A new algorithm of wind turbine blade surface damage recognition based on deep learning

The identification of surface imperfections on wind turbine blades is of significant importance for ensuring the regular operation of generator sets and achieving timely repair of damaged blades. To solve the problem, this paper first constructs a dataset of surface damages on wind turbine blades, collected by unmanned aerial vehicles (UAV), encompassing various types of defects. Secondly, an improved YOLOv5 algorithm for identifying defects on the surface of blades is proposed, which introduces deformable convolutional neural networks and Transformer modules into the original network to optimize its performance and augmenting the adaptability of feature maps in shape characteristics. Experimental results demonstrate that our model, when tested on this dataset, achieves an improvement of 0.9% in the Precision metric, an increase of 2.6% in the Recall metric, an increase of 0.4% in the mAP@50(%) metric, and an increase of 3.0% in the mAP@50:95(%) metric compared to the original model. Both qualitative and quantitative experiments indicate that our model can effectively identify surface defects on wind turbine blades, providing a reference for automatic detection methods of wind blade defects.