From Detection to Prediction: A Systematic Review of Intelligent Fault Detection in Industrial Conveyor Belt Systems

Industrial conveyor belt systems are critical infrastructure where unexpected failures cause major downtime, safety hazards, and economic losses, motivating growing research on AI-based fault detection. However, previous reviews have either focused on a single sensing modality, particularly computer vision, or considered conveyor systems within broader mining applications, without jointly assessing methodological quality, cross-modal sensing, deployment readiness, and the technical capabilities required for prognosis. Following PRISMA 2020 guidelines, this review analyzes 131 empirical studies on intelligent fault detection in industrial conveyor belt systems published between 2010 and 2026, with particular attention to sensing technologies, AI-based analytical methods, and deployment feasibility. The analysis considers the main fault types reported in the literature and the sensors and algorithms used for their detection. Deep learning was the dominant analytical approach (55.0% of studies), with CNN-based architectures such as YOLO, ResNet, and U-Net widely used for vision-based applications, while LSTM and GRU models were more commonly applied to acoustic and vibration signals. Vision/camera sensing accounted for 40.5% of studies, and mining represented 74% of applications. Despite substantial progress in automated fault detection, most studies remain focused on identifying existing faults rather than anticipating future degradation. Remaining useful life was estimated in only three studies (2.3%) and degradation trends in seven (5.3%); physics-informed machine learning and cross-site domain adaptation were absent, and principled multi-sensor fusion, uncertainty quantification, explainability, and digital-twin integration each appeared in no more than six studies. Predictive uncertainty was quantified in only one study (0.8%), and its calibration was not demonstrated. Future advances should therefore place greater emphasis on prognosis, cross-site generalization, uncertainty-aware inference, explainable models, and principled sensor fusion to support more reliable and transferable predictive maintenance systems.