With the rapid development of modern industry and artificial intelligence technology, fault diagnosis technology has become more automated and intelligent. The deep learning based fault diagnosis model has achieved significant advantages over the traditional fault diagnosis method. However, a problem has arisen when deep learning models are applied to actual industrial scenarios. In the actual process of industrial production, there are not enough fault data for deep learning, hence the accuracy will decrease because of overfitting. In this paper, aiming at the problem, the fault data generation based on deep learning is deeply studied. In this paper, the data source is the experimental data from the Fault Data Center of Case Western Reserve University. Aiming at the problem of small amount of fault data, a method of generating fault time-frequency spectrum based on improved conditional generative adversarial networks is proposed. CWGAN (Conditional Wasserstein Generative Adversarial Nets) learns the feature of time-frequency spectrum of rolling bearing fault, and generates time-frequency spectrum of corresponding fault categories according to the input categories. Experiments show that the diversity and fidelity of data generated by CWGAN is better than that of the original generative adversarial networks. The VGG-Net model is used to train the fault data enhanced by CWGAN. It is found that the data generated by CWGAN can effectively supplement the small amount of fault data, improve the training effect of the model and avoid over-fitting.