The monitoring data of permanent magnet synchronous motor (PMSM) exhibit complexities such as non-smoothness, non-linearity, multi-source heterogeneity and low value density. These characteristics make it challenging to accurately model the type and extent of motor faults using simulation data. The serious imbalance between normal and faulty data samples leads to problems such as overfitting and low accuracy in the training of fault diagnosis models. In this paper, an improved auxiliary classification generation adversarial network (ACGAN) is proposed to study the expansion of real fault data for PMSM by learning the distribution characteristics of the original samples, while the generated fault dataset provides a data base for the next fault diagnosis and health assessment. Firstly, to address the problems of poor convergence and the tendency for gradients to disappear or explode in ACGAN networks, the Wasserstein distance is used to constrain the reconstruction loss of the generated data, and the gradient penalty is used instead of weight clipping to optimize the model and mitigate model training instability. Secondly, to analyze the change relationship between data and the historical change pattern, recurrent neural network is introduced in the generator to improve the quality of the generated data. Finally, the effectiveness of four data expansion methods, ROS, SMOTE, ADASYN and improved ACGAN, is compared and analyzed in improving the performance of fault diagnosis models using fault data from PMSM inter-turn short circuits. Results show that the model trained using the improved ACGAN method is more stable, converges faster and produces expanded data of superior quality than those adopting other data expansion methods.