predictions and vehicle-level maintenance planning underscore the fundamental challenge that remains unaddressed: achieving truly integrated predictive maintenance that considers the vehicle as a complete system rather than a collection of isolated components.

4. Challenges and Future

The comprehensive analysis presented in Figure 8 synthesizes the current state of predictive maintenance research, revealing both the maturation of certain methodologies and the emergence of new challenges. This statistical overview provides crucial context for understanding the technical hurdles and future directions discussed in this section.

Predictive maintenance (PdM) systems for automotive applications encounter numerous challenges, spanning from fundamental strategic gaps to concrete technical hurdles. This is particularly evident when integrating distributed edge AI and digital twin (DT) technologies.

4.1. Transitioning to Integrated Maintenance Demand of the Entire Vehicle

A predominant focus of current research presents a fundamental strategic challenge. The field is largely characterized by a “single-component, single-model” paradigm, where studies achieve high accuracy in predicting failures for isolated subsystems—for instance, employing Support Vector Machines (SVMs) for suspension faults [46] or Long Short-Term Memory (LSTM) networks for brake wear [54] (as detailed in Section 3.2 and Section 3.3). However, a vehicle is an interconnected system where the failure of one component can induce cascading effects on others. Consequently, this siloed approach is inadequate for achieving the overarching objective of predicting the integrated maintenance demand of the entire vehicle.

To address this fundamental limitation, recent research has made significant strides toward developing integrated frameworks capable of predicting comprehensive vehicle maintenance demands. Chen et al. [60] proposed a mileage-aware model that addresses the critical challenge of integrating mileage data with maintenance project predictions. Their approach employs a specialized mileage representation method to align mileage with maintenance projects in the same vector space, combined with LSTM-attention fusion for temporal information learning and gated units for multi-source integration. This framework demonstrates that explicit consideration of mileage dynamics is essential for accurate vehicle-level predictions.

Building on this foundation, ref. [61] introduced a cooperative static and dynamic correlation-aware learning framework that captures both the inherent relationships between maintenance projects (static correlations) and their temporal evolution patterns (dynamic correlations). By leveraging co-occurrence matrices for static correlation perception and attention-weighted mechanisms for dynamic correlation awareness, their model effectively synthesizes component-level information into vehicle-level predictions.

Further advancing this integration paradigm, ref. [62] developed a collaborative multiview time series modeling approach that captures interdependencies among maintenance projects across different time periods. Their method employs multi-attention feature modeling for temporal dependency learning and dependency-aware functions to integrate information across time steps, demonstrating that comprehensive vehicle maintenance prediction requires sophisticated temporal fusion mechanisms.

For fault prediction specifically, ref. [93] proposed a collaborative multiple attention mechanisms (CoMAM) framework that simultaneously captures temporal dynamics and co-occurrence dependencies among faults. Their dual temporal correlation learning module models fault correlations from multiple perspectives, while the graph attention correlation perception module efficiently captures interdependencies between different fault types. This approach represents a significant advancement in predicting all potential vehicle faults rather than isolated subsets.

Complementing these attention-based approaches, ref. [94] introduced a multi-source data fusion network that integrates maintenance records, vehicle base information, and mileage data. Their deep fusion network utilizes co-occurrence matrices to capture maintenance project relationships and attention mechanisms to synthesize heterogeneous data sources, providing a comprehensive framework for vehicle-level maintenance prediction.

The integration of these advanced methodologies suggests a promising architectural framework for achieving true vehicle-level health management. As demonstrated by [95], the combination of graph attention mechanisms for correlation modeling and LSTM-attention networks for temporal dynamics provides a robust foundation for integrated prediction systems. These approaches collectively address the critical challenge of model fusion—developing sophisticated architectures capable of synthesizing disparate predictions from dedicated component-level models into unified and prioritized forecasts for the entire vehicle.

This paradigm shift from isolated component prediction to integrated vehicle health management is indispensable for translating academic research into tangible operational value. By adopting these integrated approaches, maintenance systems can optimize workshop scheduling, minimize overall vehicle downtime, and enhance safety through proactive identification of cascading failure risks. The pioneering work of Chen et al. [60,61,62,93,94] on comprehensive vehicle system prediction represents significant advancements in this direction, demonstrating that the fusion of temporal modeling, correlation awareness, and multi-source data integration can bridge the gap toward truly holistic Vehicle Health Management (VHM) systems.

