关键词: 永磁同步电机, 故障诊断, 多尺度卷积, 动态残差连接, 信号融合, 损失函数, 注意力机制

Abstract: Aiming at the issues of insufficient feature extraction,difficulty in capturing temporal dependencies,and class imbalance in multiple fault scenarios,on the basis of 1D Convolutional Neural Network (1D-CNN),a fault diagnosis method for permanent magnet synchronous motors based on multi-signal feature fusion mechanism was proposed. Firstly,an efficient feature extraction module for vibration and temperature signals was constructed through multi-scale convolution and dynamic residual connections to capture key information from different frequency bands and enhance the integrity of feature extraction. Then,the spatiotemporal gated recurrent unit and bidirectional attention mechanisms were introduced,fully exploring the multi-level temporal dependencies in current signals,and enhancing the ability to identify complex fault features. Finally,the contributions of various samples were effectively balanced by utilizing multi-signal feature fusion and adaptive multi-objective optimization loss function,and the feature space distribution was optimized. The experimental results showed that the proposed improved method had higher fault diagnosis accuracy and better adaptability,compared to the original 1D-CNN at different speeds and loads. The detection accuracy,recall rate,and mAP had been improved by 2.3 %,1.9 % and 2.3 %,effectively solving the problem of insufficient accuracy and robustness in identifying multiple fault types.

Key words: permanent magnet synchronous motor, fault diagnosis, multi-scale convolution, dynamic residual connection, signal fusion, loss function, attention mechanism