Fault detection and diagnosis (FDD) are critical for maintaining efficiency and operational stability of comminution systems. However, conventional methods struggle to capture their complex dynamic behaviour, while data-driven approaches are constrained by limited labelled fault data and the need for interpretable diagnostic models. Progress is further hindered by the scarcity of publicly available industrial datasets. This study presents an explainable FDD framework that integrates unsupervised autoencoder (AE)-based anomaly detection with variance-based global sensitivity analysis (GSA) for quantitative fault diagnosis. A simulated comminution control system was developed to enable controlled validation under realistic operating variability. Multiple AE architectures were trained with hyperparameters optimised using chaotic particle swarm optimisation and evaluated using statistical and reconstruction-based metrics combined with multi-criteria decision analysis. The sparse AE achieved the best performance, with an MSE of 5.6 × 10−5, F1-score of 0.9930, and accuracy of 0.986 in detecting faults in P80 and P20. To diagnose detected faults, Sobol’s variance-based GSA was applied to quantify both the main and interaction effects of operational variables on particle size distribution. The results identify circuit feed rate, ball mill critical speed, and the pulp solids fraction supplied to the hydrocyclones as dominant drivers of faults associated with product coarsening, whereas circuit feed rate and ball mill critical speed primarily govern ultrafine particle generation. By integrating deep learning with explainable sensitivity analysis, this study advances transparent and quantitative diagnosis of complex mineral processing systems.
Freddy A. Lucay (Fri,) studied this question.