Looking forward, the evolution of vehicle-level predictive maintenance is poised to follow several key trends that will fundamentally transform automotive maintenance paradigms. First, multi-modal fusion architectures will emerge as a dominant approach, integrating heterogeneous data sources including sensor streams, maintenance histories, environmental conditions, and driver behavior patterns through advanced attention mechanisms and cross-modal transformers. Second, explainable AI (XAI) will become increasingly critical as maintenance decisions affect vehicle safety and operational efficiency, necessitating transparent reasoning processes that can be validated by human experts. Third, federated learning frameworks will enable collaborative model training across vehicle fleets while preserving data privacy, allowing maintenance systems to learn from diverse operational conditions without centralized data aggregation. Fourth, causal inference methodologies will advance beyond correlation-based predictions to identify root causes and intervention effects, enabling more robust maintenance recommendations that account for complex system interdependencies. Fifth, continuous learning systems will adapt to evolving vehicle conditions and emerging failure modes through online model updates, addressing the challenge of concept drift in long-term vehicle operations. Finally, human–AI collaboration interfaces will bridge the gap between automated predictions and technician expertise, creating hybrid intelligence systems that leverage both data-driven insights and human domain knowledge. These trends collectively point toward a future where vehicle health management evolves from reactive maintenance scheduling to proactive health optimization, transforming vehicles from isolated assets into connected elements of intelligent transportation ecosystems.

4.2. Technical Hurdles in Implementing Integrated Systems

Predictive maintenance (PdM) systems for vehicular applications encounter significant challenges when integrating distributed edge AI and digital twin (DT) technologies. These challenges are multifaceted: Primarily, data heterogeneity and integration constitute significant obstacles, as vehicle sensors (e.g., for vibration, temperature, and oil pressure) generate data in diverse formats and at varying sampling rates, complicating the fusion of multimodal data streams. Relevant studies highlight these difficulties; for instance, ref. [96] underscores the challenges in fusing vibration data from distributed microcontrollers (MCUs), while the authors of [97] emphasize the necessity of precise spatiotemporal synchronization in multi-sensor DT systems. Second, resource constraints in edge deployment necessitate the development of highly lightweight deep learning models. Approaches such as the DDSNN method proposed in [96], which optimizes computational distribution across MCUs through neural network partitioning, offer solutions. However, persistent issues like power consumption disparities among nodes (e.g., 377 mW vs. 170 mW) continue to impact overall system efficiency. Furthermore, the substantial memory requirements for AI and DT integration may exceed the capabilities of resource-constrained vehicle edge devices.

A third critical challenge involves cybersecurity risks. Distributed edge networks in vehicles are susceptible to attacks, including spoofed sensor data or tampered maintenance commands. While research such as that in [97] proposes blockchain technology to mitigate analogous threats (e.g., G-code tampering in additive manufacturing), existing MCU communication protocols often lack robust security mechanisms. Real-time adaptability constitutes another key hurdle, as dynamic driving conditions (e.g., load variations, extreme weather) require models to rapidly adjust predictions without reliance on cloud computing. Current solutions may prove inadequate for high-speed requirements, underscoring the need for low-latency processing. Finally, ensuring scalability and generalization across diverse vehicle types (e.g., EVs vs. ICEs) and manufacturing variances remains a challenge. Addressing this necessitates the development of modular AI architectures [97] and sophisticated hierarchical computational partitioning strategies [96] to accommodate inherent hardware diversity.

To overcome these technical hurdles, several promising strategies are emerging that address the core limitations of current approaches; specific resolution strategies are summarized in

Table 3. Technical challenges and solution strategies in vehicle predictive maintenance.

Technical Challenge	Specific Manifestations	Potential Solutions
Data Heterogeneity	Multiple sensor formats and protocols; inconsistent data structures across different vehicle systems	Cross-modal fusion architectures; unified data preprocessing pipelines; standardized communication protocols
Edge Computing Resource Constraints	Limited memory and computational power in vehicle ECUs; real-time processing requirements	Lightweight deep learning models; model quantization and pruning; hardware–software co-design approaches
Cybersecurity Risks	Vulnerability to sensor data tampering; potential for malicious manipulation of maintenance predictions	Blockchain technology for data integrity; secure authentication protocols; encrypted communication channels
Real-time Adaptability	Dynamic environmental changes; varying operational conditions; evolving component degradation patterns	Online learning frameworks; adaptive model updating; reinforcement learning for continuous optimization
Scalability	Adaptation requirements across different vehicle models and configurations; fleet-wide deployment challenges	Modular AI architectures; transfer learning techniques; configurable model parameters for vehicle-specific tuning

