A Spectral Normalized Auxiliary Classifier GAN for Bearing Fault Detection and Diagnosis

Generative Adversarial Networks (GANs), as an unsupervised data augmentation method, exhibit significant potential in addressing the issue of insufficient samples in rotating machinery fault diagnosis. However, existing unsupervised GANs face limitations, including the inability to generate multi-modal fault samples and challenges such as mode collapse and gradient vanishing. To overcome these shortcomings, this paper proposes an improved Auxiliary Classifier GAN (SN-ACGAN) model. First, an independent classifier is introduced to construct a new ACGAN framework, enhancing compatibility between classification and discrimination. Second, the Wasserstein distance is incorporated into the loss function to effectively address mode collapse and gradient vanishing issues. Finally, spectral normalization is employed as a replacement for weight clipping, improving the stability of the discriminator's weight parameters. SN-ACGAN is applied to fault diagnosis tasks for bearings and gears, and experimental results demonstrate that, compared to existing methods, SN-ACGAN generates higher-quality mul-ti-modal fault samples more efficiently, while achieving superior accuracy and stability. These findings indicate that SN-ACGAN has promising applications in the field of rotating machinery fault diagnosis.