A hybrid non-destructive testing strategy for wind turbine blade crack prediction utilising block-designed Transformer architecture

With the maturation of wind turbine (WT) supervisory control and data acquisition (SCADA) systems and the rapid development of deep learning, big data-driven AI strategies can be applied to the prognostics and health management of WTs, laying the foundation for non-destructive testing. This study proposes a novel hybrid strategy, MIT-Block, for predicting wind turbine blade (WTB) crack occurrence within a 7-day horizon. The method employs a block design based on the Transformer model and feature separation input for processing SCADA data, and uses mutual information (MI) for feature selection. An improved minority oversampling technique is adopted to address the class imbalance between normal and abnormal data during WT operation. The proposed strategy achieves an F1-score of 98.4% on similar types of turbines, demonstrating excellent generalisability across different turbine types. Ablation experiments verify the effectiveness of the feature selection method. This strategy provides a high-performance, low-cost, real-time solution for WT maintenance.