Artificial intelligence for prognostics and health management in off-highway vehicles: a systematic review of methods, data challenges, and deployment considerations

Abstract

Off-highway machines, agricultural harvesters, construction excavators, and mining haul trucks operate under extreme load variability, harsh unstructured environments, and constrained sensor instrumentation, creating prognostic conditions fundamentally different from on-road vehicles. While AI-enabled predictive maintenance has matured for passenger vehicles and well-instrumented industrial assets, and off-highway telematics adoption is expanding rapidly, its translation to these software-defined field machines remains insufficiently addressed. This systematic review synthesizes AI-driven prognostic methods, data challenges, and deployment considerations specific to off-highway operation. Following a PRISMA 2020 protocol, the 2014–2025 literature is screened across seven databases, with the 51 studies retained for synthesis additionally quantified by method family, equipment sector, and publication year to expose the relative scarcity of off-highway-specific evidence, and a wide range of methodologies is synthesized, from foundational supervised learning (SVMs, Random Forests) and advanced deep learning (CNNs, LSTMs for RUL prediction) to unsupervised (Autoencoders), ensemble, and transfer-learning techniques. The review contrasts the primary prognostic frameworks—data-driven, physics-based, and hybrid—and the role of knowledge-based expert systems in delivering interpretable alerts. A significant focus is placed on the data pipeline, including sensor selection strategies, data quality, feature engineering, severe class imbalance, and labeling complexity. Implementation hurdles such as operating-condition variability, model validation, the computational constraints of edge devices, and Explainable AI (XAI) are further examined, with a critical analysis of where each method degrades under field variability. Finally, emerging directions are explored, including Digital Twins and Edge Computing, closing with reformulated, off-highway-specific research gaps for real-world deployment.

1 Introduction

PHM (prognostics and health management) using AI is a growing strategy for off-highway fleets due to technological advances such as telematics, edge analytics and maintenance planning being integrated into the way the equipment operates. Instead of using wide-ranging industry averages for future predictions as previously used in automotive applications will now use the telematics market for off-highway equipment as the best indicator of future growth: this market was estimated to be USD 5.94 billion at 2024 and is forecasted to reach USD 18.41 billion by 2034 with key growth areas of predictive maintenance and optimization of assets’ usage (Emergen Research, 2025).

Although there has been significant growth in the number of studies examining AI-based prognostics, there still exists a notable gap in literature. Many of the studies focus on examining the performance of individual components, subsystems (e.g., brakes, battery, etc.), or vehicles being operated in relatively predictable environments (e.g., passenger cars). Few studies have comprehensively addressed the complexities associated with off-highway field machines (e.g., agricultural combines, excavators, mining trucks, etc.) that are highly engineered. These machines often operate under great variability in loads and in harsh and unpredictable environments, resulting in data and modeling challenges that differ considerably from their road-based counterparts.

The present review addresses the various methodologies within AI and Machine Learning (ML) that are used today for prognostics. First, it describes the foundations of supervised learning such as Support Vector Machines (SVMs) and Random Forest, which have been shown to be highly accurate for classifying faults. Next, it describes the influence of deep learning, especially with respect to Convolutional Neural Networks (CNNs) to analyze multi-dimensional sensor signals and Long Short-Term Memory (LSTM) networks being used as the primary method for accurately predicting Remaining Useful Life (RUL) (Zhao et al., 2019; Khan and Yairi, 2018). Additionally, this review describes unsupervised methods such as Autoencoders that are essential for detecting anomalies when there are few labels of failure data (Bai et al., 2022; Zhao et al., 2019).

This paper investigates multiple practical and functional challenges that must be addressed in order to successfully deploy an algorithm in a real-world environment. It includes extensive analysis of the entire data pipeline beginning with data collection strategies, informed through Failure Mode and Effects Analysis (FMEA); persistent issues traditionally associated with data such as data quality and sensor noise; and the essential need for advanced feature engineering. Additionally, the paper examines several potential solutions for extreme class imbalance (a ubiquitous occurrence where the fault data is limited) and offers solutions to high-impact challenges associated with changing operating conditions (i.e., domain shift), and the complexities associated with labeling data (Theissler et al., 2021; Ward et al., 2024).

The paper aims to provide a comprehensive framework that links the limitations of AI with how they may be implemented in real life. The review will also encompass more than just isolated case studies and will thus serve to provide a thorough resource for researchers and practitioners alike. Additionally, the review will include an examination of model validation, as well the constraints of deploying (and embedding) real-time in-field devices, along with the necessary trade-off between having an accurate model and the need for interpretability (Explainable AI or XAI) (Theissler et al., 2022; Chen et al., 2023). Finally, the review looks to the future, discussing the transformative potential of advanced technologies like Digital Twins and Edge Computing architectures, and multimodal data fusion.

The novelty of this review is therefore methodological and analytical rather than purely topical: it is, to the authors’ knowledge, the first PRISMA-2020-compliant systematic synthesis to (i) quantify the off-highway AI-PHM literature by method family, equipment sector, and publication year (Section 2.6; Table 3); (ii) explicitly map each AI/ML family to characteristic failure modes under field variability rather than reporting laboratory accuracy alone (Sections 5.6–5.7; Table 6); and (iii) reconcile method selection with the edge-compute, bandwidth, and connectivity constraints distinctive to remote field operation (Sections 8.6, 9). This distinguishes the present work from prior automotive- and industrial-oriented surveys (Ucar et al., 2024; Li et al., 2024; Khan et al., 2025), which catalog methods without systematically linking them to off-highway deployment conditions (Table 4).

A direct comparison with two recent reviews is given in Table 4.

Figure 1 illustrates how technological advancements in field machines have changed throughout history from simple engine control units all the way to fully automated systems driven by artificial intelligence, while associating this change with the SAE JA6268 capability level. These Capability Levels provide a standardized definition and description of IVHM (SAE International, 2018). Figure 1 further establishes the relationship between the increasing levels of complexity and the requirement for sophisticated predictive health management (PHM) (Gong et al., 2022).

FIGURE 1

SAE International, 2018). The standard specifies six Integrated Vehicle Health Management (IVHM) capability levels spanning from “no intelligence” to “self-adaptive,” providing the standardized benchmark used here to track the progression of diagnostic and prognostic sophistication in off-highway vehicles.

2 Review methodology

This review follows the PRISMA 2020 reporting framework (Page et al., 2021). This section documents the databases and search strings, target window and justification, inclusion/exclusion criteria, and screening workflow used to make the review reproducible.

2.1 Search strategy and data sources

The following databases were searched: IEEE Xplore, ScienceDirect (Elsevier), SpringerLink, Scopus, Web of Science, MDPI, and the ACM Digital Library. Google Scholar was used for forward/backward citation chaining, and SAE/standards sources were used for standards and deployment context. Search strings combined PHM terms, AI/ML terms, and off-highway or vehicle-equipment terms with Boolean operators. The review window was 2014–2025, with pre-2014 works retained only when they were foundational standards or methods still used in PHM.

A representative full search string was: (“prognostics and health management” OR “predictive maintenance” OR “remaining useful life”) AND (“deep learning” OR “machine learning” OR “neural network”) AND (“off highway” OR “automotive” OR “heavy equipment” OR “agricultural machinery” OR “construction equipment”). Cross-cutting enabler terms (“digital twin,” “edge computing,” “sensor fusion,” “class imbalance,” “FMEA”) were run as supplementary queries to capture implementation and future-directions material.

