Abstract: Traditional permanent magnet synchronous motor fault diagnosis methods make motor fault diagnosis complicated due to the need for detailed motor parameters, in motor intelligent fault diagnosis, convolutional neural networks and recurrent neural networks can limit the parallel ability of the model and the dependence of signal features, and the limitations of the use of intrusive sensors and the cost of specific scenarios can also limit the fault diagnosis scheme. To address the above problems, an intelligent diagnosis method for permanent magnet demagnetisation and rotor dynamic eccentricity faults in permanent magnet synchronous motors is proposed in this paper by combining time-frequency diagrams of audio signals with DeiT network, which first denoises the audio signals and generates the time-frequency diagrams by using wavelet threshold denoising method and short-time Fourier transform, and then inputs them into DeiT for classification to realise the fault diagnosis. DeiT introduces the structure of knowledge distillation, which can be used in less training cycle and smaller data set to achieve the fault diagnosis. DeiT introduces a knowledge distillation structure, which can obtain better performance on fewer training cycles and smaller datasets. Experimentally, it is proved that the proposed method can effectively improve the model training efficiency and diagnostic accuracy, and at the same time, it has good anti-noise ability and better diagnostic performance compared with other mainstream diagnostic methods.