Real-Time Diagnosis of Stator Winding Faults in PMSM Based on Residual CNN With Channel Attention Mechanism

The fault detection and identification are essential for permanent magnet synchronous motors (PMSMs) to guarantee the reliable and continuous operation. As the primary energy conversion component, the stator winding integrity critically determines operational reliability. Various fault types exhibit distinct degradation patterns, requiring not only detection but precise identification for targeted maintenance. Existing data-driven fault diagnosis strategies are mostly limited to single-fault detection and offline operation. In this study, an online stator winding fault diagnosis method for PMSM is proposed based on a residual convolutional neural network with a squeezing excitation channel attention mechanism (SE-ResCNN). This method is capable of extracting both the global and local features by combining the residual branches with the CNN branches. Besides, a squeezing excitation attention mechanism is employed to realize the cross-branch channel fusion, thereby enhancing the diagnostic accuracy. Afterward, leveraging the RK3588 edge device and model conversion and migration techniques, the designed network enables efficient edge deployment while maintaining diagnostic accuracy and real-time processing speed. Finally, the proposed online diagnosis method is validated through an experimental platform of the PMSM drives.