Abstract: This paper proposes a novel PMSM fault detection scheme to address the problem of similar fault signals between windings and turns of permanent magnet synchronous motors, which makes it difficult to identify the degree of the fault. The scheme includes an improved data processing method of fully ensemble empirical mode decomposition (ICEEMDAN) permutation entropy (PE) combined with wavelet denoising, as well as a fault detection model based on residual network (ResNet) and channel and spatial attention (CPSAM). Using the data denoised by ICEEMDAN-PE combined with wavelet thresholding as the first branch input can effectively reduce noise interference, highlight data features, and preserve key signal information; The second branch inputs raw data to ensure information integrity and improve feature detail changes. The two branches extract features of different scales through a multi-scale feature extraction module, alleviate the gradient explosion problem through residual structure, and finally fuse with each other through CPSAM attention extraction channel and spatial information weight matrix to distinguish the importance of features. Finally, they are classified by a classifier. The measured data shows that the model has strong diagnostic and anti noise capabilities, with a diagnostic accuracy of 100% in ideal environments without noise, and an accuracy of 97.75% in high noise environments with 5dB.