Wind Turbine Blade Defect Detection Method based on Improved YOLOv11

Abstract

Aiming at the problems of low efficiency of manual inspection, easy missed detection of small target defects, and severe interference from complex backgrounds in wind turbine blade defect detection for wind power operation and maintenance scenarios, an improved YOLOv11 model integrating feature enhancement and dynamic sample optimization is proposed. Firstly, the local perception unit of images is constructed through superpixel texture analysis, and the texture entropy of each region is quantified combined with the Gray Level Co-occurrence Matrix (GLCM) to realize the adaptive localization of potential defect regions of blades. Secondly, the Coordinate Attention (CA) module is embedded in the backbone network to dynamically generate spatial weight masks, which strengthens the network's focusing ability on high-entropy defect regions and suppresses the artifact interference caused by complex blade textures and field environments. Meanwhile, the weighted Bidirectional Feature Pyramid Network (BiFPN) is designed to replace the original multi-scale fusion structure, which optimizes the feature aggregation process and enhances the edge consistency representation of small-sized defects. In addition, the detector head structure is optimized and the Intersection over Union (IoU) loss function is introduced. Combined with the multi-category focal loss, the penalty coefficient for easy and hard samples is dynamically adjusted according to the distribution of training samples, which alleviates the model bias caused by the long-tail distribution of industrial data and category imbalance. Experimental results show that compared with the original model and mainstream detection models such as Faster-RCNN, SSD and YOLO series, this method achieves better performance in detection speed and accuracy, and effectively improves the detection performance of the system in complex scenarios such as noise interference and low illumination.