Abstract: At present, the rolling bearing fault diagnosis method in the nuclear power industry is mainly based on inspection and human analysis and judgment, and the effective sample data of bearing fault in the actual industrial process is relatively scarce, which limits the application of data-based intelligent models such as deep learning. Therefore, in the nuclear industry, the research on few-shot fault diagnosis is of great significance. In recent years, data augmentation algorithms have been widely used in small sample fault classification and achieved good results. However, the traditional data augmentation model often has problems such as gradient explosion and gradient disappearance during training, resulting in poor image quality and low fault diagnosis accuracy, which limits its application in rolling bearing fault classification to a certain extent. To address above issues, a new model framework was proposed. The model first converted the original onedimensional vibration data of the rolling bearing into a two-dimensional image through continuous wavelet transform (CWT). Then the variational autoencoder-generative adversarial network (VAE-GAN) was uesd to do sample enhancement. Finally, the generated image and the original image were used to jointly train a convolutional neural network (CNN) fault classifier. By introducing the CWT method, the one-dimensional time-frequency image is converted into a two-dimensional image, which can not only give full play to the powerful feature representation ability of convolutional neural networks, but also enable the classification network to better extract data features, and can also directly observe when the data is generated. The quality of generated images facilitates the improvement of network models. The introduction of VAE-GAN greatly improves the stability, diversity, and clarity of the images generated by the entire model compared with traditional data enhancement methods. And the using of the convolutional autoencoder makes it work better at image generation. The proposed method is validated using a publicly available dataset from the Laboratory of Case Western Reserve University. When using 80 original pictures to train the classifier, the classification accuracy rate of the classifier for 1 200 test samples after training is 81.2%. When using 800 generated pictures to train the classifier, the classification accuracy rate can reach 94.9%. When 80 original pictures and 800 generated pictures are mixed together to train the classifier, the classification accuracy of the classifier on 1 200 test samples can reach 99.0% after the training is completed. Experimental results show that the proposed model can effectively complete the rolling bearing fault diagnosis task in the case of few-shot problems. In addition, ablation experiments were carried out, and the results show that when any module in CWT, VAE-GAN, and CNN is removed, the performance of the model decreases to a certain extent, indicating that each proposed module is helpful to improve the fault diagnosis rate.