Generating accurate damage bounding boxes can significantly improve the efficiency of WTBs (wind turbine blades) surface damage detection. A great progress has been made in the development of damage detection methods for WTBs, but there are still some open problems such as how to sufficiently extract damage features and how to efficiently match between prediction box and target box. This paper aims to offer some solutions to the two questions. In doing so, an attention-based deep learning framework with multiple attributes loss function is first proposed for surface damage detection of WTBs. In this method, YOLOv5 (You Only Look Once) is used as the detection framework. CBAM is embedded into the backbone to direct its attention to damage features. Then a multiple attributes intersection over union (MIoU) loss function is designed for bounding box regression, to better fine-tune and adjust the position of the predicted bounding box so that is closer to the ground truth bounding box. The proposed method combines CBAM with YOLOv5 to improve the network's ability to extract damage features and provide more accurate information for the generation of bounding boxes. In addition, the proposed method avoids generating bounding boxes that are inconsistent with the absolute size of the target box by designing the MIoU loss function. Experimental results on real data, collected from wind farms, show that the performance of the proposed method outperforms the existing state-of-the-art methods.

Wind turbine blade,Damage detection,Loss function,Attention mechanism,YOLO