Unsupervised anomaly detection for wind turbine blade damage: a background-agnostic approach

Wind turbine blades exhibit complex and diverse damage types. For damage types with a low occurrence probability, the scarcity of samples makes it challenging to train large-scale supervised models. To address this issue, we propose an unsupervised anomaly detection method for wind turbine blade damage. First, building on the PaDiM anomaly detection algorithm, a slicing operation is introduced to preprocess input images, enabling the model to handle higher-resolution images. Second, ConvNeXt with integrated BlurPool convolution layers is employed as the backbone network, which significantly mitigates the adverse effects of uncontrolled outdoor environments. Then, by leveraging the normal-state data distribution learned during training together with the Mahalanobis distance computed during testing, an anomaly prediction map is generated; an anomaly localization optimization module is further applied to accurately identify damage regions on the blades. Finally, comprehensive ablation and comparative experiments are conducted. Experimental results demonstrate that the proposed method improves anomaly classification accuracy and anomaly localization precision by 2.85% and 9.95%, respectively, thereby fully validating its effectiveness.