Data augmentation methods have recently been widely used in fault classification under limited data. However, the traditional data augmentation model often suffers from the problems of gradient explosion and gradient disappearance in training, which limits its application to rolling bearing fault classification to some extent. A new modeling framework is proposed in this paper to address the aforementioned issues. The model first converts the original one-dimensional vibration data of rolling bearings into frequency-domain signals by Fourier transform and then does sample augmentation of the small amount of fault data by using conditional variational auto-encoding generative adversarial network (CVAE-GAN), and then finally trains convolutional neural network with a first wide convolutional layer by using the generative fault data and the original fault data together. The proposed method is validated using a publicly available dataset from the University of Paderborn Laboratory. The experimental results show that the CVAE-GAN method significantly improves fault diagnosis accuracy for rolling bearings under limited data.