Data-driven prognostics and health management for mining processing machinery: a systematic review

1 Introduction

Operations in mining processing rely on large and heavy mechanical machinery whose ability to perform their required function is critical for efficient and continuous production. The required function of a system refers to a function, a combination of functions, or the set of functions of an item that is considered necessary to fulfil a specified requirement (IEC 60050-192, 2015). In the mining industry, once the ore is transported to the processing plant, it undergoes a series of transformation stages that convert the heterogeneous properties of ore into a homogeneous, marketable product by separating valuable minerals from waste materials (Hodouin, 2011). To achieve this, raw ore undergoes a well-established processing stage (Hodouin, 2011; Sbarbaro and del Villar, 2010) as shown in Figure 1.

• Crushing and grinding: Process of reducing the large rock fragments into smaller particles for subsequent processing.
• Sizing and classification: Process of separating particles into size-based fractions using screens or hydrocyclones.
• Concentration: Process of ore enrichment by using methods such as froth flotation, gravimetric separation, magnetic and electrostatic separation aimed to keep as much as possible of the valuable minerals.
• Dewatering: Process of removing excess water through thickening, filtration or drying process to reach the level of concentration in the final product.

In a mining environment, each machine or system operates continuously under harsh, variable and load-intensive conditions. Components of equipment inevitably degrade over time and usage, which leads to reduced functional capacity and an increase in the likelihood of failure. Failure is the termination of the ability of an item to perform the required function (IEC 60050-192, 2015). System failure often results from factors such as ageing, design configuration, environment, or misuse of the system (Söderholm, 2004). Failure to deliver the required function might lead to consequences such as unplanned downtime, reduced productivity, higher maintenance costs, or exposure of personnel to hazards (Dhillon, 2008; Jardine et al., 2006). Consequently, a system often needs a certain type of support to function as intended throughout its operational life. While Prognostics and Health Management (PHM) has demonstrated significant benefits in asset-intensive industries such as aerospace, automotive, and railways, its application to mining processing machines remains comparatively limited. At the same time, growth in sensor technologies, the Industrial Internet of Things (IIoT), advances in machine learning (ML) and deep learning (DL) have created new opportunities. A recent review within the mining domain highlights that automation and Artificial Intelligence (AI) driven technologies are increasingly being adopted across mining operations (Codoceo-Contreras et al., 2024). This motivates a timely and systematic review of data-driven approaches applied to mining processing machinery.

Common maintenance strategies encompass Corrective Maintenance (CM), which restores functionality after failure occurs, and scheduled Preventive Maintenance (PM), which attempts to avoid failure by scheduling interventions at fixed time intervals. CM might result in unexpected failures and production losses, while scheduled PM can lead to premature replacements and unnecessary maintenance, thereby increasing costs (Jardine et al., 2006). These limitations have motivated a transition towards Condition-Based Maintenance (CBM) and Predictive Maintenance (PdM), where decisions are taken by monitoring the actual health of the system (Lee et al., 2006). PHM is an enabler to achieve PdM, which represents a maintenance strategy that uses real-time condition information and predictive models trained on historical data to determine the optimal time of maintenance actions. PHM is a framework that integrates sensing technologies, data analytics, and modelling techniques to detect anomalies or faults, diagnose their causes, and predict remaining useful life (RUL) (Lee et al., 2014). PHM enables timely, targeted and cost-effective maintenance actions by continuously monitoring the system, identifying degradation and predicting future performance (Helu and Weiss, 2016; Rezaeianjouybari and Shang, 2020). This transition is unfolding within a broader paradigm, in which industrial assets are instrumented as cyber-physical systems that couple physical machinery with real-world sensing, the Industrial Internet of Things, and data-driven intelligence.

One principle of PHM is to convert a set of raw data collected from the asset into one or more meaningful features and possibly to build health indicators (HI) (Atamuradov et al., 2020; Gouriveau et al., 2016). This process initially depends on feature engineering, relying on domain experts. However, with the introduction of DL, feature engineering has shifted towards feature learning, allowing models to learn informative representations for HI construction (Fink et al., 2020). The analysis of these HIs over time enables the definition of a maintenance strategy and informed decision-making (Gouriveau et al., 2016). The following are the three tasks in PHM analysis: fault detection, diagnostics, and prognostics (Gouriveau et al., 2016; Zhang et al., 2019; IEEE, 1856, 2017). Fault detection identifies deviations in performance by analyzing the data to detect anomalies. Early detection is crucial for mining machinery, where subtle changes can lead to unexpected downtime (Prabhu et al., 2025). Diagnostics aims to identify and isolate the component where an anomaly occurs. It identifies the likely cause for a given effect. Advanced diagnostic approaches include DL to classify and identify the failure component and sensor fusion that integrates heterogeneous data to enhance diagnostic certainty (Kolodenkova et al., 2019; Calikus et al., 2019). Lastly, given the current system state, prognostics aim to predict the future state of the system, focusing on predicting the RUL. Data-driven methods, such as DL models, autoencoder-based HI construction, and sequential techniques, have shown capability in modelling complex degradation patterns (Li and Zhang, 2021; Fink et al., 2020). Building on these tasks, health management refers to a decision-making process that uses estimated system health and expected future operating conditions to guide maintenance and operational actions (IEEE, 1856, 2017). Health management supports optimal maintenance and logistics decisions by considering the available resources, the operational context and the economic implications of different faults (Fink et al., 2020). It enables risk-informed decision-making and undertaking actions based on prognostics information, thereby supporting an effective maintenance scheme (Huang C. et al., 2024). In industrial settings, ML and DL models trained on condition-monitoring data have become central to fault detection, diagnosis, and RUL estimation, shifting maintenance from reactive and scheduled regimes toward proactive, data-driven decision-making (June, 2022; Azeta et al., 2026). The mining sector is increasingly receptive to ML-based decision support (Chimunhu et al., 2024), creating a foundation on which PHM frameworks can build.

Although PHM has shown progress in recent years, most developments have been concentrated in the aerospace domain, automotive systems, railway, and general rotating machinery (Jardine et al., 2006; Rezaeianjouybari and Shang, 2020; Lee et al., 2014). The mentioned domains benefit from abundant high-quality datasets, advanced sensing methods, and established domain-specific reliability standards. While mining processing machines share some characteristics with these industries, they also present a distinct combination of challenges. These include the processing of heterogeneous and abrasive raw material, cascading effects across interdependent equipment, scarcity of labelled fault data and limited sensor infrastructure (Rezaeianjouybari and Shang, 2020; Kolodenkova et al., 2019; Li and Zhang, 2021). Furthermore, mining processing involves domain-specific machines that have no direct counterpart in other industries. These characteristics motivate the need for exploring data-driven PHM approaches focused on the mining production process, where an increase in system availability contributes to operational continuity and production cost control.

