兰州理工大学学报

兰州理工大学学报 ›› 2026, Vol. 52 ›› Issue (4): 52-58.doi: 10.13295/j.cnki.issn1673-5196.2026.04.006

• 机械工程与动力工程 • 上一篇    下一篇

风电叶片表面缺陷检测算法轻量化研究

魏泰*1, 刘宇航2, 薛文杰3, 陈学友1   

  1. 1.甘肃省特种设备检验检测研究院, 甘肃 兰州 730050;
    2.兰州理工大学 机电工程学院, 甘肃 兰州 730050;
    3.东营市海科瑞林化工有限公司, 山东 东营 257200
  • 收稿日期:2024-04-17 出版日期:2026-08-28 发布日期:2026-09-03
  • 通讯作者: 魏泰(1983-),男,甘肃庆阳人,高级工程师.Email:weitai0931@163.com
  • 基金资助:
    国家自然科学基金(51965034),兰州市科技计划项目(2023-3-98)

Research on lightweighting of wind turbine blade surface defect detection algorithm

WEI Tai1, LIU Yu-hang2, XUE Wen-jie3, CHEN Xue-you1   

  1. 1. Gansu Province Special Equipment Inspection and Testing institute, Lanzhou 730050, China;
    2. School of Mechanical and Electrical Engineering, Lanzhou University of Technology, Lanzhou 730050, China;
    3. Dongying Haike Ruilin Chemical Industry Co. Ltd., Dongying 257200, China
  • Received:2024-04-17 Online:2026-08-28 Published:2026-09-03

PDF

15

摘要： 基于人工和传统自动化检测算法在风电叶片表面裂纹和剥落腐蚀缺陷检测任务中存在精度低和效率差等问题,提出了基于优化YOLOv5s的风电叶片表面缺陷检测轻量化算法.首先,构建不同自然环境条件下叶片表面缺陷数据集;其次,在YOLOv5s中引入GhostNet的Ghost模块和Ghost BotleNeck结构;最后,针对真实框与预测框不一致的问题,引入损失函数SIoU,通过聚焦普通质量锚框的预测回归提升叶片表面缺陷识别准确率和定位精度.优化后模型的参数量、计算量和权重文件分别为YOLOv5s的52.6%、51.6%和55.2%,叶片表面缺陷识别平均精度均值为88.3%.与Faster-RCNN、SSD、YOLOv4、YOLOv5s和YOLOv8s模型相比,平均精度均值分别提高了34.2%、36.5%、24.5%、1.9%和1.8%.结果表明,YOLOv5s优化后成功防止漏检和误检,提升了目标定位精度,更适用于检测复杂工况下叶片表面裂纹和剥落腐蚀缺陷.

关键词: 风电叶片, 图像识别, YOLOv5s, 目标检测, 轻量化算法

Abstract: Given the low accuracy and poor efficiency of manual and traditional automated detection algorithms for identifying surface cracks and exfoliation corrosion defects in wind turbine blades, a lightweight algorithm based on optimized YOLOv5s for surface defect detection in wind turbine blades has been proposed. This algorithm involves the creation of a dataset encompassing blade surface defects observed under various natural conditions. Ghost modules and Ghost Bottleneck structures from the GhostNet are integrated into YOLOv5s to enhance its capabilities. Soft Intersection over Union (SIoU) is introduced to improve prediction regression, with a focus on standard quality anchor boxes, bridging the gap between actual and predicted bounding boxes. This enhancement significantly boosts the accuracy of blade surface defect recognition and improves localization precision. Furthermore, the optimized model reduces parameters, computational load, and the weight file size of YOLOv5s by 52.6%, 51.6%, and 55.2%, respectively. The algorithm achieves an average precision of 88.3% in recognizing blade surface defects. Comparative analysis against Faster-RCNN, SSD, YOLOv4, YOLOv5s, and YOLOv8s algorithms reveals improvements in average precision by 34.2%, 36.5%, 24.5%, 1.9%, and 1.8%, respectively. The experimental results indicate that the algorithm successfully prevents both missed detections and false positives after optimization, while also improving the accuracy of target localization, making it more suitable for detecting surface cracks and exfoliation corrosion defects on blades under complex working conditions.

Key words: wind turbine blade, image recognition, YOLOv5s, object detection, lightweight algorithm

中图分类号: