基于Faster R-CNN的风机叶片缺陷图像多点检测系统设计  

Design of fan blade defect image multi-point detection system based on Faster R-CNN

作　　者：曲忠侃 李学峰 李凤俊 谢宝瑜 马长啸 QU Zhongkan;LI Xuefeng;LI Fengjun;XIE Baoyu;MA Changxiao(New Energy Business Department of Guodian Power Development Co.,Ltd.,Beijing 100101,China;Guodian Power Inner Mongolia New Energy Development Co.,Ltd.,Huhhot 010020,China;Guodian Power Hunan New Energy Development Co.,Ltd.,Changsha 410016,China)

机构地区：[1]国电电力发展股份有限公司新能源事业部,北京100101 [2]国电电力内蒙古新能源开发有限公司,内蒙古呼和浩特010020 [3]国电电力湖南新能源开发有限公司,湖南长沙410016

出　　处：《电子设计工程》2021年第4期57-61,共5页Electronic Design Engineering

基　　金：北京自然科学基金(51766520)。

摘　　要：随着新能源风力发电生产规模逐渐扩大,对生产效益精细化的要求不断提升。针对现有电机叶片检查方式造成经济成本高、检查效率低、高空坠落安全隐患等问题,基于深度学习理论,提出了基于Faster R-CNN的胶衣脱落缺陷图像多点检测系统设计。通过对风机叶片的缺陷图片进行数据增广,丰富缺陷数据,以此来增加模型的泛化能力,同时引入GIoU、DCNv2。测试结果表明,所提算法得到的模型平均类别精度(mAP)达到92.6%,可显著提高胶衣检测的精度。As the scale of new energy wind power production gradually expands,the requirements for refined production efficiency continue to increase.Aiming at the problems of high economic cost,low inspection efficiency,and high-altitude fall safety hazards caused by the existing motor blade inspection methods,based on deep learning theory,a detection method for fan gel coat loss based on Faster R-CNN was proposed.By augmenting the defect pictures of the fan gel coat falling off and enriching the defect data,the generalization ability of the model is increased.Test results show that the mean Average Precision(mAP)of the model obtained by the proposed algorithm reaches 92.6%,which can significantly improve the accuracy of gel coat detection.

关 键 词：风机叶片 无人机图像 深度学习 缺陷检测 Faster R-CNN 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]