As a critical component of wind turbines, the scarcity of fault data for the main bearing presents significant challenges for fault diagnosis. Generative adversarial network (GAN) presents a new approach to tackle fault diagnosis with small samples but still faces issues such as training instability. Therefore, this paper proposes a fault diagnosis method based on WSRGAN-GP and LSTM for time-series data. Time-series data are transformed into image data, and WSRGAN-GP is utilized for data augmentation. Fault diagnosis is achieved using an LSTM model. Firstly, a loss function is designed by incorporating Wasserstein distance and gradient penalty. Secondly, a generator is constructed using a residual block, and a GAN model with a self-attention mechanism is developed. Finally, an LSTM model is employed for fault diagnosis. Experimental results demonstrate that the proposed method effectively diagnoses faults in wind turbines under small sample conditions, validating the efficacy of the model.