A Comparative Multi-Model Framework with Robustness Validation and SHAP Explanation for Belt Conveyor Fault Diagnosis

Belt conveyors are essential to continuous ore and bulk-material transport, yet the diagnostic signatures of slippage, belt breakage, deviation and overheating tend to appear as coupled changes across multiple variables rather than isolated in a single signal. This study presents a multi-model and explainable diagnostic framework using belt speed, motor temperature, motor current, drum temperature and belt tension. A balanced dataset containing 1500 observations from five operating states was analyzed under a unified protocol. Eleven classifiers were compared using identical stratified partitioning and cross-validation, and a perturbation-based validation set was introduced to test whether near-ceiling baseline scores would hold under sensor noise and boundary-state drift. The measurements were preserved without alteration, and the perturbation set was treated explicitly as a sensitivity analysis. The baseline test confirmed strong class separability; the robustness-validation setting, in contrast, gave a more conservative picture of model behavior and revealed differences invisible in the ordinary hold-out split. Logistic regression reached accuracy = 0.800, macro-F1 = 0.798 and weighted AUC = 0.946 under perturbation. SHAP analysis showed that motor current, belt tension, motor temperature and belt speed were the dominant contributors, while drum temperature played a much weaker role. The results suggest that conveyor fault diagnosis should be reported with model comparison, robustness analysis and feature-level explanation together, particularly when ordinary ROC curves approach one and the practical value of a model must be weighed against uncertainty.