Valves are common devices in nuclear power plants, primarily used for controlling fluid transport. In nuclear power plants, some valves operate in harsh environments with high temperature, high pressure, and high radiation. Prolonged operation in such conditions can lead to performance degradation or loss of certain functions. In order to better detect valve failures in advance, this study focuses on the early fault diagnosis techniques for electric gate valves. An experimental setup was constructed to simulate normal operation, three-phase voltage imbalance, and packing damage in electric gate valves. The XGBoost method is utilized to diagnose the collected fault data, and to address the issue of poor fault diagnosis performance caused by imbalanced data, the Generative Adversarial Network (GAN) method is employed to generate sample data, thereby increasing the number of minority class samples and resolving the sample imbalance problem. Experimental results demonstrated that the proposed method effectively identifies fault states of electric gate valves and provides accurate fault diagnosis predictions under sample imbalance conditions.