Research on wind power equipment fault diagnosis method based on small sample data

Targeting issues like small sample size, class imbalance in wind turbine fault diagnosis, and poor quality/unstable training of traditional GANs, this paper proposes a lightweight fault diagnosis method combining Curvature Regularized Wasserstein GAN with Gradient Penalty (CR-WGAN-GP) and Channel Attention Multiscale CNN (CAMS-CNN). Initially, Adaptive Wavelet Transform (AWT) extracts preliminary features from raw vibration signals. Then, CR-WGAN-GP, by incorporating curvature regularization into the WGAN-GP framework, generates high-quality fault data from small samples, suppressing mode collapse and enhancing GAN training stability. Subsequently, the CAMS-CNN classifier, leveraging its channel attention mechanism to focus on key features and using a multiscale convolutional structure to capture diverse fault information, achieves accurate and efficient fault classification on the CR-WGAN-GP augmented dataset. Experimental verification on the Case Western Reserve University bearing dataset shows that this method delivers high accuracy and strong generalization in small-sample fault diagnosis tasks.