Few-Shot Fault Diagnosis of Wind Turbine Bearings Using HAGATCN-ProtoNet

Abstract

To address the challenges of limited fault samples and severe class imbalance in wind turbine bearing fault diagnosis, a few-shot fault diagnosis method based on a hierarchical attention gated temporal convolutional network and prototype learning (HAGATCN-ProtoNet) is proposed. The method constructs a unified framework that combines feature embedding with prototypical metric learning, enabling accurate fault identification under limited labeled samples and varying operating conditions. For feature extraction, a hierarchical attention gated temporal convolutional network is developed. A gating mechanism is employed to adaptively filter temporal information, while a parallel attention module enhances feature representation from both channel and temporal dimensions. In addition, a self-attention module is introduced to capture long-range temporal dependencies and improve global feature modeling. For classification, prototype learning is adopted to perform prototype based metric classification in the embedding space, reducing reliance on large scale labeled data. To address class imbalance, an episodic training strategy and a class-adaptive weighting scheme based on the effective number of samples are incorporated into the learning process. Experimental results on the CWRU bearing dataset and a real wind turbine bearing dataset demonstrate that the proposed method achieves superior performance under few-shot, cross-condition, and imbalanced scenarios, exhibiting strong robustness and generalization capability.