Interpretable fault diagnosis of insulated bearings in offshore wind turbines: the DRSI-former framework for complex operational regimes

Deep learning models can automatically learn deep-level features from massive data, completely eliminating the reliance on expert experience and laborious manual feature engineering. However, the complex and variable operating conditions of offshore wind turbines pose significant challenges to existing deep learning models, including low recognition accuracy for bearing electrical erosion faults, difficulties in effective feature extraction, and poor model interpretability. To address these challenges, this paper innovatively proposes a bearing electrical erosion fault diagnosis framework for offshore wind turbines, named the DRSI-Former framework. Leveraging its powerful temporal modeling and global feature capture capabilities, the DRSI-Former framework accurately separates and enhances fault features related to electrical erosion in wind turbine bearings. Additionally, an adaptive soft thresholding mechanism based on deep residual shrinkage networks is introduced to dynamically filter redundant noise while completely preserving sensitive fault features, thereby mitigating the issue of fault feature distortion. To validate the generalization performance of the DRSI-Former framework under variable load conditions, an experimental platform based on the FYKK16 doubly-fed wind turbine system was established for real-world data testing. The diagnostic performance of the DRSI-Former framework was systematically evaluated and compared with five other mainstream fault diagnosis models for rolling bearings. Experimental results demonstrate that the DRSI-Former framework achieves an accuracy of 99.9% in identifying electrical erosion faults in wind turbine bearings and maintains excellent performance under variable load conditions, fully proving its generalization capability, robustness, and technical superiority. Finally, by employing t-SNE for visualizing the fault features extracted by the framework, clear clustering separability of different fault modes in the feature space is revealed, further verifying the reliability, accuracy, and engineering credibility of the diagnostic results from an interpretability perspective.