To address this need, the present study conducts a review focused on PHM in the mining production process. The main contributions of this paper are.

• A comprehensive Systematic Literature Review (SLR) dedicated to PHM methods applied to machinery used in the mining production process.
• A structured synthesis of existing studies according to the three tasks of PHM, while mapping them to corresponding data types, modelling approaches, equipment categories, and health indicator construction strategies.
• Based on the observations of the reviewed studies, this paper highlights some trends that motivate a set of open questions for advancing PHM implementation in mining processing environments.

The manuscript is organized as follows: Section 2 describes the methodology; Section 3 presents the findings of the review process. Section 4 presents the bibliometric characteristics of the corpus, the observed trends, a critical analysis of the reviewed studies, and the identified research gaps. Finally, Section 5 concludes the paper.

2 Methodology

This study adopts an SLR methodology (Kitchenham, 2012; Kitchenham and Charters, 2007) to ensure a transparent, reproducible, and comprehensive synthesis of research on data-driven PHM for mining processing machinery. SLR is a recognized methodology that is widely used to identify, analyze and interpret relevant parts of research about specific issues, areas or phenomena of interest. The review process followed the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines (Page et al., 2021; Moher et al., 2009) which provide a standardized reporting structure to ensure transparency and reproducibility in systematic reviews. Figure 2 shows the PRISMA framework, which comprises the following phases.

• Identification: retrieving relevant studies from scientific databases and removing duplicates.
• Screening: excluding papers based on title and then abstract.
• Eligibility: assessing the full-text of the remaining papers against the inclusion criteria.
• Included: final set of studies used for qualitative and/or quantitative synthesis.

To structure the review, the following research questions are formulated.

• RQ1: What data modalities and equipment types are commonly investigated in PHM studies for mining processing machinery?
• RQ2: What are common data-driven modelling approaches applied to perform PHM?
• RQ3: Which PHM tasks (fault detection, diagnostics, and prognostics) are primarily addressed in the existing literature?

These research questions are classificatory in character, and the review incorporates elements of a systematic mapping study (Petersen et al., 2015), which characterizes the extent and distribution of research activity. This work is reported as a systematic literature review as it synthesizes all included studies thematically in Section 2.4, applies a quality assessment, discussed in Section 2.5, and analyzes their methodological limitations in Section 4.3.

2.1 Identification

2.1.1 Databases for literature search

The literature search is conducted exclusively in Scopus. This was a deliberate protocol decision rather than a constraint of convenience. Scopus provides broad multidisciplinary coverage of the engineering literature and indexes conference proceedings extensively, which is material for this review. Scopus also overlaps substantially with the discipline-specific engineering databases, so additional sources would be expected to yield duplicate records. Restricting the search to a single database allows the review to be reproduced from one unambiguous search chain. However, the consequences of this decision for the completeness of the corpus are discussed in Section 2.6.

2.1.2 Search strategy

For SLR execution, the choice of a specific keyword string is formulated and described as:

Article title, Abstract, Keywords

(“Mining” OR “Mine” OR “Mining Processing” OR “Mining Production Process”) AND (“Machine” OR “Equipment” OR “Plant” OR “Machinery”) AND (“PHM” OR “Prognostics and Health Management” OR “Prognostics” OR “Diagnostics”) AND NOT “data mining” AND (“Artificial intelligence” OR “Machine learning” OR “deep learning”)

2.1.3 Automatic exclusion criteria (EC)

• EC1: Publications prior to the year 2015. This is to ensure the review reflects existing data-driven PHM techniques, which have evolved significantly over the past decade.
• EC2: Non-English language publications.

The search was executed using the Scopus advanced search interface as a TITLE-ABS-KEY query, with EC1 and EC2 applied as database-level filters within the query itself. No document-type restriction was applied at the database level; journal articles, conference papers, and book chapters were all eligible, and non-research items such as editorials and errata were removed during title screening. The query returned 1,285 records, which were exported to Mendeley in RIS format. Since the search was confined to a single database, no duplicate records were identified or removed.

2.2 Screening

The screening process is carried out in two stages: (1) title screening and (2) abstract screening. Screening was performed by the first author, with records of uncertain relevance resolved through discussion with the co-authors. Although the search query was designed to target mining-related machinery, the initial results still included papers from unrelated domains, possibly due to overlapping terminology across disciplines. Therefore, the screening phase involves assessing titles and abstracts against the following EC to ensure relevance.

• EC3: Papers from unrelated subject areas such as Medicine, Biochemistry, Genetics, Molecular biology; Chemistry; Agricultural and Biological Sciences, Immunology and Microbiology, Pharmacology, Toxicology and Pharmaceutics, Veterinary, Nursing, Earth and Planetary Sciences, Environmental Science, Social Sciences, and Chemical Engineering.
• EC4: Papers with titles referring to unrelated domains, such as medical, parkinsonian freezing, e-learning, ADHD, text classification, wind power generation, and nuclear.

Title screening against EC3 and EC4 excluded 1,010 records, reducing the set to 275. During abstract screening, a further 179 records were excluded for the following reasons: the study addressed underground mining, drilling machines, proton exchange membranes, or other equipment not used for processing raw mining materials; the study addressed general PHM without a mining application; or the study addressed mining optimization without addressing a PHM task. Records were retained where they applied PHM, predictive maintenance, or health monitoring to the mining production process or its associated equipment. This left 96 articles to be assessed in full text.

2.3 Eligibility

The following inclusion criteria (IC) are applied to ensure relevance to the PHM in ore processing in the mining industry.

• IC1: Studies explicitly applying PHM, condition monitoring, and predictive maintenance to mining production process machinery.
• IC2: Papers involving assets used for mining ore processing.
• IC3: Studies addressing at least one PHM task: fault detection, diagnostics, or prognostics.
• IC4: Research focusing specifically on data-driven approaches, including machine learning and deep learning.

The four inclusion criteria IC1–IC4 were applied during full-text assessment, as determining whether an article addresses a defined PHM task and employs a data-driven approach generally requires examination of the methods section. Of the 96 articles assessed, 22 were excluded because the equipment studied fell outside the scope of mining ore processing, no PHM task was addressed, or the approach applied was not data-driven. A final set of 74 studies met all inclusion criteria and forms the corpus synthesized in this review.

