Purpose <p>The paper proposes a gear fault diagnosis method to identify fault types with different loads and degradation degrees under small sample data.</p> Method <p>Firstly, the multi-source sparse dictionary feature extraction algorithm is used to extract the key features from the collected signals effectively. Then, the feature data is input into the sample enhancement module built based on the Wasserstein Generative Adversarial Network to generate new samples that retain the features of the original samples to solve the problem of insufficient data. Finally, a new cross-domain diagnosis module based on the domain adaptive migration network is proposed to improve the accuracy of fault data identification under different loads and degradation degrees.</p> Results and Validation <p>13 kinds of gear vibration data sets with different failure degrees were collected by using the self-developed gear comprehensive vibration test platform. By using the proposed method, the generated samples retain the features of the original samples completely, and the gear fault identification accuracy reaches 90%.</p> Conclusion <p>Compared with other algorithms, it is concluded that the proposed algorithm has better recognition accuracy and generalization ability.</p>