Abstract: In this study, a generative model-auxiliary classifier generative adversarial network (ACGAN) for bearing fault diagnosis is proposed, as the traditional discriminative bearing fault diagnosis algorithm relies on artificial feature extraction and has poor diagnostic effects in complex working conditions. First, the bearing vibration signal is converted into a two-dimensional frequency-domain feature grayscale image using a fast Fourier transform. Subsequently, the convolution neural network is designed as the main structure of the model, and batch normalization and LeakyReLU activation functions are added to mitigate the gradient disappearance issue. Second, the self-attention (SA) mechanism is introduced to correlate the features far away from each other in the data. A new SA-ACGAN model is designed to realize the effective learning of the original data distribution features in the multi-classification scene. Finally, the model is applied to the motor bearing for comparison and verification. The results show that the fault diagnostic accuracy of the proposed method is as high as 99.7%, with considerable robustness and generalization.