A structured data extraction process is used by the first author during full-text assessment to capture information relevant to the PHM implementation of mining processing machines. Extraction was carried out uniformly across all the included studies. Extracted attributes included: mining subsystem or equipment studied, data type used, data acquisition sources, PHM tasks investigated, and data-driven PHM models applied. Papers are also categorized to highlight common challenges with a lack of labelled data, sensor drift, and domain shift.

2.4 Included

The selected studies are synthesized using a thematic analysis approach, following Braun and Clarke (2006). Thematic analysis enables the systematic identification of recurring patterns, methodological trends, and conceptual similarities across diverse literature. Given the variability across studies in equipment type, data characteristics, and modelling techniques, thematic analysis allowed the findings to be grouped into higher-level analytical themes. These themes included.

• the PHM tasks addressed (detection, diagnosis, prognosis),
• the data-driven modelling category (machine learning, deep learning), and
• the equipment category within mining processing (e.g., conveyors, crushers, rotating machinery, pumps, screens).

The three PHM tasks are not treated as mutually exclusive. Each study was assigned to every task it addressed, so a study that both detects and classifies faults is recorded under detection and diagnostics and appears in both corresponding summary tables.

This thematic synthesis provides a coherent structure for integrating diverse PHM techniques and enables comparative analysis across studies. This theme forms the foundation of the literature review and observation section, where insights from related studies are consolidated and critically discussed. Descriptive bibliometric indicators, comprising publication year and the specificity of the equipment addressed, were extracted alongside the thematic synthesis and are reported in Section 4.1.

2.5 Quality assessment

The included studies were assessed using a short checklist adapted from the criteria used in engineering systematic reviews (Kitchenham and Charters, 2007; Dybå and Dingsøyr, 2008). Three criteria were retained as relevant to this corpus: whether the aims and scope are clearly stated; whether the data source is described, with the dataset, test rig, or industrial source identified; and whether the results are validated and the limitations acknowledged. Each question was scored as fully reported, partially reported, or not reported, giving a maximum of three points per study. The assessment was carried out by the first author. It was used to characterize how completely the reviewed studies report their methods and data and was not applied as an exclusion criterion; therefore, all 74 studies are retained in the synthesis.

2.6 Methodological boundaries

This review has certain methodological boundaries. As the review targets data-driven PHM approaches, the search string includes Artificial Intelligence-, Machine Learning- and Deep Learning-specific terms. The relevant studies employing non-AI data-driven methods may not have been captured. The restriction to a single database is the most consequential boundary of this review and introduces several forms of selection bias. Studies indexed only in Web of Science, IEEE Xplore, or Engineering Village/Compendex are absent by construction, as are those in journals with limited international indexing. The EC2 likely excludes a non-random subset, given that much of the corpus originates from contexts where results are also published in the national language. Literature such as industrial technical reports and vendor documentation falls outside the search entirely. Screening, data extraction, and quality assessment were performed by a single author, with the co-authors consulted when clarification was needed. The EC1 is applied to focus the review on recent data-driven PHM techniques. This cutoff may exclude earlier studies that employed foundational methods. However, the introduction and background sections include pre-2015 works to provide contextual grounding on PHM concepts. The review therefore characterizes the Scopus-indexed, English-language, peer-reviewed literature rather than the complete body of work, and the prevalence findings in Section 4 should be read accordingly. The qualitative gaps identified in Section 4.4 are less sensitive to this boundary, since they concern what the reviewed studies do not attempt rather than how frequently particular approaches occur.

3 Literature review

The reviewed papers on data-driven PHM are organized according to the three core PHM tasks: fault detection, diagnostics, and prognostics. Within each task, studies are further refined by the type of equipment being monitored, the data modalities used to monitor, modelling approaches used and the HIs. The review includes both studies explicitly focused on mining processing machines and studies on general industrial machinery. Table 1, Table 2 and 3 provide a structured summary of data-driven PHM studies in mining production machines. The tables map studies to dimensions such as equipment, modelling approaches, data modality and HIs across the three PHM tasks. Studies addressing more than one task appear in each corresponding table. In cases where specific information is not reported in the original studies, the corresponding table cells are left blank to reflect the unavailability of data. The final column reports what each study uses to represent machine condition. Entries span a range of constructs: raw measurements, statistical descriptors of the signal, and learned representations extracted by deep networks. Relatively few studies construct a health indicator in the stricter sense of a single quantity.

TABLE 1. Summary of reviewed studies for the fault detection task.

TABLE 2. Summary of reviewed studies for the fault diagnostics task.

TABLE 3. Summary of reviewed studies for the fault prognostics task.

Table 4 provides a comparative overview of the three tasks across the reviewed studies, summarizing the analysis level, dominant data modalities, modelling approaches, health indicator strategies, and reported challenges associated with each.

TABLE 4. Comparative summary of the three PHM tasks across the reviewed studies.

3.1 Studies focused on fault detection in mining processing machinery

For the detection task, the focus is on identifying abnormal conditions or signs of degradation before significant damage happens. The reviewed studies are organized by equipment category: conveyor systems, infrastructure and environmental systems, and other subsystems. Within each category, the progression from single-sensor to multimodal and learning-based detection methods is discussed.

For conveyor-related equipment, vibration, acoustic, and vision-based sensing are commonly used for early fault detection. Vibration-based condition monitoring of conveyor belt drum bearings using long-term operational data demonstrates that random forest outperforms other models in accuracy and minimizing false alarms (Afonso Bauler et al., 2024). Acoustic signal analysis with Polar K-Nearest Neighbor (KNN) improves traditional KNN by including distance and angular similarity for anomaly detection of conveyor-roller (Liu et al., 2025), while a non-destructive ultrasonic diagnosis with sensors mounted above and below the belt provides real-time monitoring of belt thickness, wear, and local damage (Kirjanów-Błażej et al., 2022). Vision-based smoke detection along coal conveyor belts combines features from optical flow, RGB color histograms, and contour irregularity and fuses them through rule-based logic to reliably distinguish smoke from dust, achieving high detection accuracy (Huang X. et al., 2024). These studies highlight the effectiveness of combining multiple sensing modalities for improving early detection capabilities in conveyor systems.

In addition to equipment-level monitoring, several studies focus on mine infrastructure and environmental systems where fault detection supports safety and operational awareness. A machine-vision-based framework for scene classification in open-pit coal mines fuses shallow texture features extracted using Gray-Level Co-occurrence Matrix (GLCM), and Gabor filters with deep features from VGG-16 and Graph Convolutional Network (GCN), and classifies scenes using Particle Swarm Optimized-Support Vector Machine (PSO-SVM), enabling situational awareness (Liu and Zhang, 2024). Similarly, data-driven approaches applied to mine ventilation systems use multi-label classification techniques, including KNN, decision trees, and a multi-layer perceptron, to detect and localize multiple concurrent faults (Wang D. et al., 2022; 2023).

