Abstract: Aiming at the problems of bearings working in complex environments， where fault data are difficult to obtain in large quantities and the serious imbalance between the ratio of normal data and fault data resulting in insufficient in-depth model training and low diagnostic accuracy， a bearing fault diagnosis method based on LSGAN-Swin Transformer is proposed. The least-squares generative adversarial network is utilized to expand the imbalanced or lack of bearing dataset， and the windowed self-attentive network is introduced for bearing fault state identification. The proposed method is validated by using two date sets， and compared with SGAN and WGAN respectively. It is demonstrated that LSGAN generates data training models with higher accuracy. The proposed Swin Transformer （Swin-T） model is compared with CNN， AlexNet and SqueezeNet under small sample conditions， and the accuracy is improved by 34.85%， 13.45%， and 12.95%， respectively. The classification effect of the model is evaluated by t-SNE visualization， and the results show that the LSGAN-Swin-T model can still meet the requirements in fault diagnosis better when the number of training samples is small， which provides a new idea for the research of bearing fault diagnosis under unbalanced data.