Fault diagnosis of wind turbine pitch bearings via transfer learning and an improved residual network

Aiming to address the challenges of difficult fault signal extraction, limited sample availability, and reduced recognition accuracy due to noise interference in wind turbine pitch bearings under low-speed, heavy-load conditions, this paper proposes an enhanced deep residual network (PRSN) approach. The method integrates Mel-frequency cepstral coefficient (MFCC) feature extraction, multiscale feature analysis, soft-threshold denoising, and transfer learning. The model incorporates a pyramid split attention (PSA) mechanism to extract spatial and channel-wise multiscale features, combined with a DRSN-based soft-threshold denoising module to suppress irrelevant signals and enhance noise resilience. Additionally, a transfer learning strategy is employed to retain and fine-tune the weights of the pre-trained ResNet50 model, thereby improving recognition of early-stage impacts and complex fault patterns. Experimental results demonstrate that the proposed method achieves a fault recognition accuracy of 97% in noiseless conditions and an average accuracy of 80% under multi-noise environments, significantly outperforming traditional approaches. Furthermore, the integration of LIME and t-SNE for visualization elucidates the critical fault regions identified by the model, enhancing interpretability and offering a robust, efficient solution for intelligent fault diagnosis in wind turbine pitch bearings.