Other machines and subsystems are investigated using alternative sensing modalities such as stress-wave, pressure, and acoustic-based approaches. By extracting time-frequency features, early fault signatures are detected via stress waves, using polynomial neural networks and Bayesian data fusion for condition classification to provide early indicators of degradation (Wu et al., 2018). Desmet and Delore (2017) proposed an unsupervised anomaly detection method combining wavelet-based pressure features, random forest-based feature selection, local outlier factor, and autoencoders to detect developing leaks for a compressed air system. More recently, the application of Large Language Models (LLMs) to detect leaks in natural gas infrastructure is presented in (Wei et al., 2024). The proposed approach integrates acoustic and infrared temperature monitoring with a pre-trained sound classifier, dynamic time warping for temperature analysis, and an LLM-based agent with retrieval-augmented generation for knowledge-based decision support.

3.2 Studies focused on fault diagnostics in mining processing machinery

Diagnostics is the most extensively addressed PHM task in the reviewed literature. The studies are broadly categorized based on modelling approaches, including classical ML methods, DL architectures, and multi-sensor or graph-based techniques. For bearings and rotating components, a wide range of ML and DL architectures is reported. Classical diagnostic frameworks use time-domain and frequency-domain features such as root mean square (RMS), crest factor, entropy, and wavelet energy, combined with fuzzy logic or tree-based classifiers. Zhang F. et al. (2024) classifies rolling bearing failures by integrating real-time monitoring, feature extraction, and fuzzy control logic. Similarly, vibration data is decomposed using Discrete Wavelet Transform (DWT) and wavelet energy features are extracted, followed by J48 feature selection and Random Forest classification, which are used for bearing fault diagnosis in an induction motor (Arunkumar and Maniunath, 2018). Hybrid approaches combining backpropagation neural networks, fuzzy inference, and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) demonstrate improved diagnostic accuracy by balancing learning capability and interpretability (Xu et al., 2020). Stress-wave-based diagnostics using ultrasonic sensors and Bayesian fusion enable the classification of multiple fault sources (Wu et al., 2018). Signal processing techniques, including envelope spectra, cyclic spectral coherence, spectrograms, and phase-space plots, are applied to identify bearing faults in vibrating screens (Wodecki et al., 2023). These approaches highlight the relevance of feature engineering and interpretable models in diagnostic tasks.

Deep learning methods provide more automated diagnostic solutions by reducing reliance on manual feature extraction. End-to-end 1D Convolutional Neural Network (CNN) eliminates manual feature engineering by learning directly from raw vibration data to classify bearing fault types (Wang Y. et al., 2022). Yang et al. (2022) introduces architectures, including an enhanced Deep Belief Network (DBN) with joint class-metric and sparse representation regularization; while Xue et al. (2023) apply multi-scale feature fusion with adaptive weighting. A hybrid model combining Convolutional Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) with attention mechanism jointly learns local and global temporal features, enhancing fault-relevant representation (Xu and Yang, 2022). To improve the diagnosis of bearing and rotor imbalance faults (Meng et al., 2019), combines learning segmentation aligned with rotational speed and data augmentation strategies with sparse-autoencoder-based Deep Neural Networks (Yang et al., 2022), while (Cai et al., 2022) proposes that an ensemble of dilated CNNs exploits multi-scale receptive fields for bearings and gearboxes. These studies demonstrate the increasing adoption of DL for handling diagnostic tasks.

Multi-sensor and graph-based diagnostic models are applied to capture dependencies in complex machinery. Graph-based models, such as GCNs and Graph Attention Networks (GATs), represent sensors as nodes to capture dependencies and improve diagnostic accuracy (Sun et al., 2024; Li C. et al., 2022). Multi-sensor fusion frameworks using deep features extracted by CNN enhanced with a hybrid attention mechanism to capture inter-sensor correlations to enhance feature representation (Wang et al., 2025). Data-Enhanced models, DBN and 1D CNN, for fault diagnosis in rotatory machines are proposed in (Han et al., 2024). Data-Enhanced Squeeze-and-excitation Deep Belief Network (DEDBN) uses a Restricted Boltzmann Machine for acoustic signals, whereas for vibration signals, a Data Enhanced 1D CNN is implemented, which uses a squeeze-and-excitation network to emphasize fault-relevant information. Dimensionality reduction techniques, such as principal component analysis (PCA), combined with enhanced sparse filtering with softmax classifiers to demonstrate that efficient low-dimensional features can support real-time diagnostic deployment (Ling et al., 2024).

Beyond rotating machines, diagnostic approaches are also applied to a variety of mining subsystems. Fault diagnosis of coal mine belt conveyors uses PCA to reduce 19 sensor features and SVM parameters tuned by a Hybrid Grey Wolf Optimizer to classify seven conveyor states (Li et al., 2020). Mine main ventilator-related faults are addressed in multiple studies, Tian et al. (2024) combines vibration and wind-pressure data using ensemble empirical mode decomposition to extract intrinsic mode functions and classify faults via back propagation neural networks. Wang et al. (2023) applies a data-driven multilabel classification of airflow measurements with decision trees, MLP, and ranking SVM, enabling simultaneous identification of multiple fault locations in a ventilation network, while Wang D. et al. (2022) addresses fault localization. Hydrocyclone and grinding systems are addressed through multi-sensor condition monitoring, where vibration, acoustic emission, force, strain, power, and machine control data are synchronized with grinding cycles, and a quadratic SVM classifier is used to distinguish fault conditions (Ahmer et al., 2020). For hoisting components (Li Q. et al., 2024), presents a wear detection framework using a vision method, comprising an industrial camera, a laser and a light source to capture wear of guide pairs to measure wear position and magnitude under harsh operating conditions. A non-invasive diagnostic method, such as thermography, vibrometer pens, and state testers, is applied to bearings, electrical connections, and other components to enable early wear detection and condition assessment in real-time (Bialy, 2024). Coal bursting liability is also treated as a diagnostic classification problem, where an index system based on dynamic fracture duration, elastic strain energy index, bursting energy index, and uniaxial compressive strength is modelled using a back propagation neural network (BPNN) and SVM (Wang et al., 2024a).