2.2 Target years and justification

The main time frame of interest for this review is 2014–2025. The adoption of deep learning within PHM increased dramatically around 2014, primarily due to the development of state-of-the-art CNN and LSTM architecture for predicting RUL and for diagnosing faults. As a result, the methodological and current literature contained within this time frame will form the basis of the contributions presented here. A handful of works from before 2014 were kept as exceptions because of their foundational nature (and have been cited) so that they can provide the theoretical foundation for methods that are still considered canonical today; these are intentional citations and are not the result of some form of oversight.

2.3 Inclusion and exclusion criteria

Inclusion summary: Peer-reviewed journal articles, conference proceedings and authoritative review articles, published in English, on substantive topics involving AI/ML/DL methodologies for PHM, predictive maintenance/fault diagnosis/RUL estimation, related to off-highway, automotive or transferable industrial/rotating machinery applications, and making either a methodological or an applied contribution. Exclusion summary: Works published prior to 2014 (with the exception of a few seminal foundation works noted above), writing not in English, duplicates across databases, any material not peer-reviewed (with a very small number of reputable preprint and standard exceptions), studies that only cover sensors and mechanical designs without providing any PHM analytics, and papers in off-topic domains that do not offer relevant or transferable value to vehicular PHM, as well as any abstract-only entry, poster, or piece that does not provide access to the entire work and any venue that does not have methodological rigor and are considered predatory and do not warrant consideration.

2.4 Screening workflow (PRISMA 2020)

To classify the information retrieved, a four-phase process was used: (1) records were identified from multiple databases or sources; (2) duplicate records were excluded from consideration; (3) titles/abstracts were screened based on inclusion/exclusion criteria; and finally (4) remaining records were assessed for full-text eligibility based on documented reasons for exclusion. In total, there were 612 bibliographic records identified, 38 additional records identified through citation chaining or industry/standards sources, 126 duplicates removed, 524 screened using titles/abstracts, 418 excluded, 106 assessed for full-text eligibility, and 51 remaining for qualitative synthesis. Of the excluded full-text records, most were excluded due to focus unrelated to PHM, methodological inadequacy, lack of applicability to vehicles/off-highway PHM, or inaccessible full-texts. Title/abstract and full-text screens were done by two separate reviewers independently; disagreements among reviewers were resolved through discussion and/or reviewing with a third reviewer. No formal inter-rater reliability statistic was calculated as screening used consensus method after duplicate removal. The corresponding PRISMA flow diagram is shown in Figure 2, and the counts are summarized in Table 1.

FIGURE 2

2.5 Quality appraisal and risk-of-bias assessment

Because the studies retained for synthesis are predominantly data-driven prediction models (fault classifiers and remaining-useful-life estimators), risk of bias was appraised using a framework adapted from PROBAST (Prediction model Risk of Bias Assessment Tool, Wolff et al., 2019). PROBAST’s four standard domains data source, predictors/features, outcome (label) definition, and analysis/validation were retained, and a fifth domain, field-representativeness of validation, was added to reflect the central concern of this review: whether a model was validated under conditions representative of off-highway field operation rather than bench, benchmark, or single-condition settings. Each study was rated Low concern, some concern, or High concern within every domain, and an overall judgment was assigned, with any High-concern domain capable of driving the overall rating. Two reviewers applied the framework independently to the applied evidence set characterized in Table 6; disagreements were resolved by discussion, consistent with the consensus procedure used during screening (Section 2.4). Appraisal outcomes were not used to exclude studies but to weight the strength of their evidence in the synthesis: findings from studies rated High concern chiefly on field-representativeness are treated as directional and as upper-bound performance estimates rather than as field-representative results, a position carried through the failure-mode analysis (Sections 5.6, 5.7) and the research gaps (Section 11). Consistent with PRISMA 2020 Items 11 and 15, this provides both an individual-study risk-of-bias assessment and a basis for the confidence placed in the overall body of evidence.

Across the appraised set, no study reached an overall Low-concern rating (Table 2). Three studies were rated Some concern (Shafi et al., 2018; Pavlopoulos et al., 2024; Said and Nicoletti, 2015) and the remaining six were rated High concern. The field-representativeness domain was the dominant driver: eight of nine studies were validated under bench, benchmark, controlled-fault, or offline conditions, with only the field-oriented construction-telematics study (Said and Nicoletti, 2015) and, partially, the real-service-record nowcasting study (Pavlopoulos et al., 2024) drawing on operational field data. This concentration of risk in a single domain is itself a finding: it indicates that the off-highway AI-PHM evidence base is methodologically sound in its analytical execution but weakly validated against true field variability, which is precisely the standardized field-validation gap formulated in Section 11 (Gap 4).

2.6 Relation to prior reviews and recent evidence

Recent comprehensive surveys reinforce both the momentum and the open challenges in this field.

Ucar and others have reviewed advanced predictive maintenance with the use of Artificial Intelligence, including explanations on trust, explainability and new technologies such as Digital Twins and IIOT (Ucar et al., 2024). Li et al. conducted review of deep-learning-based architectures for predictive maintenance in engineering applications, with an emphasis on data-driven degradation modeling and Remaining Useful Life (RUL) estimation (Li et al., 2024). A recent review provides detailed information about the types of models, paradigms for training them, datasets, and evaluation metrics that currently exist in this field. This illustrates the explosive growth in the published literature on this topic (Khan et al., 2025). Critically, none of these three surveys differentiates method performance by field-operating condition: Ucar et al. (2024) and Khan et al. (2025) report accuracy, trust, and explainability metrics largely as reported by the source studies, without re-examining how dust, vibration, load variability, or intermittent connectivity would be expected to degrade the reported performance, and Li et al. (2024) evaluates deep-learning architectures principally against benchmark run-to-failure datasets rather than field-deployed off-highway equipment. This omission is the central gap the present review addresses through the failure-mode analysis in Sections 5.6, 5.7 and Table 6, where each method family is assessed explicitly against off-road operating conditions rather than reported benchmark accuracy alone.

Research into specific fields provides more support at the component level: Singh et al. (2023) conducted a survey of AI methods to detect faults in rotating equipment such as engines that use bearings, drivetrains, and hydraulic actuators. Off-highway applications account for most of the reasons why bearing, drivetrain and hydraulic actuator failures occur (based on the faults identified using AI). Booyse et al. (2020) provided evidence using deep digital twins to identify failures and correctly diagnose them so that future failures do not occur (discussed in Section 9). These works support the trend or methodology being discussed in this overall report, but they demonstrate that implementing all or a portion of this methodology for off-highway applications has not been well-studied.

Regarding the structured screening process outlined above, the reference set was developed from a combination of representative high impact, and methodologically rigorous works, and therefore is not the result of exhaustive enumeration. The synthesis-based type of this review covers algorithms, data pipelines, delivery methods, and advanced technologies; to complement this thematic synthesis with a quantitative view of the field, the 51 studies retained for qualitative synthesis were additionally classified by AI/ML method family, equipment sector, and publication year (Table 3). Table 4 positions the present review against two recent surveys.

This quantitative breakdown shows that deep learning and classical supervised methods together account for roughly three in five retained studies, that off-highway-specific evidence (agricultural, construction, and mining equipment) represents little over a third of the synthesis set even after deliberately weighting the search strategy toward off-highway terms, and that the literature has grown markedly since 2020—trends that motivate the off-highway-specific research gaps formulated in Section 11.

