Interpretable Harmonic-Aware Dual-Branch Neural Network for Trustworthy Diagnosis of OCFs in DTP-PMSMs With Enhanced Disturbance Robustness

Boyuan Zheng, Bingtao Liu, Junyu Yan, Mi Tang, Pericle Zanchetta, Hao Yu

Abstract: Data-driven fault diagnosis models for power-electronic systems have gained significant attention in industrial applications, particularly for dual-three-phase permanent magnet synchronous motor (DTP-PMSM) systems with numerous switching devices. However, robustness, reliability, and interpretability remain key challenges for neural network-based methods. This study first analyzes the impact of open-circuit faults on DTP-PMSM phase currents. Based on this analysis, a harmonic-aware dual-branch neural network is proposed, combining time-domain and frequency-domain (harmonic) features via an attention-based feature-level fusion strategy. The proposed method achieves 99.90% overall accuracy, a 99.91% F1 score, and a LogLoss of 0.0077 with 271 frames per second under disturbance conditions, outperforming existing methods. Moreover, by leveraging fast-Fourier transform-extracted harmonics, the model effectively filters out white noise and outliers, improving robustness and interpretability, as supported by shapley additive explanations (SHAP) analysis. In addition, the proposed approach significantly reduces epistemic uncertainty in fault diagnosis, thereby enhancing diagnostic reliability.

External IDs:doi:10.1109/tpel.2025.3599747