AQUADA-DTEC: Curriculum-Learning-Based Thermographic Blade Anomaly Detection in Normal Wind Turbine Operation

Blade damage inspection without stopping the normal operation of wind turbines has significant economic value for wind farm operation and maintenance. This study proposes AQUADA-DTEC, a curriculum-learning-based AI approach that can accurately and robustly detect blade anomalies without stopping the normal operation of wind turbines in field. AQUADA-DTEC introduces a Swin-Transformer-based industrial image anomaly detection model to detect blade structural anomaly, and it employs a curriculum learning strategy that contains three meticulously designed curriculums to train the model, thus enabling it to learn in an easy-to-hard way and to detect small anomalies with low thermal contrast. To train and evaluate the approach, we collected a large-scale wind turbine blade dataset that contains over 17 000 thermal images of wind turbine blades. Experimental results show that AQUADA-DTEC outperforms state-of-the-art methods in thermographic blade structural anomaly detection by 6.6% in AUROC and 4.3% in average precision. In addition, the proposed curriculum learning strategy significantly enhances the model’s ability to detect small anomalies with low thermal contrast and completely elevates the applicability level of the model, making it practically useful in real-world applications. Overall, AQUADA-DTEC represents an important step toward practical, in-field thermographic inspection of blade structural damage.