2.7 Search log and reproducibility

Exclusion criteria were kept within the screening log at the record level and included reasons such as: out of scope, no AI/PHM method, non-mobile equipment, technical detail incomplete, duplicate record or not accessible/not academic source. The final synthesis aggregated academic literature, standards and examples of industrial PHM implementations to support claims made by peer-reviewed scholarly literature about methods and validation, with grey literature (e.g., market-sizing reports such as Emergen Research, 2025) used only to contextualize deployment and market-adoption constraints and never as evidence for AI/ML method performance. PRISMA Flow and Table 1 and the comparison of applied studies have represented this boundary.

3 Machine health maintenance strategies

FIGURE 3

In addition to being reactive, preventive, and predictive, three structured frameworks are regularly used within heavy equipment fleets. They support condition-based maintenance. The first is Total Productive Maintenance (TPM) which encompasses an organizational philosophy of maximizing overall equipment effectiveness by encouraging operators to do autonomous maintenance as well as planned maintenance and continuous improvement efforts.

Secondly, Reliability-Centered Maintenance (RCM) helps determine which maintenance task would be most cost-effective for a particular failure mode based on its consequence. Usually, this is achieved through a Failure Mode Effect Analysis (FMEA), which is focused on keeping the system functioning, not just maintaining individual components.

Finally, Prescriptive Maintenance (RxM) provides further development of Predictive Maintenance (PdM) by not simply predicting failure or estimating RUL, but also identifying and automatically scheduling the best course of action with respect to the timing of that course of action based on a cost/availability/risk balance. Thus, RxM represents the ultimate target for all AI-enabled off-highway fleets, providing the means by which prognostic outputs can be translated into established fleet-wide maintenance policies.

The transition between these strategies is summarized in Table 5, which highlights the trade-off between cost, information value, and operational risk.

4 The prognostics market landscape

This section describes how AI-based PHM relates to the characteristics of commercial adoption related to off-highway fleets. The first part of Figure 4 compares the reactive, preventive, predictive, and prescriptive maintenance paradigms based on the operational characteristics that are driving the change from reactive to predictive (i.e., prognostics).

FIGURE 4

Agricultural, construction, mining, and forestry equipment do not directly reflect the amount of passenger vehicles in the prognostic market, so off-highway-specific market data is critical. The off-highway vehicle (non-road) telematics (electronics installed on vehicles that serve as an interface between machines and systems and contain various features to monitor machine performance automatically) are projected to grow from about 5.9 billion US dollars in 2024 to nearly 18.4 billion US dollars by 2034 (with a CAGR of close to 12%) based on the increasing demand for real-time location tracking, predictive maintenance, and optimizing asset utilization, according to (Emergen Research, 2025). This demand indicates that the focus of field-machine PHM development will shift away from only looking at passenger vehicles for prognostic purposes.

The application of predictive maintenance is becoming an important component of off-road equipment, including agricultural equipment, construction equipment, and mining equipment. Many of the leading equipment manufacturers have implemented AI-based predictive maintenance systems in their vehicles, using onboard sensors to monitor the condition of critical components by gathering data on component wear and failure rates, temperature, vibration, etc., and using machine learning models to predict failures and establish optimal maintenance schedules. By using these models, manufacturers can reduce the cost of downtime during peak utilization periods (Ranasinghe et al., 2020; Said and Nicoletti, 2015).

5 Machine learning approaches for prognostics

5.1 Supervised learning methods

The foundation of many prognostic systems is based on supervised learning methods that make use of datasets with labels, mapping sensor readings to fault conditions for interpretation. Support Vector Machines (SVM) provide an established supervised learning baseline for the classification of faults; although they have been used to classify faults in diesel engines and to perform fast fault detection studies, their accuracy has been shown to be limited to the dataset used. Therefore, the reported accuracy of SVMs used for the classification of faults cannot be applied across all off-highway work cycles without verification via field testing (Wang, 2005; Bai et al., 2022; Pavan Kumar et al., 2022). Random Forest classifiers are useful for handling multiple operating conditions and ranking feature importance, particularly when maintenance teams need interpretable indicators rather than black-box outputs.

The use of neural networks in predictive maintenance has been very successful, but because they are black boxes, it is often difficult to understand how they arrive at their predictions. As an alternative, decision trees are much easier to interpret compared to most neural networks, and have been integrated with expert systems for the purpose of diagnosing faults through rule-based reasoning that is verifiable by domain experts (Butler, 1996; Gearhart, 1989).

5.2 Deep learning architectures

The revolution in automatic feature extraction (removing manual engineering) means that deep learning methods, such as Convolutional Neural Networks (CNNs) have had a dramatic impact on prognostics. CNNs have been shown to effectively process large amounts of high-dimensional data collected from sensors on software-defined vehicles, and there are reports showing improvement in fault classification accuracy as well as fewer false alarm rates compared to traditional feature-based methods; however, the magnitude of these improvements will depend on the specific dataset and application, and therefore cannot be generalized without validation in the field. CNNs excel at extracting spatial features from time-series sensor data through convolution operations, with applications ranging from motor fault diagnosis to battery voltage anomaly detection (Polverino et al., 2023).

LSTM networks along with other similar algorithms have been widely employed for RUL forecasting, as they are able to capture time-dependent relationships within a set of sequential sensor measurements. Compared to the older statistical linear modeling techniques and models built on the CNN architecture without any other approaches, LSTM-based estimates of RUL are better able to leverage the sequential nature of the data when predicting remaining useful life from benchmark datasets containing PHM measurement data (Zheng et al., 2017; Khan and Yairi, 2018). The hybrid architecture elements have been combined to produce an architecture that combines CNNs and LSTMs (CNN-LSTMs) to gain the advantages of both architectures to combine local (spatial) extraction of data by the CNN to provide temporal (sequence) model for use in improving the predictive capability of health monitoring systems (battery) and how components in engines degrade (Zhao et al., 2019; Li et al., 2024).

Complex degradation trends can be accounted for using recurrent neural networks (RNNs) and their bidirectional variants when operating conditions vary. The addition of attention mechanisms to long short-term memory (LSTM) networks has greatly improved interpretability due to their ability to identify which time steps and features are most significant (Priyatno and Widiyaningtyas, 2024).

5.3 Unsupervised and semi-supervised learning

The lack of labeled failure data presents a major obstacle that can be addressed using unsupervised techniques. Anomaly detection has benefited from the development of autoencoders as effective means to reconstruct normal operating behavior and identify variations from this as anomalies. Convolutional autoencoders combine the benefits of convolutional neural networks (CNNs) with the unsupervised learning approach; therefore, they will be very useful where there is limited availability of run-to-failure data (Pavlopoulos et al., 2024; Zhao et al., 2019).

Clustering algorithms such as k-means and k-medoids segment data by operating condition, enabling condition-specific prognostic models. Semi-supervised methods bridge labeled and unlabeled data, particularly valuable where labeling requires expert knowledge (Guo et al., 2020; Ward et al., 2024).

5.4 Ensemble methods

Ensemble Learning combines a lot of base learners to become more robust and accurate. Bagging techniques like Random Forests reduce variance and deal with noise by using bootstrapped samples and averaging predictions from them. Boosting methods like Gradient Boosting Machines and XGBoost build stronger versions of weak learners one after another and provide much better RUL predictions and fault classifications (Zhao et al., 2019; Guo et al., 2020). Stacking approaches create meta-learners that combine diverse base models, often yielding gains beyond individual algorithms; these strategies are particularly effective for the heterogeneity of automotive sensor data.

