In order to improve the recognition accuracy of single faults and compound faults for gear in different gearboxes,and overcome the difficulty of traditional fault feature extraction methods relying too much on empirical judgment, a semi-supervised convolutional generative adversarial network (SCGAN) model is proposed based on deep learning, which combines two deep neural network features, convolutional neural network (CNN) and generative adversarial network (GAN). This model uses two CNN networks as the generation network (G) and the discriminant network (D) of the GAN network, and improves its internal structure to realize the transformation of GAN from an unsupervised learning mechanism to a semi-supervised learning mechanism. The gear fault signals collected in the power transmission simulation test-bed are made into time-domain,frequency-domain and time-frequency sample sets,and the SCGAN model is constructed for fault diagnosis. By comparing the effects of several situations,including different network model,different sample set and different iteration times, the results show that the diagnostic accuracy of the SCGAN is significantly higher than that of CNN and recurrent neural network (RNN), and the convergence speed is faster.