关键词: 永磁同步电机, 在线故障诊断, 高分辨率时频增强, 跨模态多尺度子带对齐, 门控多头注意力, 结构化判别

Abstract: In response to the problem of decreased recognition accuracy of fault feature details in permanent magnet synchronous motors under complex operating conditions,a novel fault diagnosis method was proposed by improving the Transformer architecture and integrating cross-modal vibration and current signals. First,a high-resolution time-frequency enhancement mechanism was constructed to improve the time-frequency representation accuracy and modal discriminability of vibration and current signals. Then,a cross-modal multi-scale sub-band alignment mechanism was introduced,and strengthened the coupling consistency and alignment capability between multimodal features by a phase-guided gated multi-head attention strategy. Finally,a structural discriminative loss function was designed to optimize the feature distribution from multiple perspectives,including intra-class compactness,inter-class separability,and cross-modal consistency,thereby reinforcing the fault discrimination boundaries. Experimental results demonstrated that the three proposed enhancement strategies improved the model mAP by 3.3 %,3.6 %,and 2.3 %,respectively,significantly enhancing the stability and structural discriminative capability of feature extraction. The improved model achieved a precision,recall,mAP,and F1-score of 92.2 %,94.6 %,93.9 %,and 93.1 %,respectively. Moreover,the mAP decreased by only 0.5 % under complex noise interference,indicating that the proposed method effectively alleviated feature aliasing and modal mismatch and provided reliable technical support for online intelligent fault diagnosis of permanent magnet synchronous motors.

Key words: permanent magnet synchronous motor, online fault diagnosis, high-resolution time-frequency enhancement, cross-modal multi-scale sub-band alignment, gated attention, structural discriminative