3.3 Studies focused on fault prognostics in mining processing machinery

Many papers focus on bearings and rotating machinery, where degradation is captured via vibration, torque, or multi-sensor features and transformed into HIs. These studies are broadly classified along a technical development path from traditional time-series methods, to DL approaches and, more recently, techniques for handling uncertainty and limited data. Traditional approaches rely on statistical modelling for degradation analysis. For instance, to forecast bearing failures (Primera et al., 2024), use traditional time-series forecasting methods, such as Autoregressive Integrated Moving Average (ARIMA) and Holt–Winters, applied to wireless vibration measurements in roller mills. Similarly, a prognostics framework is proposed in (Bejaoui et al., 2020) to extract torque-based features, select them via monotonicity, fuse them into a single HI using PCA, and fit an exponential stochastic degradation model to estimate RUL for motors. DL-based methods focus on learning patterns directly from data. For bearing RUL prediction, Ma and Mao (2021) transforms run-to-failure vibration data into time-frequency representations via Short Time Fourier Transform (STFT) are used in a convolution-based LSTM network that jointly learns temporal and time-frequency degradation patterns. Similarly (Liu et al., 2021), focuses on constructing smooth, monotonic HIs from vibration features and leveraging an LSTM encoder-decoder with attention to learn degradation sequences and forecast future health states.

Health indicator construction and degradation modelling remain central themes in prognostics studies. Akpudo and Hur (2020) extracts statistical features, selects those with high monotonicity and fuses them via PCA into a single HI. To identify degradation onset, Fuzzy C-Means is used, and gaussian process regression is compared with DBN for RUL forecasting of rolling bearings in (Akpudo and Hur, 2020). For scenarios with limited data, for RUL prediction of bearings under few-shot conditions, (Zhuang et al., 2024) proposed a novel Transferable Autoregressive Recurrent Adaptation (TARA) framework, which combines adversarial data augmentation, deep autoregressive recurrent modelling, a semantic attention module, and meta-learning.

To address challenges related to noise and complex operating conditions, several studies propose adaptive modelling techniques. Under non-Gaussian noise conditions, (Shiri et al., 2024) replaces the standard Extended Kalman Filter (EKF) with a Robust-EKF based on the maximum correntropy criterion, which improves noise resilience and prediction accuracy on benchmark datasets. A DBN with a feedforward neural network architecture is proposed in (Deutsch and He, 2018) to automatically learn degradation features from continuous vibration data and predict RUL of gears and bearings. Advanced RUL architectures such as the Gated Transient Fluctuation Dual Attention Unit (GTFDAU) explicitly model transient fluctuations in HIs through a combination of transient gating, historical attention, and current attention mechanisms, applied to industrial IoT data for rotating machinery (Li S. et al., 2024).

Process machines and system-level prognostics are addressed in a smaller number of studies. For hydrocyclone prognostics, (Liang et al., 2024) integrates a mechanism-based simulation model with TimeGAN to generate realistic degradation sequences for health state assessment and supports both direct RUL prediction and uncertainty-aware RUL estimation. At the system level, (Malozyomov et al., 2024) explores statistical reliability models for mine hoists. Key reliability metrics, including failure intensity, mean time to failure, and repair time, are derived from historical records to assess equipment controllability and predict failures. Frameworks that combine fault classification with degradation-trend estimation, such as stress-wave-based PHM (Wu et al., 2018) using ultrasonic sensors and Bayesian fusion, link diagnostics and prognostics by providing both current condition assessment and early estimation of future deterioration.

3.4 Studies discussing challenges relevant to mining processing machines

A large portion of reviewed studies focused on challenges in applying PHM in mining processing machines, including limited labelled data, small-sample learning, noisy vibration environments, imbalanced fault classes, and inconsistent data dimensions or sampling rates. These challenges are addressed through a range of advanced modelling techniques, which can be grouped based on the nature of the problem. One of the primary challenges is the scarcity of labelled data. To address this, pseudo-labels generated from time and frequency domains with an invariance constraint are used to strengthen self-supervised learning to reduce information loss from inconsistent time-frequency representations (Tang J. et al., 2024). Meta-learning strategies support learning with few source samples, such as the optimal transport meta-attention network learning the fault knowledge from the source machine and transferring it to diagnose the target machine (Wu et al., 2024). To enable RUL prediction from limited labelled data, supported by a lightweight contextual memory module for rapid task adaptation, Yang and Wang (2024) proposes a meta-learning with a flow-enhanced Gaussian process. To overcome incomplete information from a single sensor, Lu et al. (2024) fuses operating condition and vibration data via interactive spatiotemporal attention.

Noisy and non-stationary operating environments present another major challenge. Under small-sample and noisy conditions (Ye et al., 2025), introduces a multi-sensor residual convolutional fusion network that combines vibration and acoustic signals to suppress noise and extract more discriminative features. Robust feature learning under noise is addressed in (Zhang Y. et al., 2024) using multiscale dilated convolution and swin-transformer architectures with coordinate reconstruction attention. Empirical-mode-decomposition-based multi-sensor graph networks decompose raw signals into intrinsic modes before fusion to mitigate noisy and non-stationary features (Sun and Yin, 2025). A hybrid CNN-LSTM model with an attention mechanism fuses time-domain and frequency-domain features to address few-sample composite faults (Liu et al., 2023).

Imbalanced datasets are tackled using attention-based multiscale feature aggregation combined with dual focal loss to emphasize minority and hard-to-classify samples (Xu et al., 2024), weight-activation modules in multi-sensor graph networks to reduce majority bias, (Deng et al., 2023) and a knowledge discovery in databases framework that integrates data balancing at multiple stages (Wagner et al., 2016). Distributed sampling-rate inconsistencies are managed via multi-rate CNN-LSTM fusion frameworks with batch normalization (Huang et al., 2021).

Mining processing machines rarely operate under fixed conditions, which causes a domain shift between source (training) and target (deployment) data. Some PHM studies focused on the inability of models to generalize across domains. Vision Transformer-based feature extraction combined with partial adversarial training and knowledge distillation aligns marginal and conditional distributions for cross-domain diagnosis (Yu K. et al., 2024). Ensemble learning integrates angular signal processing, CNN features, and unsupervised clustering to detect unseen compound faults (Jiang and Xiang, 2023). Multiscale transfer networks that integrate multiscale feature extraction with transfer learning are proposed to reduce negative transfer effects between mismatched domains (Ge et al., 2024). In (Yu Y. et al., 2024), a dual-graph strategy that evaluates sensor importance before extracting domain-invariant multi-sensor features to address multi-source learning is discussed. Shi et al. (2023a) proposes MARTAN (Multi-representation Transferable Attention Networks), which combines convolutional degradation-feature extraction, transferable attention, and representation alignment to improve cross-domain robustness. Noise-aware uncertainty-perception metric networks estimate class-prototype similarities using probability distributions to handle noisy, scarce target domains (Wang et al., 2024b).

