DiffGAN: A Fault Diagnosis Data Augmentation Method for Hydropower Units Based on Adversarial Training and Diffusion Model

In industrial applications, fault diagnosis can effectively enable early warning. However, in practical production scenarios, fault data is difficult to obtain, and traditional fault diagnosis methods require a large quantity of high-quality data, severely limiting their diagnostic performance. To address this issue, this paper proposes a data augmentation method based on adversarial training and diffusion models—DiffGAN. DiffGAN alleviates the data imbalance problem by generating high-quality fault data, thereby improving fault diagnosis performance. It takes diffusion noise as the input for adversarial training, enhances the generator’s high-speed generation capability through adversarial training, optimizes the efficiency of the reverse generation process in traditional denoising diffusion models, and mitigates gradient vanishing and mode collapse issues in generative adversarial networks. Experiments were conducted on an actual hydropower unit dataset, and the results demonstrate that DiffGAN can efficiently generate fault data and significantly improve the sample distribution in imbalanced datasets. Furthermore, fault diagnosis experiments incorporating the ResNet model further verify the effectiveness of DiffGAN in enhancing diagnostic performance. This provides a novel and practical solution for data augmentation and fault diagnosis in hydropower units.