In practical engineering applications, the short duration of rolling bearing fault states leads to imbalanced datasets, making it difficult to use deep learning algorithms for fault diagnosis. In this paper, a n infrared diagnosis method for rolling bearing faults based on the combination of the Wasserstein distance-based gradient penalty generative adversarial network (WGAN-GP) and a support vector machine-based convolutional neural network (CNN-SVM) is proposed. The imbalanced dataset is constructed from infrared thermal images, and WGAN-GP is used to augment the imbalanced data to achieve dataset balance, after which the CNN-SVM model is then applied to the dataset to extract deep features and complete fault classification. The experimental results show that the model combining WGAN-GP with CNN-SVM performs well under imbalanced datasets, with better fault diagnosis capability compared to other models, and reduces the time spent in the fault classification stage by more than 16.89%.