To address subdomain discrepancies across heterogeneous operating conditions, (Lu and Lee, 2024) use a Feature-Enhanced Multi-source Subdomain Adaptation (FEMSA) framework. This method redefines multiple source domains into subdomains using clustering and aligns them with the target domain through Maximum Mean Discrepancy (MMD) and correlation alignment (CORAL) loss metrics. In (Shi et al., 2023b), target-free generalization is explored through Domain Transferability-based Deep Domain Generalization (DT-DDG), which learns universal diagnostic features without access to target samples via dynamic weighting and spectral penalization. Causal disentanglement networks separate causal and domain-specific components of features to support continuous degradation and cross-machine diagnosis (Li et al., 2023), while (Li J. et al., 2022) discusses compound-domain whitening and causal learning, and extracts invariant features without retraining across working conditions. Representation Gradient Muting Paradigm (RGMP) proposed by Tang et al. (2022) suppresses dominant gradients to improve generalization under multiple unseen conditions, and Zhuang et al. (2022) introduces an Adversarial Domain Generalization framework with Self-Supervised Regularization Learning (ADGR), which generates latent domains and reduces feature drift under unseen environments. A dense-connection CNN with depth-wise separable convolutions, combined with fine-tuning, which updates only a small portion of parameters to improve generalization across machines and datasets without losing previously learned features, is discussed in (Qin et al., 2024). To handle distribution shifts between training and test data, (Tang Z. et al., 2024) proposes a Regularized Domain Adaptive Weight Optimization (RDAWO) integrated with 1D CNNs, combining hyperparameter regularization and domain-adaptive weighting of training samples for robust fault diagnosis under unseen conditions.

Finally, several works focus on data with strong temporal dependencies, cross-sensor interactions, and multi-scale fault signatures. Multi-sensor interactive graph attention networks treat sensors as graph nodes and use attention to learn spatial interactions while filtering redundant sensors for RUL estimation (Ding et al., 2022). To capture higher-order correlations among samples at different scales, (Yan et al., 2022) proposed a multiresolution hypergraph neural network that connects samples via KNN relations, capturing both coarse and fine structural information for fault diagnosis. Planetary gear systems, which produce complex coupled vibrations, are addressed through comprehensive reviews (Wang et al., 2016) combining dynamic modelling with signal-processing techniques, including the Fourier Transform (FT), Wavelet Transform (WT), and Empirical Mode Decomposition (EMD) for extracting multi-fault features. Unsupervised Disentangled Time Mining (DTM) (Qu et al., 2018), an unsupervised DL method, uses deep sparse autoencoders to extract fault signatures from frequency spectra while separating noise components. Multi-scale spatiotemporal attention networks (MSTAN) that combine multi-scale feature perception, adaptive relationship mining, and spatiotemporal attention are proposed in (Huang K. et al., 2024) to learn sensor correlations for accurate degradation forecasting and RUL prediction. For machines under variable speed and limited sample conditions, Zhang et al. (2023) proposes a hybrid framework based on Variation Amplitude Normalization (VAN) that stabilizes vibration amplitudes and supports both shallow and deep diagnostic approaches. For prognostics under variable operating conditions and limited, variable-length sequences, (Ding et al., 2022) proposed meta-learning with an attention mechanism combined in a Meta-Attention Recurrent Neural Network (MARNN). The reviewed studies highlight that addressing real-world challenges in mining processing machines requires robust and adaptive approaches.

4 Observations

4.1 Bibliometric characteristics of the corpus

The 74 included studies are unevenly distributed over the review window, as shown in Figure 3. No study meeting the inclusion criteria was published in 2015, and only 16 studies (22%) appeared between 2016 and 2021. From 2022, the corpus expands sharply: 58 studies (78%) date from 2022 or later, and 34 studies (46%) of the corpus were published in 2024 alone. Figure 4 classifies the studies by the equipment addressed. Nineteen studies investigate equipment specific to mining or mineral processing, including conveyors, mine ventilators, hydrocyclones, grinding machines, hoisting systems, sieving screens, and roller mills. A further 26 address generic rotating machinery, principally bearings and gearboxes, and the remaining 29 address domain-independent methodological challenges, of which 17 concern data limitations and 12 concern domain shift.

4.2 Trends of the reviewed literature

The analysis of the selected literature reveals dominant trends and gaps in data-driven PHM research for mining processing machines. The observations reflect several recurring patterns across modelling approaches, sensing modalities, equipment types, and maintenance objectives. These patterns help in identifying the existing research gaps and opportunities for advancing PHM in mining processing environments.

4.2.1 Asset coverage

Among the 74 papers reviewed, some papers focus on general PHM challenges, whereas some studies specifically address data-driven PHM for mining processing machines. Within mining-related studies, the focus is predominantly on mechanical components, as shown in Figure 5, where bearings, gearboxes, shafts, rollers, and ventilation equipment appear most prominently.

Figure 6 displays a strong emphasis on component-level analysis. While components such as bearings, gearboxes, and shafts, which dominate the investigated assets, are critical to mining production, this narrow scope means that system-level PHM is comparatively underexplored. While mechanical components dominate the assets investigated in the reviewed PHM studies, other subsystems, such as electrical, hydraulic, and control systems, receive limited attention in comparison. The word cloud from Figure 5 shows no visible representation of these non-mechanical subsystems. This gap is notable given that mining processing machines are complex systems comprising not only mechanical components but also electrical, hydraulic, pneumatic, control and software-based systems (Hodouin, 2011; Sbarbaro and del Villar, 2010; Dhillon, 2008), which are also susceptible to failure.

4.2.2 Data characteristics and sensing modalities

As shown in Figure 7, 50% of the reviewed papers rely on public datasets, while 25% use laboratory test rigs. Only around 16% of the studies utilize real industrial data. This trend can be linked to the fact that many mining operations still rely on scheduled preventive maintenance, where continuous condition monitoring systems are not yet deployed, limiting access to degradation data. Real industrial datasets remain limited, which might be due to accessibility, confidentiality constraints, and the difficulty of collecting long-term labelled degradation data, making them costly to reproduce. As a result, lab test rigs facilitate the development and validation of experimentation; however, they may not fully represent the harsh variability and load fluctuations inherent to machines in a mining environment.

Specifically, laboratory data are typically generated under controlled conditions with stable loads, limited environmental disturbances, and well-defined faults, whereas in-situ mining data involves variable, multi-component load variations, contaminated signals and unknown disturbances. This discrepancy poses risks for models trained on lab data when deployed on real industrial data.

