Objective: To improve the accuracy and comprehensiveness of permanent magnet motor stator winding fault diagnosis. Methods: A fault diagnosis model of permanent magnet motor stator winding based on stack autoencoder （SAE） was proposed, and a neural network composed of SAE and Softmax classifier was used to train the network with fault sample data. The simulated annealing particle swarm optimization （SAPSO） algorithm was used to optimize the connection weight and bias of the network, and determined the optimal network structure. Results: The network had been used to realize the fault diagnosis of inter-turn short-circuit, inter-phase short-circuit, inter-phase insulation reduction, and poor contact of the terminals of the permanent magnet motor stator windings. Compared with wavelet analysis +Softmax, spectrum analysis +Softmax and SAE+Softmax, the diagnostic accuracy of this method was the highest, and the diagnostic rate was 99.40%. Conclusion: The optimized SAE+Softmax fault diagnosis model has good robustness and is less affected by motor speed and load changes, which can improve the accuracy of permanent magnet motor stator winding fault diagnosis.