摘要
Abstract
In order to address the problems of insufficient detection accuracy for small target defects in wind turbine blade surface defect detection under complex environmental conditions,this paper proposes an efficient method for detecting defects in wind turbine blades,aimed to enhance detection accuracy and reduce computational load and model parameters.In the method,feature extraction in complex scenarios is improved by integrating a deformable attention mechanism;a novel NSC2f module is designed,which combines a normalized attention module with the SE attention mechanism within the C2f structure,effectively suppressing non-significant features and enhancing the detection capability for small targets.Additionally,the Slim Neck network module is employed to simplify the neck architecture and introduce a dynamic upsampling operator,thereby reducing interference from redundant information and capturing richer semantic features.Finally,the improved network is pruned by using magnitude-based layer adaptive sparsification pruning to further reduce both model parameters and computational load.Experimental results are demonstrated as follows:The WTB-YOLO model achieves a detection accuracy of 86.8%and an mAP@0.5 of 88.7%.Compared to the original YOLOv8n model,detection accuracy has been improved by 2.1%,mAP@0.5 increased by 2.6%,model parameters decreased by 46.7%,computational load reduced by 48.1%,and model size shrunk by 42.9%,achieving superior lightweight performance while maintaining improved detection accuracy.关键词
风力发电/风机叶片/小尺度缺陷检测/深度学习/YOLOv8/注意力机制/新能源/可再生能源Key words
wind power generation/wind turbine blades/small-scale defect detection/deep learning/YOLOv8n/attention mechanism/new energy/renewable energy分类
信息技术与安全科学