5.5 Transfer learning and domain adaptation

Utilization of existing information from related sources will help solve the problem of insufficient labelled data in other target areas. These data transfer techniques include Domian Adaptation techniques to reduce distribution differences that may exist between these two sets of data via Maximum Mean Differences (MMD) use, Domain-Adversal Nueral Network’s (DANN), and correlate their members via the Correlation Alignment (CORAL) techniques. All the above have shown to be effective for Fault Diagnosis in changing operating conditions and help provide knowledge that can be shared between different vehicles (Guo et al., 2020; Khan and Yairi, 2018). Deep imbalanced domain-adaptation frameworks specifically address severe class imbalance across working conditions, a common real-world situation (Theissler et al., 2021).

5.6 Domain-specific behavior in off-highway settings

The common attributes cited above manifest very clearly, and quite often negatively, when operating in the off-highway environment. A support vector machine (SVM) trained using bench-stable signatures performs poorly with the broad-band, non-stationary vibration data generated by a mining excavator. This is because the impact of loading and the way that tracks interact with the surface of the ground causes the kernel decision boundary to become blurred, leading to the need for frequent re-tuning of the SVM. A convolutional neural network (CNN) trained on a single harvesting process will misclassify the same combine as it transitions to road transport, again because the convolutional filter is generating spectral signatures associated with a single operating point. LSTMs focused on RUL tend to perform poorly when applied across different crops, soils, or digging methods because the temporal degradation patterns they acquire are influenced by load histories that vary with each task. It is crucial to identify these specific failure modes related to the domain instead of just addressing general scalability issues when choosing and testing models for use in the field.

5.7 When methods fail under field variability

The performance of these methods degrades in characteristic ways under field conditions. CNNs perform well when the spectral or spatial signature of a fault is stable, but they fail when load, speed, temperature, terrain, or sensor mounting changes shift the input distribution away from the training regime; the practical symptom is often a higher false-alarm rate or missed incipient faults after deployment. When the sensors on the source and target machines are comparable, the underlying physics of the faults are similar, and the operating envelopes are similar transfer learning is beneficial. This may not necessarily be the case when the source-target alignment is disrupted due to drift from the off-highway operating condition, resulting in negative transfer. The RUL modeling process can be further compromised by rights censored trajectories, infrequent failures, and inaccurate logbooks of the maintenance performed on the assets. A significant portion of the assets are taken out of service before they reach failure, RUL modeling can only be developed using a very small number of failures, while work orders can document the completion of the repair, i.e., restoration of function but not the onset of failure. The use of physics-based or physics-informed models can help to mitigate some of these limitations, as both types of models can be used to limit the predictions to only those degradation paths that are physically realistic and can also provide a means of extrapolating for determining the RUL when a labeled run-to-failure data is relatively scarce. Taken together, classical supervised, deep-learning, and transfer-learning families share a common critical limitation under realistic off-highway conditions: each is validated predominantly against datasets collected under a comparatively narrow operating envelope (Table 6), so reported accuracy is best read as an upper bound rather than a field-representative estimate. This shared limitation, rather than any single algorithm’s shortcoming, is the basis for the standardized field-validation gap formulated in Section 11 (Gap 4).

This organizational chart categorizes prognostic methods into Physics-Based, Data-Driven, and Knowledge-Based approaches, setting the stage for Section 6 (Melli and Sciubba, 2020).

6 Prognostic methodologies and frameworks

Prognostic approaches can be broadly categorized into four families that mirror the taxonomy shown in Figure 5.

FIGURE 5

6.1 Data-driven prognostics

Historically driven and real-time data retrieved from sensors are used by means of statistical learning, machine learning and deep learning as methods for identifying data patterns associated with the degradation of components and predicting failures without needing a comprehensive physical model of a component (Sikorska et al., 2011; Guo et al., 2020). The pipeline (Figure 6) typically consists of (i) signal and feature selection, (ii) model development, and (iii) model selection and validation for deployment.

FIGURE 6

The Pipeline has a hierarchical structure (a), starting with a collection of raw data, preprocessed and passed through any number of modeling layers that use various algorithms (i.e., LSTM, 2D CNN) to produce desired outputs. There are multiple feedback loops throughout the Pipeline, such as a layer for Bayesian-optimization and hyperparameters tuning, and an API layer for deploying the model’s output predictions. The Sequential Workflow (b) makes it clear that development is an iterative process and includes a train/validation loop to determine the best-trained model prior to deployment. A key output of this iterative process is the Training vs. Validation RMSE curve (c), which provides essential information to diagnose model overfitting and choose a model that provides generalizable results on unseen data.

An important benefit associated with data-driven methodologies is their adaptability to changing patterns of behavior via ongoing experiential education; and also their capacity to model complex, non-linear interactions within HD (high dimensionality) data. Data-driven methodologies do, however, necessitate acquiring large amounts of representative, training data representing different fault types and operating environments (Ranasinghe et al., 2020).

6.2 Physics-based prognostics

Physics-based approaches can include knowledge about a system’s environment through mathematical representations (e.g., physical models of degradation) that represent the various ways components within the system degrade. This allows for the calculation of degradation parameters and predictions of degradation rate. Examples of techniques that combine model-based and data-based estimates to combine estimates based on observed data are Particle Filtering, Extended Kalman Filtering (EKF), and Unscented Kalman Filtering (UKF) (Liao and Koettig, 2014). The methods can be valuable when historical data is limited and provide interpretable estimates based upon engineering theory; however, an additional challenge with these techniques is the complexity of models for many systems, making producing accurate models of complex systems difficult (Cubillo et al., 2016).

6.3 Hybrid approaches

Hybrid frameworks take advantage of both data-driven and physics-based approaches to modelling and understanding degradation mechanisms. The physics-based models can describe the inherent degradation processes while the data-driven models can learn the complexity of terminal degradation as well as compensate for uncertainties associated with either of the two approaches being used (Liao and Koettig, 2014; Ademujimi and Prabhu, 2021). Hybrid frameworks have been shown to outperform either data or physics-based frameworks alone with limited training data available or where the physics is not well understood. One example is combining unsupervised machine learning with domain knowledge to develop a hybrid framework for predicting remaining useful life (RUL) (Ochella et al., 2022).

6.4 Knowledge-based and expert systems

Expert systems implement the knowledge-based branch of the taxonomy and use a combination of both AI and domain expertise to provide diagnostic and prognostic recommendations that can be understood by users. The knowledge bases of these expert systems are built from interviews with subject matter experts, maintenance records, and technical documentation represented as rules, cases and frames; technicians can understand and independently validate the transparent nature of decisions based on rule-based reasoning (Butler, 1996; Nadakatti, 2006).

Knowledge representation and reasoning. Expert systems encode domain heuristics as production (IF–THEN) rules with explicit relational operators, for example: IF engine temperature > 110 °C AND oil pressure < 20 psi THEN risk of bearing failure = HIGH.

Advanced systems incorporate fuzzy logic to handle uncertainty in sensor data (Pang et al., 2020).

Integration of operations with ML: In a hybrid pipeline, the rule engine consumes the numerical output of a learned model instead of only raw sensor data; an LSTM will produce a continuous estimated remaining useful life (RUL) or degradation index, and an autoencoder will produce an anomaly score, whereas a classifier will produce probabilities of faults. After thresholding, the expert system will combine these outputs with contextual rules based on the duty cycle, ambient conditions, service history, and FMEA criticality to generate interpretable alerts such as “hydraulic-pump cavitation risk increasing” or “bearing wear trend requires inspection within the next service window.” Butler (1996) presents a methodology for developing a framework involving incipient failure detection and predictive maintenance using Expert Systems. Ademujimi and Prabhu (2021) demonstrate that the integration of bayesian networks for the purposes of fault diagnostics also allows for learned outcomes as well as probabilistically based reasoning to be combined. Booyse et al. (2020) show that model based/digital twin reasoning can also assist with detection, diagnostics, and prognostics. This will provide both the necessary level of accuracy related to data and a transparent logic behind the actions taken by the Expert System when performing maintenance.

