Fault Detection and Diagnosis for Multi‐Faults of PMSM‐Drive Systems Using a Hybrid Machine Learning Method

ABSTRACT This paper presents a non‐invasive fault detection and diagnosis (FDD) methodology for permanent magnet synchronous machine (PMSM) drives, using low‐frequency phase current signals. Specifically, this work focuses on the detection and diagnosis of power electronics‐related inverter faults, which are a common source of system failures. The proposed framework introduces a pairwise feature fusion technique to enhance class separability and employs a three‐stage selection process to distil a compact, discriminative feature set from Clarke‐transformed current data. Diagnosis is performed by a hybrid machine learning model that ensembles the predictions of random forest, histogram‐based gradient boosting, and k‐nearest neighbours classifiers via a late‐fusion strategy. The performance of the proposed method is evaluated on a publicly available experimental dataset containing nine operational states (one healthy and eight distinct inverter faults). The proposed method achieves an overall accuracy of 93.3% and a macro F1‐score of 95.91%. The results demonstrate that the proposed approach can accurately diagnose multiple inverter faults without requiring high‐frequency data acquisition or additional sensors, offering a cost‐effective solution for enhancing the reliability of PMSM drives.