Data sensing modalities, illustrated in Figure 8, show a strong skew towards vibration signals, which account for around 50% and dominate the PHM studies in mining processing machines. Acoustic, torque, and temperature measurements appear to a much lesser extent, while a broad group of additional modalities, including force, belt tension, airflow, and others, collectively form the rest. The word cloud further reinforces this imbalance in data modalities, highlighting vibration-related terms as the most prominent inputs. Multimodal sensing, which combines the dominant vibration signals with electrical signature, thermal imaging, visual or 3D data, remains uncommon in mining processing machines. This narrow focus on vibration-based analysis limits the capability of fault detection and prediction that manifests in electrical, hydraulic, or software/control subsystems, which are increasingly relevant in modern mining machines.

4.2.3 Modelling approach

The modelling landscape, as shown in Figure 9, reveals that supervised ML and DL methods dominate, particularly for fault detection and diagnostics. SVM, CNN, LSTM networks, and DBN appear most frequently. These supervised algorithms depend heavily on high-quality labelled datasets, which in reality are limited given the scarcity of real-world fault annotations.

4.2.4 PHM task focus and modelling trends

As shown in Figure 10, diagnostics receives the highest research attention with 49.2% of task assignments, followed by fault detection at 35.6%. Prognostics remains the least explored task, representing only 15.3%. This imbalance suggests that while there is substantial progress in identifying and classifying faults, predicting RUL under varying loads, material properties, and environmental conditions remains a significant challenge for a mining processing machine in real-world setting. Modelling techniques used across different PHM tasks reveal preferences for certain algorithms. Figure 11 shows the top 3 algorithms used in the PHM tasks. CNN and SVM appear as the most frequently adopted algorithms for fault detection and diagnostics. In contrast, prognostics rely mainly on temporal DL models like LSTM and less on traditional statistical approaches like ARIMA.

4.2.5 Coverage of health management in PHM

Across the reviewed studies, only a subset of papers explicitly relates to health management. Health management is the process of decision-making based on the estimated health state and the expected future use of the asset (IEEE, 1856, 2017). Health management might include maintenance scheduling optimization, spare parts logistics and resource allocation, cost-risk trade-off analysis, and operational and maintenance decision-support systems. Among the reviewed studies, only a few demonstrate the health management integration.

A conveyor-belt thickness monitoring system in Kirjanów-Błażej et al. (2022) links degradation to replacement planning and life-cycle resource allocation (Wu et al., 2018). offers a complete PHM workflow combining diagnostics, prognostics, thresholding, and maintenance action guidance. By producing RUL, several studies support health management that might enable timely interventions. Reliability-oriented work (Malozyomov et al., 2024) and prognostics architectures included in (Lu et al., 2024; Yang and Wang, 2024; Deng et al., 2023; Zhuang et al., 2022; Huang K. et al., 2024) contribute to health management by informed maintenance scheduling, availability optimization, and operational planning.

As shown in Figure 12, only 27% of the reviewed papers explicitly address health management by linking PHM outputs to decision-making. The majority of papers focus on health monitoring, detection, or diagnosis without extending their contributions to actionable operations and maintenance planning.

4.3 Critical analysis of the reviewed literature

Reporting in the reviewed corpus is generally complete. All 74 studies clearly stated their aims and scope; 69 identify their data source fully and four partially and one does not report it; 58 both validated their results and acknowledged limitations, with the remaining 16 doing one of the two. Scores are correspondingly clustered towards the upper end, with 58 studies at the maximum of three points and a mean of 2.85. These figures indicate that the reviewed studies describe their own work adequately; they do not indicate that results are comparable between studies.

Beyond the distribution of topics and methods, the reviewed studies exhibit several recurring methodological limitations that constrain how far their reported results can be transferred to operating mining processing plants. Reported performance is not comparable across studies. Almost all reviewed papers report high classification accuracy on their respective evaluation sets, yet these figures are not mutually comparable. Studies differ in the dataset used, the number and definition of fault classes, the severity of the seeded faults, and the evaluation metric reported. A diagnostic accuracy obtained on a small set of well-separated seeded faults in a laboratory rig is not equivalent to the same figure obtained on naturally occurring faults in industrial data, even where both are reported as classification accuracy. Consequently, the high accuracies that dominate the literature should be considered as evidence that the methods function under the conditions tested, rather than as evidence of comparative superiority or of readiness for deployment.

Validation data is largely disconnected from mining operating conditions. As shown in Figure 7, half of the reviewed studies for which the data source was recorded rely on public datasets and a quarter on laboratory test rigs, against approximately 16% using real industrial data. The public datasets most commonly used in this literature are general rotating-machinery bearing datasets rather than mining-specific ones. The studies that do use industrial mining data are, correspondingly, the most informative about achievable performance in practice and remain a small minority. Increasing model complexity is not consistently matched by demonstrated benefit. The reviewed corpus displays a clear methodological progression, from feature engineering with classical classifiers, through convolutional and recurrent architectures to attention mechanisms, graph networks, transformers, and meta-learning. However, comparative baselines in the reviewed studies are typically drawn from the same or an adjacent model family, and well-tuned classical alternatives are rarely retained as competitive baselines. Where classical methods are evaluated directly against learned representations on identical data, they remain competitive: random forest is reported as the strongest performer among the models. The literature therefore does not currently establish that architectural advancement is the binding constraint on PHM performance in this domain; data availability and quality appear to be the more restrictive factor.

Health indicator construction is seldom validated against physical degradation. As Tables 1–3 show, most studies represent machine condition through statistical descriptors of the vibration signal or through representations learned by the model, rather than through a constructed HI. Where an indicator is constructed, it is typically assembled by fusing selected features, with dimensionality reduction used for fusion and monotonicity used for selection. More commonly, the constructed indicator is assessed by the accuracy of the downstream prediction rather than by correspondence to an independently established degradation state.

Deployment feasibility is largely unaddressed. Few reviewed studies report inference cost, memory footprint, latency, retraining requirements, or the sensor and computing infrastructure their method presupposes. For mining processing plants, where instrumentation is frequently sparse and legacy control infrastructure is common, these constraints determine whether a method can be adopted at all.

4.4 Gaps and future directions

Based on the observed trends and the limitations identified in Section 4.3, the following research gaps and future directions are proposed to guide future advancement in implementing PHM in mining processing machines. They are then considered in terms of expected practical impact and feasibility. The first three directions follow from patterns observed in the reviewed studies. The fourth is prospective and is drawn from developments outside the reviewed literature.