7 The critical role of data in prognostics

The implementation of an efficient prognostic process requires robust data collection and preparation protocols as the main foundation of success. The diagram below illustrates the iterative and sequential steps necessary to create a predictive maintenance solution beginning with collecting and preparing the required data until its monitored deployment at the defined application site. The steps within the process flow relate to the content on data handling, modeling, and validation discussed below.

7.1 Data collection strategies

Off-highway vehicles today have mixed sensors that can monitor multiple types of sensors such as engine temperatures, oil pressures, hydraulics, vibrations, fuel efficiency, duty cycle, location, and electrical measurements. There are practical issues in determining the appropriate sensors to be used, their sampling rates, and the signals that should be transmitted from remote locations given bandwidth and power limitations (Chen et al., 2023; Galego et al., 2024; Ranasinghe et al., 2020). Sensor configuration should be tied to FMEA so that each monitored signal maps to a defined component, failure mode, and maintenance action (Wang et al., 2018).

7.2 Data quality and preprocessing

Model performance is greatly affected by the quality of the data being fed into the model. Data quality issues that can impact model performance include missing values, sensor noise, outliers, duplicate data points, and measurement drift. Preprocessing techniques can be used to mitigate these data quality issues through techniques such as interpolation, noise filtering and removal, outlier detection (using Z-score and IQR), and normalization (e.g., using StandardScaler) (Fernandes et al., 2023). Temporal data typically requires sliding window techniques, time series segmentation, and resampling to create training sequences in the appropriate format (Wojenski et al., 2020; Martins et al., 2023). Figure 7 illustrates End-to-end development and deployment workflow for prognostic systems.

FIGURE 7

7.3 Feature engineering and selection

Feature engineering transforms raw sensor data into meaningful representations that expose degradation patterns more clearly than raw sensor streams alone (Fernandes et al., 2023; Zhao et al., 2019).

Statistical functions such as mean, variance, skewness, and kurtosis are referred to as time-domain characteristics. On the other hand, frequency-domain data characteristics can be obtained through Fourier transformation and wavelet decomposition (Zhao et al., 2019). Dimensionality reduction is achieved with feature selection using three different techniques: filtering through correlation with a feature set; wrapping through recursive feature elimination; or embedding using LASSO or Random Forest importances (Priyatno and Widiyaningtyas, 2024). Using PCA achieves lower dimensional representation of data and greater retention of variance; however, it gives up some level of explanatory power (Priyatno and Widiyaningtyas, 2024).

The validation and data-preparation stages are summarized in Tables 7, 8: Table 7 lists validation settings and metrics for off-highway PHM, while Table 8 lists the EDA steps used before model training.

The EDA summary in Table 8 lists the main operations used to clean, inspect, and transform sensor and maintenance data before prognostic model development. The broader AI/DL taxonomy is shown in Figure 8.

FIGURE 8

8 Challenges in off-highway prognostics

8.1 Limited sensors and data availability

One of the primary challenges is that there are not enough sensors and not enough failure information for the entire population. Field vehicles are subject to a wide range of environmental conditions, unlike a controlled laboratory environment; therefore, they are often equipped with limited instrumentation because of cost, space, or power limitations (Woll et al., 2017). Furthermore, many significant failure modes are infrequent events, which leads to a significant class imbalance between the classes (the failure modes) (Theissler et al., 2021; Ward et al., 2024). Supervised training is complicated because there is no run-to-failure data available and maintenance intervention creates right-censoring; the component is typically replaced before its natural failure time can be observed (Ward et al., 2024). Poorly recorded logs also confuse labels because work orders will often use some combination of inspection date, repair date or billing date rather than the physical time of fault-onset. Each of these conditions produces a bias in the classifiers and in the RUL regression analysis unless validation of the data is done using separation of alert time, repair time and time of confirmed failure evidence.

8.2 Data imbalance and class distribution

Class imbalance biases fault classifiers toward majority (normal) classes; fault rates may range from fractions of a percent to single digits. Addressing imbalance requires resampling (oversampling via SMOTE, under sampling), cost-sensitive learning, and metrics beyond accuracy (precision, recall, F1) (Sekar et al., 2021; Wojenski et al., 2020).

There is one major caveat concerning time series prognostics: naive SMOTE functionally generates interpolated samples from one-to-another completely disregarding time. Based on this, it destroys the sequential degradation structure on which RUL models utilize; resulting in generation of synthetically feasible but unrealistic trajectories. For sequential data, using temporally aware alternative methods that maintain trajectory continuity for sequence-level augmentation (i.e., time series-GANs), and window-level resampling methods, are preferential over naive applications and there is limited application of resampling methods to classification type framing where temporal dependence is not part of the learnable variable.

8.3 Data labeling complexity

Supervised learning requires a lot of labeled data - labeling faults requires expert knowledge and takes a long time. In a well-maintained fleet, faults are rare, so few labeled examples may be available. Labeling is a subjective process (Theissler et al., 2021; Fernandes et al., 2023).

Semi-supervised, unsupervised, and active-learning strategies partially address this.

8.4 Operating condition variability

Off-road vehicles work under highly variable loads, speeds, extremes of temperature, and types of terrain which will affect the sensor readings and the degradation rates; this complicates the problem of generalization. Therefore, a model developed for one operating regime may perform poorly when subjected to a different operating regime (due to a domain shift) (Guo et al., 2020; Fernandes et al., 2023). Addressing this requires comprehensive datasets or domain-adaptation techniques.

These factors directly impact specific components and the entire sensing chain. Hydraulic valves, cylinder seals, bearing races experience abrasive wear from dust and particulates. This same material will foul/abrade the sensor faces causing bias on vibration and pressure readings. High humidity as well as exposure to water and mud corrode electrical components (e.g., connectors), create intermittent electrical faults and jeopardize connector integrity. Extreme ambient/process temperature impacts lubricant viscosity which increases/decreases wear rates of bearings and gears. In addition, extreme temperatures cause thermal drift in MEMS accelerometers, strain gauges and pressure transducers. Continuous broadband vibration and shock can cause looseness of mounts, bracket fatigue, structural member fatigue and mechanical noise which masks early fault signatures and reduces overall sensor life expectancy. Given that all stressors will result in changes to both the degradation rate of the monitored component and the fidelity of the signal being monitored, the backhoe prognostic models must validate (test) on the full envelope of dust, moisture, temperature, vibration vs. only being tested under bench conditions.

8.5 Model validation and field deployment

Validating a prognostic model is challenging because of the temporality of the outcome predicted and a high cost associated with false positive/negative on either end of the scale. Laboratory testing provides an initial evaluation, but field validation over a deployment period of a length of time will allow for observation of real-world failures within that period and against a range of variability (Said and Nicoletti, 2015; Ranasinghe et al., 2020). Deployment introduces constraints limited compute in embedded systems, real-time processing, and connectivity limits necessitating optimization such as quantization and pruning (Chen et al., 2023; Theissler et al., 2022).