. For addressing cybersecurity vulnerabilities, blockchain-enhanced digital twin architectures offer a robust solution by providing immutable audit trails for maintenance actions and secure data provenance verification. These distributed ledger technologies can prevent unauthorized modifications to digital twin models and ensure the integrity of predictive maintenance recommendations. For tackling computational constraints, lightweight deep learning models specifically designed for edge deployment are being developed through techniques such as neural architecture search (NAS), model pruning, and quantization. These optimized models maintain high predictive accuracy while reducing computational requirements, making them suitable for real-time inference on vehicle-grade hardware. To enhance model interpretability and trust, Explainable AI (XAI) methodologies are being integrated into predictive maintenance systems, providing transparent reasoning behind maintenance recommendations and enabling human experts to validate AI-driven decisions. Techniques such as attention mechanism visualization, feature importance analysis, and counterfactual explanations help bridge the gap between complex deep learning models and practical maintenance decision-making. For improving real-time adaptability, online learning frameworks with incremental model updates allow predictive maintenance systems to continuously adapt to changing vehicle conditions and operational patterns without requiring complete model retraining. These adaptive systems can detect concept drift and automatically adjust prediction models to maintain accuracy over the vehicle’s lifecycle. Finally, to address data heterogeneity, cross-modal fusion architectures employing transformer-based attention mechanisms can effectively integrate diverse data types—including time-series sensor data, maintenance records, and environmental context—into unified representations that capture the complex interdependencies within vehicle systems. These strategies collectively represent a comprehensive approach to overcoming the technical barriers in implementing robust and scalable vehicle health management systems.

5. Conclusions

The domain of predictive maintenance (PdM) in the automotive industry has experienced substantial evolution, transitioning from traditional reactive and preventive approaches to sophisticated data-driven methodologies. As evidenced by the comprehensive analysis in Figure 1, Figure 4 and Figure 5, recent advancements—spanning physics-based models, knowledge-based systems, statistical and machine learning techniques, and digital twin technology—have demonstrated substantial improvements in maintenance efficiency, cost reduction, and overall system reliability.

The temporal development of these methodologies, illustrated in Figure 6, reveals a clear progression toward increasingly sophisticated approaches, with deep learning architectures now dominating the research landscape. However, as highlighted in Figure 8, significant challenges remain in integrating these component-level predictions into holistic vehicle health management systems.

Physics-based models facilitate precise degradation analysis but encounter challenges related to scalability and implementation costs. Knowledge-based approaches, particularly when integrated with machine learning, offer interpretable and adaptable solutions for fault diagnosis. Data-driven methods, encompassing statistical models, traditional machine learning algorithms (e.g., SVM, ANN, decision trees), and deep learning architectures (e.g., LSTM, CNN, GANs), have proven highly effective in predicting component failures and optimizing maintenance schedules. Digital twin technology further augments these predictive capabilities by creating a dynamic link between real-time operational data and high-fidelity virtual simulations, thereby enabling proactive and informed decision-making.

Notwithstanding these advancements, several research challenges persist, as detailed in the previous section. Nevertheless, predictive maintenance is poised to revolutionize vehicle servicing paradigms, promising substantial cost savings, extended asset lifespans, and enhanced safety. The continued maturation and integration of artificial intelligence (AI) and Internet of Things (IoT) technologies are anticipated to further refine the accuracy and reliability of PdM systems, thereby solidifying predictive maintenance as an indispensable strategy for the future of the automotive industry.

This survey has reviewed the state-of-the-art predictive maintenance methodologies for vehicles, highlighting the transition from component-level to vehicle-level health management. While significant progress has been made in data-driven and digital twin approaches, several challenges remain, including data heterogeneity, model integration, and real-time adaptability. A limitation of this survey is its focus on academic literature, which may not fully capture industrial practices. Future work should focus on developing unified frameworks that combine component-level models into holistic vehicle health management systems, with an emphasis on explainability and edge deployment. The realization of such systems will substantially enhance operational efficiency, safety, and sustainability in the automotive industry.