基于改进YOLOv5的航空发动机叶片表面缺陷检测方法研究  

Research on surface defect detection method of aero-engine blade based on improved YOLOv5

作　　者：刘宜轩 程志江[1] 吴动波 梁嘉炜 王辉[3] LIU Yixuan;CHENG Zhijiang;WU Dongbo;LIANG Jiawei;WANG Hui(College of Electrical Engineering,Xinjiang University,Urumqi 830017,China;Institute for Aero Engine,Tsinghua University,Beijing 100086,China;Department of Mechanical Engineering,Tsinghua University,Beijing 100086,China)

机构地区：[1]新疆大学电气工程学院,乌鲁木齐830017 [2]清华大学航空发动机研究院,北京100086 [3]清华大学机械工程系,北京100086

出　　处：《激光杂志》2023年第7期57-61,共5页Laser Journal

基　　金：中国航空发动机集团有限公司产学研项目(No.HFZL2020CXY020)。

摘　　要：针对航空发动机叶片缺陷检测过程中表面噪声较大以及检测精度较低的问题,提出了一种基于改进YOLOv5的叶片表面缺陷检测方法。通过叶片表面缺陷图像采集和典型缺陷标注构建了航空发动机叶片表面缺陷数据集;采用K-means++算法代替K-means算法对标记框进行聚类,获得最适合该数据集中标记框的锚框;在主干网络中结合CBAM注意力机制模块,并且采用EIoU损失函数替换原CIoU损失函数,提高对叶片表面缺陷特征的提取能力。对比实验结果表明所提出的方法相较于YOLOv5算法在平均精度均值上提升了1.4%,相较于FasterRCNN和YOLOv3,本方法在平均精度均值上分别提升了13%和2.9%。Aiming at the issues of high surface noise and low detection accuracy in the process of aero-engine blade defect detection,a blade surface defect detection method based on improved YOLOv5 is proposed.The aero-engine blade surface defect dataset is constructed by collecting blade surface defect images and labeling typical defects.K-means++algorithm is used instead of K-means algorithm to cluster the labeled boxes,and the most suitable anchor boxes are obtained.The convolutional block attention module is combined in the backbone,and the EIoU loss function is used to replace the original CIoU loss function to improve the extraction ability of blade surface defect features.The experimental results show that the proposed method improves the average accuracy by 1.4%compared with YOLOv5 algorithm.Compared with FasterRCNN and YOLOv3,the average accuracy of the proposed method is improved by 13%and 2.9%,respectively.

关 键 词：航空发动机叶片 表面缺陷检测 深度学习 YOLOv5 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]