Deep beneath the surface of the mining industry, a quiet data revolution is stalling at the component level. Ore processing plants depend on enormous, heavy, interconnected machines — crushers, grinding mills, conveyor belts, ventilation systems, hydrocyclones — running continuously under brutal conditions. When one of these assets degrades, the consequences cascade fast: unplanned downtime, production loss, safety risks, and ballooning maintenance costs. A new systematic review published in Frontiers in Manufacturing Technology by researchers at Luleå University of Technology in Sweden reveals that while artificial intelligence has transformed maintenance in aerospace, automotive, and railway industries, its application to mining processing machinery remains surprisingly narrow, fragmented, and — critically — disconnected from real-world decision-making.
The review, led by Sameer Prabhu with co-authors Ravdeep Kour, Pierre Dersin, and Ramin Karim, screened 1,285 research articles published since 2015 and ultimately synthesized 74 studies that applied data-driven Prognostics and Health Management (PHM) to mining processing machines. PHM is the engineering framework that converts raw sensor data into meaningful health indicators, detects anomalies, diagnoses their causes, and predicts how much longer a component can function before failure — its remaining useful life, or RUL. The approach underpins predictive maintenance, where interventions are scheduled based on the actual health of the machine rather than fixed calendar intervals or reactive repairs after breakdown.
The mining processing pipeline itself is a well-established sequence of transformations. Raw ore arriving at the plant undergoes crushing and grinding to reduce large rock fragments into smaller particles, sizing and classification to separate particles into size-based fractions using screens or hydrocyclones, concentration to enrich the ore through methods like froth flotation, gravimetric separation, and magnetic or electrostatic separation, and finally dewatering through thickening, filtration, or drying. Each stage relies on massive mechanical equipment operating under harsh, variable, load-intensive conditions. Components inevitably degrade over time, and failure — the termination of an item’s ability to perform its required function — can expose personnel to hazards while halting production across the entire plant.
What the researchers found is a field heavily skewed toward one narrow slice of the problem. Fault diagnostics — identifying and isolating which component is failing — received the highest research attention, accounting for 49.2 percent of task assignments across the reviewed studies. Fault detection, spotting abnormal conditions before significant damage occurs, came second at 35.6 percent. Prognostics — actually predicting the future state of the system and its remaining useful life — represented just 15.3 percent of the literature. This imbalance means that while substantial progress exists in identifying and classifying faults, predicting RUL under varying loads, heterogeneous material properties, and environmental conditions remains a significant unsolved challenge for mining processing machines in real-world settings.
The sensing landscape is equally lopsided. Vibration signals account for roughly half of all data modalities used across the reviewed studies, dominating the field to a degree that limits what can be observed. Acoustic, torque, and temperature measurements appear far less frequently, while force, belt tension, airflow, and other modalities collectively form the remainder. Multimodal sensing — combining vibration with electrical signatures, thermal imaging, visual, or 3D data — remains uncommon in mining processing machines. This narrow focus restricts the capability of fault detection and prediction for failures that manifest in electrical, hydraulic, pneumatic, control, or software subsystems, which are increasingly relevant in modern mining equipment yet barely represented in the word cloud of investigated assets.
On the modelling side, supervised machine learning and deep learning methods dominate, particularly for detection and diagnostics. Support vector machines, convolutional neural networks, long short-term memory networks, and deep belief networks appear most frequently. These algorithms depend heavily on high-quality labelled datasets — which are scarce in reality given the confidentiality constraints, safety-critical environments, and the predominance of scheduled preventive maintenance practices in mining operations. Half of the reviewed studies relied on public datasets, a quarter on laboratory test rigs, and only around 16 percent used real industrial data. Public datasets most commonly employed are general rotating-machinery bearing collections rather than mining-specific ones, raising serious questions about whether models trained under controlled conditions with stable loads and well-defined faults can generalize to the variable, multi-component load fluctuations, contaminated signals, and unknown disturbances of in-situ mining environments.
Perhaps the most striking finding concerns health management — the decision-making process that translates estimated system health and expected future operating conditions into maintenance and operational actions. Only 27 percent of the reviewed papers explicitly link PHM outputs to decision-making such as maintenance scheduling optimization, spare parts logistics, resource allocation, or cost-risk trade-off analysis. The majority focus on monitoring, detection, or diagnosis without extending their contributions to actionable operations and maintenance planning. This gap means that even where RUL estimates are produced, they rarely inform when to repair, replace, shut down, or reallocate resources — the very decisions that determine system availability and operational continuity in a processing plant.
The review also uncovered a clear methodological progression in the literature, from feature engineering with classical classifiers through convolutional and recurrent architectures to attention mechanisms, graph networks, transformers, and meta-learning. Yet comparative baselines are typically drawn from the same or adjacent model families, and well-tuned classical alternatives are rarely retained as competitive benchmarks. Where classical methods are evaluated directly against learned representations on identical data, they remain competitive — random forest was reported as the strongest performer among models in at least one conveyor bearing study. The literature 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.
Looking forward, the researchers identify four research directions ordered by expected practical impact and feasibility. Reducing dependence on labelled data appears the most immediately tractable, as self-supervised learning, meta-learning, and generative approaches are already represented in the 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, demanding closer collaboration with industry. Moving beyond component-centric monitoring depends on multi-sensor infrastructure not yet widely deployed in processing plants. Exploiting emerging technologies — large language models, agentic AI, augmented reality interfaces, LiDAR, and thermal cameras — is the least evidenced direction, with only one of the 74 reviewed studies applying a large language model and none addressing agentic AI or augmented reality.
The stakes could hardly be higher. Mining is increasingly receptive to AI-driven decision support, and the convergence of digital technologies, data analytics, and machine learning has positioned data-driven approaches as a central enabler of intelligent asset management. But this review makes clear that the field remains fragmented and component-centric, with most studies focusing on isolated elements like bearings, gearboxes, or motors rather than the interconnected systems that actually determine plant performance. Until PHM research in mining processing machinery matures from laboratory benchmarks into system-level health management — where prognostics information is translated into maintenance and operational decision-making — the industry will continue leaving availability, safety, and cost improvements on the table. The machines are ready to talk. The question is whether the field is ready to listen at scale.
Subject of Research: Artificial Intelligence
Article Title: Data-driven prognostics and health management for mining processing machinery: a systematic review
Article References: Prabhu, S., Kour, R., Dersin, P., & Karim, R. (2026). Data-driven prognostics and health management for mining processing machinery: a systematic review. Frontiers in Manufacturing Technology, 6, Article 1936330. https://doi.org/10.3389/fmtec.2026.1936330
Image Credits: AI Generated
DOI: 10.3389/fmtec.2026.1936330
Keywords: artificial intelligence, machine learning, deep learning, predictive maintenance, mining, prognostics and health management, industrial IoT, systematic review, Data-driven, prognostics, health, management
News Source: Blake Davidson. (October 15, 2026). AI Can Predict When Mining Machines Will Fail — But Almost Nobody Uses It Yet. Scienmag.
