Abstract: Aiming at the problem of model overfitting caused by scarce samples in rolling bearing datasets, this study proposes a two-stage fault diagnosis model framework （DCGAN-CNN） that integrates deep convolutional generative adversarial networks and convolutional neural networks. First, a DCGAN structure with a gradient penalty mechanism is constructed to generate one-dimensional vibration signals consistent with the real data distribution via adversarial training. Then, a multi-scale feature fusion CNN classifier is designed to improve the generalization ability of the model using the generated data. Experimental results show that the diagnostic accuracy of the proposed model reaches 97.9% under insufficient data samples, which is 1.2% higher than that of the traditional overlapping sampling method. This verifies the advantages of the model framework in the field of rolling bearing fault diagnosis.