Hierarchical multi-fault diagnosis of IPMSM drives using speed-normalized wavelet features and random forests

Abstract

Purpose Interior permanent magnet synchronous motor (IPMSM) drives may hold speed regulation even after electrical or mechanical faults have changed the stator-current signature. This paper aims to develop a ten-class diagnosis framework that separates healthy, single-fault and compound-fault operation under varying speed and load. Design/methodology/approach A closed-loop IPMSM simulation model generates healthy operation, three inter-turn short-circuit severities, three bearing-related severities and three compound fault cases. From stator-current windows, the features combine wavelet descriptors, speed-normalized frequency bands and electrical–mechanical interaction indicators. ReliefF selects informative variables from the training data, and a hierarchical random forest separates fault-family recognition from severity classification. Direct SVM, k-NN, RF and DQN benchmarks are compared. Findings Under grouped testing, the hierarchical RF with ReliefF gives the best ten-class performance and a clear fault-family path. On the unseen speed/load grid it stays close to the slightly more robust direct RF, and high load is difficult for all classifiers. Family-level recognition on that grid remains at 80.67% accuracy, so the first hierarchical stage generalizes better than the full ten-class decision. Repeating the grouped partition ten times preserves the ordering of the four feature-based classifiers. Noisy currents degrade all models markedly. Originality/value The study combines speed-normalized wavelet features, compound-fault indicators and hierarchical RF classification for IPMSM multi-fault diagnosis; a cautious DQN comparison indicates reinforcement learning is not necessarily superior here. Evaluation uses grouped testing, an unseen speed/load grid, repeated partitions and noise sensitivity.