8.6 Computational and real-time constraints

Embedded systems impose strict computing, memory, and power budgets (Galego et al., 2024; Chen et al., 2023). Edge-computing architectures address these by performing critical processing onboard while offloading intensive tasks to the cloud, balancing latency and bandwidth.

8.7 Interpretability and trust

High-performance models that behave like a black box make it difficult for technicians and fleet managers to get clear explanations (Theissler et al., 2022; Chen et al., 2023). Explainable AI (XAI) helps to provide transparency to these users using methods like SHAP and LIME, as part of model governance and robust analytics (Theissler et al., 2022; Chen et al., 2023).

9 Advanced technologies and future directions

9.1 Digital twin technology

Digital twins are virtual representations of physical assets that leverage real-time sensor data to allow for ongoing analysis of physical items to help improve predictive maintenance practices. Digital twins can also be used to simulate different maintenance-related scenarios, which is useful when there are not many data points available regarding the physical repair or operational failure of an item (El Hadraoui et al., 2024; Ranasinghe et al., 2020; Booyse et al., 2020). When these digital twins simulate the degradation of a component, this provides additional support to assist in determining the remaining useful life (RUL) for the item.

Since high-bandwidth continuous synchronization is impractical for remote mining, construction and agricultural applications, an event-driven synchronization model is applied to the mobile-equipment digital twinning process. As a lightweight edge twin is running on the vehicle, it will be updating health state information from CAN/telematics, vibration, temperature, pressure, and hydraulic signals as well as buffering data until a connection is available. When a connection is established, the system will upload compressed health indicators, model residuals, and selective high-resolution snapshots to the cloud twin, which will update degradation parameters and simulate wear using alternative haul-road grades, soil loads, ambient temperatures, duty cycles and service schedules. Through this architecture, safety-critical onboard alerts are decoupled from intermittent connectivity while providing additional richer cloud-based what-if analysis (Ranasinghe et al., 2020; Booyse et al., 2020; El Hadraoui et al., 2024).

9.2 Edge computing and deployment of hybrid Expert–ML systems

As fleet management is evolving due to integrated IoT and Edge Computing, they are both interrelated and connected. The use of Edge Computing in real-time processing at or close to the vehicle reduces latency and lessening dependency upon the cloud (Galego et al., 2024; Chen et al., 2023). The architecture is designed such that processing is distributed across the onboard devices, local edge servers and the cloud, requiring the development and use of light weight models, OTA (over the air) updates and fail-safe local processing (Chen et al., 2023).

To reconcile the hybrid expert/machine learning systems in Section 6.4 with the edge constraints detailed in Section 8 requires dividing up the functions instead of deploying everything on the vehicle. Compared to larger systems for storing knowledge bases, there are lightweight, compressed inference models, and rule engines that operate on the vehicle for decision-making at the low level of until their first-level fault. Larger systems are used for updating knowledge bases, retraining the model, or executing full probabilistic inference are accomplished while using the cloud or local gateways, respectively. The use of quantization, pruning, and knowledge distillation help keep the on-board footprint limits within CPU, memory, and power budgets. Upgrades made through OTA will spread out the enhancements to the rules and weights used in the hybrid architecture and will allow them to be deployed realistically on low-capacity off-road hardware.

9.3 Multimodal and fusion approaches

Future systems will increasingly leverage multimodal data fusion, integrating vibration, acoustic, thermal, and visual data for comprehensive health assessment (Guo et al., 2020; Khan and Yairi, 2018). Combining statistical methods, ML, and physics-based models represents a systematic approach to effective PHM (Tao et al., 2025; Ochella et al., 2022).

10 Case studies and applications

10.1 Off-highway agricultural equipment

The use of AI-based predictive maintenance by OEMs has been developed by using continuous sensor monitoring as well as machine learning to accurately predict failures and determine optimal timing to schedule activities during critical times for planting or harvesting (Ranasinghe et al., 2020; Shafi et al., 2018). Real-time decision-making occurs within strict latency constraints and autonomous systems utilize edge computing to display improved productivity and efficiency (Said and Nicoletti, 2015).

Operational advantages from early maintenance alerts and planning were observed in field-based studies evaluating vehicle health; however, many of the results from deployments are either proprietary or presented without a complete validation protocol (Ranasinghe et al., 2020; Shafi et al., 2018). This lack of empirical evidence is considered a research gap, rather than independently validated quantitative evidence.

10.2 Construction and mining equipment

Another important application area for mining and construction fleets (that includes haulage equipment, articulated dump trucks, excavators, etc.) is the construction and mining construction side of the world (Said and Nicoletti, 2015; Woll et al., 2017) as they work in an organized environment but also have to operate within extreme (dust, vibration) environments, thus impacting the reliability of sensors to measure these parameters. Predictive maintenance for these fleets typically targets high-stress components like hydraulic systems and drivetrains; fleet-level prognostics allow for maintenance activities to be coordinated across the entire fleet (Polverino et al., 2023; El Hadraoui et al., 2024).

10.3 Electric and hybrid vehicles

For proactive assessment, battery health monitoring and remaining useful life (RUL) of the battery are two important predictive elements in electric vehicles (EV) (El Hadraoui et al., 2024). How battery degrades is determined by a combination of charge cycles, temperature, and the depth of discharge, necessitating the need for advanced RUL models to address this (Liao and Koettig, 2014; Zhu et al., 2019; Yan et al., 2022).

Techniques that use LSTM with autoencoders or CNN-LSTM hybrids have given good results for detecting faults on batteries and estimating their remaining useful life (RUL) (Zhao et al., 2019). In this review, the automobile and electric vehicle (EV) batteries will be considered as examples for transferability of techniques rather than the focus of analysis (see Section 11).

11 Synthesis and research gaps

While substantial progress has been made, several critical gaps remain. They are stated here as explicitly scoped problems for off-highway PHM.

Gap 1 - The limited focus on integrated, complex off-highway systems in existing literature mostly addresses separate subsystems or vehicles rather than complex integrated machines used in fields. Any automotive or electric vehicle battery example used in this review is to show how to use mature and well-documented transfer methodologies for off-highway prognostics, rather than as the subject of discussion. The contribution of this article is the multi-subsystem off-highway context and this gap in existing literature applies to the consolidated literature in this review.

Gap 2 - Absence of systematic end-to-end frameworks. Most studies present isolated use cases without systematic frameworks spanning data collection, sensor selection, model validation, and deployment.

Gap 3 - Cost–accuracy sensor selection for a defined off-highway failure mode.

The primary objective of this problem is to develop a way of determining, for only a specific platform and mode of failure, the minimum number of sensors that will provide an acceptable cost for reaching a desired level of prognostic performance. Examples of this type of problem include hydraulic-pump cavitation in a wheel-loader application, brake systems overheating in an articulated dump truck, bearing wear in a swing drive of an excavator, and planetary-gear pitting problems in final drives on combines. The literature currently focuses more on general discussion on selecting sensors without adequately describing their costs, sampling rates, sensor placements, and performance trade-offs associated with using the various sensors.

Gap 4 - Standardized field-validation methodologies. Approaches for assessing prognostic performance under real operational variability remain underdeveloped (Emergen Research, 2025).

Gap 5 - End-to-end validated frameworks for a bounded vehicle/component task.

A needed framework will be illustrated via a bounded case: RUL of truck drivetrains in regard to haul-road grades; hydraulic pump deterioration on wheel loaders during their duty cycle; and bearing wear in excavators subject to both vibration and dust/mud. Additionally, this framework will handle data scarcity/censoring/domain shift; underlying physics-based constraints; expert-rule interpretation; and, lastly, field validation jointly rather than as separate issues.

