关键词: 时频特征融合, 时频联合建模, 时频权重, 跨尺度残差补偿

Abstract: Aiming at the problems of the existing algorithms, such as missing time domain features, lack of model robustness in high noise environment, and decline of fault classification accuracy in small sample data, this paper proposes an audio fault diagnosis model of permanent magnet synchronous motor based on time–frequency joint modeling and prior guided attention mechanism. First, the convolutional block attention mechanism is used to fuse the logmel spectrum and time-frequency domain features. Secondly, the time-frequency domain coupling relationship is constructed and the time-frequency weight matrix is calculated through the time-frequency dual-dimension memory network. Through the time-frequency weight matrix, the 8×8Patch layer of Vision Transformer is guided to extract key features, and the cross-scale residual compensation module is combined to avoid feature loss in the image block merging process. The experimental results show that under the condition of high noise and small sample data, the proposed algorithm model reaches 94.32% in the average F1 index, significantly surpassing other audio fault diagnosis algorithms compared, and significantly improving the diagnostic accuracy in the fault detection of demagnetization, rotor eccentricity, interturn short circuit and other categories.

Key words: Time-frequency feature fusion, Time–Frequency Joint Modeling, Time-frequency weight, Cross-scale residual compensation