Abstract: Timely diagnosis of potential faults in hydroelectric generating units is crucial for ensuring the safe and stable operation of power plants. In practical engineering applications, however, fault samples are scarce, categories are imbalanced, and signal-coupled noise is intense, making it difficult for traditional methods to achieve reliable diagnosis under small-sample conditions. In light of this, this paper proposes a novel fault diagnosis model for hydroelectric generating units that integrates data augmentation and ensemble learning by combining VMD, multi-channel CNN, and WGAN. First, Variational Mode Decomposition (VMD) is used to decompose the vibration signal into several Intrinsic Mode Function (IMF) components, thereby achieving feature decoupling. Next, a data augmentation model based on a multi-layer one-dimensional Convolutional Neural Network (CNN) and a Generative Adversarial Network (GAN) is constructed to generate new samples using the decoupled signals as the data source. Finally, an ensemble learning strategy is employed to perform fault classification. The model achieved an accuracy of over 99% on the test bench, with a frequency-domain correlation of 0.96 for the generated samples and an RMSE of < 0.02, outperforming traditional Generative Adversarial Networks (GANs). Empirical results show that the model proposed in this paper achieves an accuracy approximately 28% higher than conventional methods in imbalanced scenarios and maintains an accuracy of 85.56% even under complex background noise. This study provides a highly accurate and robust solution for fault diagnosis of hydropower units with limited data, which is of great value for ensuring the safe and stable operation of the units.