Gap 6 - Sociotechnical adoption. User acceptance, organizational change management, and integration with existing maintenance workflows receive limited attention despite their importance in deployment (Guo et al., 2020; Ochella et al., 2022).

Gap 7 - Reasons for wanting to improve benchmarking procedures in the off-highway product sector include the fact that many performance numbers of headlines are reported from vendor disclosure (the origin of many of the performance claims) and thus, there is a need for document off-highway benchmarks to aid in establishing empirical transparency and reproducibility of performance.

All the reviewed literature’s major limitations were not because of no available high-performance algorithms but were due to the instability of reported performance for that algorithm once deployed under changing business environments. For example, studies conducted with controlled datasets and generated high levels of accuracy for algorithms do not typically contain cross condition field validation, while operational studies often report operational performance without sufficient model, dataset, and failure model detail to permit replication by other researchers. For future off-highway PHM research to have maximum value, researchers developing their benchmarking methods should provide all necessary documentation (vehicle class, fault type, sensor set, operating envelope, censoring method, validation set, and deployment method) to allow their models to be replicated by subsequent researchers.

12 Limitations of the review

There are three major limitations of this study. The first limitation is that most of the validated PHM studies published for off-highway applications do not share the same validation conditions, fault definitions and levels of field exposure which prevents quantitative comparison across the various algorithms. The second limitation is that most of the fleet or program results are not reported using fully reproducible benchmark datasets and should be interpreted only as evidence of the direction of implementation and not a performance guarantee across the board. The third limitation results from an emphasis placed on mobile machinery used in agricultural, construction and mining applications while stationary industrial machinery is included only if it provides a transferable methodology or data governance challenge to illustrate. These limitations provide further evidence for a need for commonly accepted off-highway benchmark metrics, transparent data censoring criteria and common reporting standards linking a model’s metrics to decisions made regarding the operation of maintained equipment.

13 Conclusion

This review conclusively demonstrates that AI-based PHM systems employed in off-highway applications differ markedly from predictive maintenance systems utilized in automobiles and fixed-site industrial applications. This is because field machines operate in non-stationary duty cycles, are exposed to harsh environmental conditions, often have an intermittent connection to a network, and rely heavily on both sparse labeled data and incomplete or inconsistent records of maintenance and previous failures. As such, all these elements will affect both the validity of AI-based models and their usefulness when generating alerts.

Based on the literature reviewed, classical machine-learning models can still be used for fault classification with interpretability when there are well-defined features and reliable labels. In contrast, deep learning, transfer learning, digital twins and physics-aware models will show the greatest promise when they relate to realistic validation protocols. Evidence of their effectiveness comes from not only algorithmic accuracy but also from studies that correlate model outputs to components, fault modes, operating conditions, maintenance actions and deployment scenarios.

To facilitate research and publication, the urgent requirement is a reproducible evidence base from off-highway PHM. This needs to include transparent, PRISMA-like screening of studies, well-documented study references (traceable references), validated metrics of field samples, explicit treatment of infrequent events, an understanding of censoring, and reporting of boundary/cloud limitations.

There is a way to improve this: combine physics-informed design (an understanding of how a system works), use learning from uncertainty, develop rule-based expert systems, and use digital twins to update the workflow for maintenance of fleets (of agricultural, construction or mining vehicles) that can be performed based on limited amounts of actual data.

References

• 1

  AdemujimiT.PrabhuV. V. (2021). Fusion-learning of Bayesian network models for fault diagnostics. Sensors21, 7633. 10.3390/s21227633

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  BaiH.ZhanX.YanH.WenL.YanY.JiaX. (2022). Research on diesel engine fault diagnosis method based on stacked sparse autoencoder and support vector machine. Electronics11, 2249. 10.3390/electronics11142249

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  BooyseW.WilkeD. N.HeynsS. (2020). Deep digital twins for detection, diagnostics and prognostics. Mech. Syst. Signal Process.140, 106612. 10.1016/j.ymssp.2019.106612

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  ButlerK. L. (1996). “An expert system based framework for an incipient failure detection and predictive maintenance system,” in Proceedings of the International Conference on Intelligent System Application to Power Systems, 321–326. 10.1109/ISAP.1996.501092

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  ChenW.MilosevicZ.RabhiF. A.BerryA. (2023). Real-time analytics: concepts, architectures, and ML/AI considerations. IEEE Access11, 71634–71657. 10.1109/ACCESS.2023.3295694

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  CubilloA.PerinpanayagamS.Esperon-MiguezM. (2016). A review of physics-based models in prognostics: application to gears and bearings of rotating machinery. Adv. Mech. Eng.8, 1–21. 10.1177/1687814016664660

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  El HadraouiH.OuahabiN.El BaziN.LaayatiO.ZegrariM.ChebakA. (2024). Toward an intelligent diagnosis and prognostic health management system for autonomous electric vehicle powertrains: a novel distributed intelligent digital twin-based architecture. IEEE Access12, 113314–113333. 10.1109/ACCESS.2024.3441517

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  Emergen Research (2025). Off-Highway Vehicle Telematics Market Size Worth USD 18.4 Billion by 2034. Vancouver, BC: Emergen Research.

  • Google Scholar
  • View reference in article

• 9

  FernandesA. A. A.KoehlerM.KonstantinouN.PatonN. W.SakellariouR. (2023). Data preparation: a technological perspective and review. SN Comput. Sci.4, 425. 10.1007/s42979-023-01828-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  GalegoN. M. C.MartinhoD. S.DuarteN. (2024). Cloud computing for big data analytics: how cloud computing can handle processing large amounts of data and improve real-time data analytics. Procedia Comput. Sci.237, 297–304. 10.1016/j.procs.2024.05.108

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  GearhartL. M. (1989). “Managing the development and deployment of expert systems,” in National Aerospace and Electronics Conference, 1004–1010. 10.1109/NAECON.1989.40333

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  GongC.-S. A.SuC.-H. S.ChenY.-H.GuuD.-Y. (2022). How to implement automotive fault diagnosis using artificial intelligence scheme. Micromachines13, 1380. 10.3390/mi13091380

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  GuoJ.LiZ.LiM. (2020). A review on prognostics methods for engineering systems. IEEE Trans. Reliab.69, 1110–1129. 10.1109/TR.2019.2957965

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  KhanS.YairiT. (2018). A review on the application of deep learning in system health management. Mech. Syst. Signal Process.107, 241–265. 10.1016/j.ymssp.2017.11.024

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  KhanU.ChengD.SettiF.FummiF.CristaniM.CapogrossoL. (2025). A comprehensive survey on deep learning-based predictive maintenance. ACM Trans. Embed. Comput. Syst.25, 1–43. 10.1145/3732287

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  LeeJ.WuF.ZhaoW.GhaffariM.LiaoL.SiegelD. (2014). Prognostics and health management design for rotary machinery systems: reviews, methodology and applications. Mech. Syst. Signal Process.42, 314–334. 10.1016/j.ymssp.2013.06.004

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  LiZ.HeQ.LiJ. (2024). A survey of deep learning-driven architecture for predictive maintenance. Eng. Appl. Artif. Intell.133, 108285. 10.1016/j.engappai.2024.108285

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  LiaoL.KoettigF. (2014). Review of hybrid prognostics approaches for remaining useful life prediction of engineered systems, and an application to battery life prediction. IEEE Trans. Reliab.63, 191–207. 10.1109/TR.2014.2299152