How can PHM move beyond component-centric monitoring? Most studies focus on monitoring individual mechanical components, such as bearings or gears, while electrical, hydraulic, and software subsystems receive less attention. This leaves a significant gap in the development of PHM methods for integrated system-level maintenance. Multi-modal sensing and data-fusion techniques can improve observability of system behaviour, enhance diagnostic accuracy, and support holistic health estimation. To move beyond component-centric monitoring, future research can explore multi-sensor data fusion at both the feature and model levels. Such approaches can be evaluated using metrics including robustness under varying operating conditions, the ability to detect interdependent faults, and system-level performance indicators.

How can PHM research reduce its dependence on supervised learning, given the scarcity of labelled industrial fault data? Supervised learning, which most of the studies focused on, requires a high-quality labelled dataset. However, in real-world applications, due to confidentiality constraints, safety-critical environments, and the predominance of preventive maintenance practices, fault data is rarely available. As a result, many studies depend on open public datasets or laboratory test rigs. To reduce dependence on labelled data, future research can explore self-supervised learning approaches, as well as meta-learning techniques for learning under limited samples. Generative models can be used to augment the fault data; such data should be verified with domain experts. Collaborative efforts to develop anonymized industrial datasets need to be promoted. Also, there is a need for the industry to transition from legacy systems to adapt advanced technologies. Researchers could explore domain adaptation techniques to transfer knowledge from the lab to real field settings.

How can health management be strengthened? Several studies produce RUL estimates, but only a few integrate the output with maintenance planning, scheduling, and operational actions. Health management requires deciding and executing responses such as repair, replacement, shutdown, and resource reallocation based on current health and future requirements. To move beyond prognostics, future research might focus on coupling prediction models with cost-risk optimization that balances intervention costs against failure risks. Uncertainty-aware maintenance policies could also be incorporated, as confidence in the RUL prediction directly influences scheduling decisions. Additionally, developing scenario-based decision support would allow planners to evaluate alternative maintenance strategies. Adopting a structured framework could provide a systematic basis for connecting sensing and data acquisition, analytics, and decision-making to support the maintenance process. System availability and decision effectiveness under uncertainty can be used as evaluation metrics.

How can some emerging technologies be exploited to advance PHM in mining processing machinery? This direction has a different basis from the preceding three. Of the 74 reviewed studies, one applies a large language model, and none address agentic AI or augmented reality (AR). The rapid evolution of sensing, modelling, and visualization technologies presents new opportunities for PHM. Advanced sensors such as LiDAR, thermal cameras, and acoustic measurements can provide richer information; recent developments in modelling techniques, including LLMs and agentic AI, can support intelligent reasoning. Visualization of machine health and maintenance instructions can be enhanced using an AR interface. Exploring how these technologies can be effectively integrated into the PHM framework may enhance machine health awareness, automation, and decision support. This can be achieved through combining heterogeneous sensor inputs and advanced models for monitoring and predicting. In addition, the AR interface can be linked to provide context-aware decision support.

The four directions above differ in the extent to which addressing them would affect maintenance practice and in how readily they can be pursued with currently available methods and data. Table 5 summarizes this. Reducing dependence on labelled data appears the most immediately tractable, as suitable methods are already present in the reviewed corpus and the main constraint is access to data rather than modelling capability. Strengthening health management would connect prognostic outputs to maintenance decisions, but requires cost and operational information alongside condition data and, therefore, closer collaboration with industry. Moving beyond component-centric monitoring depends on sensing coverage that is not yet common in processing plants. The exploitation of emerging technologies is the least evidenced in the reviewed literature and rests to a large extent on progress in the preceding three directions.

TABLE 5. Research directions ordered by expected practical impact and feasibility.

5 Conclusion

This systematic review presents a comprehensive analysis of research studies related to data-driven PHM applied to mining processing machinery. Using a structured strategy and following the PRISMA guideline, 1,285 articles are screened on title, 275 are retained for abstract screening, 96 articles are assessed in full text, and a final set of 74 studies meeting the inclusion criteria related to PHM tasks, mining processing relevance, and data-driven methodologies is identified. Each paper is examined in depth with respect to sensing modalities, modelling approaches, equipment categories, and health indicator construction strategies across the PHM tasks of fault detection, diagnosis, and prognosis, enabling a detailed characterization of the current research landscape. The synthesis is guided by a set of research questions and supported by a thematic analysis of the selected papers. This approach enabled the identification of recurring methodological patterns, emerging trends, and notable gaps, which can be of interest to both researchers and practitioners.

The findings show that data-driven PHM in mining processing machines remains fragmented and component-centric, with most studies focusing on isolated equipment elements such as bearings, gearboxes, or motors. Vibration data dominates the sensing modality, though studies using torque, temperature, acoustic, and visual data highlight the potential of multimodal sensing. Health indicators are predominantly constructed from statistical features like RMS, entropy and fused with methods such as PCA, though some studies explore deep learning-based approaches for automated health indicator construction. In terms of modelling approaches, supervised learning is the preferred modelling paradigm and shows strong performance on public and laboratory datasets. However, the real-world adoption of PHM solutions remains limited, often due to data scarcity, harsh environmental conditions, heterogeneous equipment, and low generalizability of supervised models. Despite these challenges, the reviewed articles demonstrate encouraging progress, including the use of attention-based architectures, domain adaptation methods, and multimodal feature learning to address issues such as noise, non-stationarity, and cross-domain variability.

By synthesizing existing methods and challenges, this review also identifies broader research gaps and some open questions that can guide future research. System-level PHM is underexplored compared to isolated component-level studies. Furthermore, electrical, hydraulic, pneumatic, control, and software subsystems are comparatively underrepresented in the existing data-driven PHM literature. Health management is seldom addressed, with limited work linking prognostics to actionable decision-making. Addressing this may require a shift towards multimodal sensing, reducing reliance on labelled fault data through semi-supervised and self-supervised learning with the integration of generative models. Emerging technologies such as advanced sensors, recent modelling developments, and enhanced visualization tools could offer further opportunities, although these are largely absent from the reviewed studies and are identified here as an outlook rather than as a finding of the review. Progress in these directions may support more resilient, continuous, and cost-effective operation of mining processing machinery.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used ChatGPT, a language model developed by OpenAI, and Grammarly to assist with grammar refinement and text clarity. All content was reviewed and edited by the authors, who take full responsibility for the content of this publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Viet Q. Vu, Thai Nguyen University of Technology, Vietnam

Reviewed by: Nagesh Tambake, Walchand Institute of Technology, India

Mohamed Lamine Soudani, Universite Hassan II Casablanca Ecole Nationale Superieure de l'Enseignement Technique, Morocco