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  MartinsA. S.FonsecaI.FarinhaJ. T.ReisJ.CardosoA. J. M. (2023). Online monitoring of sensor calibration status to support condition-based maintenance. Sensors23, 2402. 10.3390/s23052402

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  MelliR.SciubbaE. (2020). Diagnostics and prognostics of energy conversion processes via knowledge-based systems. Proceedings4, 1. 10.3390/WEF-06937

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  NadakattiM. M. (2006). A knowledge-based approach in maintenance engineering. Manuf. Eng.85, 24–27. 10.1049/ME:20060203

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  OchellaS.ShafieeM.DinmohammadiF. (2022). Artificial intelligence in prognostics and health management of engineering systems. Eng. Appl. Artif. Intell.108, 104552. 10.1016/j.engappai.2021.104552

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  PageM. J.McKenzieJ. E.BossuytP. M.BoutronI.HoffmannT. C.MulrowC. D.et al (2021). The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ372, n71. 10.1136/bmj.n71

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  PangN.PengJ.LiuP.FengY.ZhouL.WangL.et al (2020). A fuzzy markov model for risk and reliability prediction of engineering systems: a case study of a subsea wellhead connector. Appl. Sci.10, 6902. 10.3390/app10196902

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  Pavan KumarC. V.PandyaV.ValaK.VilasA.SantK. (2022). Support vector machine for fast fault detection and classification in modern power systems using quarter cycle data. Energy Rep.8, 92–98. 10.1016/j.egyr.2022.10.279

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  PavlopoulosJ.RomellA.CurmanJ.SteinertO.LindgrenT.BorgM.et al (2024). Automotive fault nowcasting with machine learning and natural language processing. Mach. Learn.113, 843–861. 10.1007/s10994-023-06398-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  PolverinoL.AbbateR.MancoP.PerfettoD.CaputoF.MacchiaroliR.et al (2023). Machine learning for prognostics and health management of industrial mechanical systems and equipment: a systematic literature review. Int. J. Eng. Bus. Manag.15, 18479790231186848. 10.1177/18479790231186848

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  PriyatnoA. M.WidiyaningtyasT. (2024). A systematic literature review: recursive feature elimination algorithms. JITK J. Ilmu Pengetah. Dan Teknol. Komput.9, 196–207. 10.33480/jitk.v9i2.5015

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  RanasingheK.KapoorR.GardiA.SabatiniR.WickramanayakeV.LudoviciD. (2020). Vehicular sensor network and data analytics for a health and usage management system. Sensors20, 5892. 10.3390/s20205892

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  SAE International (2018). “Design and run-time information exchange for health-ready components,” in Surface Vehicle/Aerospace Recommended Practice JA6268_201804. Warrendale, PA: SAE International. 10.4271/JA6268_201804

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  SaidH.NicolettiT. (2015). Telematics data-driven prognostics system for construction heavy equipment health monitoring and assessment. Vancouver, BC, Canada: University of British Columbia. 10.14288/1.0076365

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  SekarK.ShahS. A.AthithanA. A.MukilA. (2021). “Role of machine learning approaches in remaining useful life prediction: a review,” in Advances in Intelligent Systems and Computing, 361–370. 10.1007/978-981-16-3153-5_39

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  ShafiU.SafiA.ShahidA. R.ZiauddinS.SaleemM. Q. (2018). Vehicle remote health monitoring and prognostic maintenance system. J. Adv. Transp.2018, 8061514. 10.1155/2018/8061514

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  SikorskaJ.HodkiewiczM.MaL. (2011). Prognostic modelling options for remaining useful life estimation by industry. Mech. Syst. Signal Process.25, 1803–1836. 10.1016/j.ymssp.2010.11.018

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  SinghV.GangsarP.PorwalR.AtulkarA. (2023). Artificial intelligence application in fault diagnostics of rotating industrial machines: a state-of-the-art review. J. Intelligent Manuf.34, 931–960. 10.1007/s10845-021-01861-5

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  TaoL.LiS.LiuH.HuangQ.MaL.NingG.et al (2025). An outline of prognostics and health management large model: concepts, paradigms, and challenges. Mech. Syst. Signal Process.232, 112683. 10.1016/j.ymssp.2025.112683

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  TheisslerA.Perez-VelazquezJ.KettelgerdesM.ElgerG. (2021). Predictive maintenance enabled by machine learning: use cases and challenges in the automotive industry. Reliab. Eng. and Syst. Saf.215, 107864. 10.1016/j.ress.2021.107864

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  TheisslerA.ThomasM.BurchM.GerschnerF. (2022). ConfusionVis: comparative evaluation and selection of multi-class classifiers based on confusion matrices. Knowledge-Based Syst.247, 108651. 10.1016/j.knosys.2022.108651

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  UcarA.KarakoseM.KirimcaN. (2024). Artificial intelligence for predictive maintenance applications: key components, trustworthiness, and future trends. Appl. Sci.14, 898. 10.3390/app14020898

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  WangL. (2005). Support Vector Machines: Theory and Applications. Berlin: Springer.

  • Google Scholar
  • View reference in article

• 41

  WangZ.GaoJ.WangR.ChenK.GaoZ.ZhengW. (2018). Failure mode and effects analysis by using the house of reliability-based rough VIKOR approach. IEEE Trans. Reliab.67, 230–248. 10.1109/TR.2017.2778316

  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  WardT. J.JenabK.Ortega-MoodyJ.StaubS. (2024). Comprehensive review of machine learning techniques for condition-based maintenance. Int. J. Prognostics Health Manag.15. 10.36001/ijphm.2024.v15i2.3850

  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  WojenskiA.PozniakK. T.LinczukP.ChernyshovaM.KasprowiczG.MazonD.et al (2020). Data quality monitoring considerations for implementation in high-performance raw signal processing real-time systems with use in tokamak facilities. J. Fusion Energy39, 221–229. 10.1007/s10894-020-00243-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  WolffR. F.MoonsK. G. M.RileyR. D.WhitingP. F.WestwoodM.CollinsG. S.et al (2019). PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann. Intern. Med.170, 51–58. 10.7326/M18-1376

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 45

  WollL.SchickK.JacobsG.KramerA.NeumannS. (2017). “Reliability evaluation of drivetrains: challenges for off-highway machines,” in Reliability and Safety Engineering (London: IntechOpen). 10.5772/intechopen.70280

  • CrossRef
  • Google Scholar
  • View reference in article

• 46

  YanN.YaoY.-B.JiaZ.-D.LiuL.DaiC.-T.LiZ.-G.et al (2022). Online battery health diagnosis for electric vehicles based on DTW-XGBoost. Energy Rep.8, 121–128. 10.1016/j.egyr.2022.09.126

  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  ZhaoR.YanR.ChenZ.MaoK.WangP.GaoR. X. (2019). Deep learning and its applications to machine health monitoring. Mech. Syst. Signal Process.115, 213–237. 10.1016/j.ymssp.2018.05.050

  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  ZhengS.RistovskiK.FarahatA.GuptaC. (2017). “Long short-term memory network for remaining useful life estimation,” in 2017 IEEE International Conference on Prognostics and Health Management (ICPHM), 88–95. 10.1109/ICPHM.2017.7998311

  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  ZhuJ.ChenN.PengW. (2019). Estimation of bearing remaining useful life based on multiscale convolutional neural network. IEEE Trans. Industrial Electron.66, 3208–3216. 10.1109/TIE.2018.2844856

  • CrossRef
  • Google Scholar
  • View reference in article