改进YOLOv7的航空发动机叶片损伤检测方法  

Improved YOLOv7 Damage Detection Method for Aeroengine Blade

作　　者：魏永超 刘嘉欣 朱泓超 朱姿翰 刘伟杰 WEI Yong-chao;LIU Jia-xin;ZHU Hong-chao;ZHU Zi-han;LIU Wei-jie(Department of Scientific Research Office1,School of Civil Aviation Safet Engineering,Deyang Sichuan 618307,China;Department of Scientific Research Office1,School of Avionics and Electrical3,Deyang Sichuan 618307,China;Department of Scientific Research Office1,Civil Aviation Flight Academy of China,Deyang Sichuan 618307,China)

机构地区：[1]中国民用航空飞行学院科研处,四川德阳618307 [2]中国民用航空飞行学院民航安全工程学院,四川德阳618307 [3]中国民用航空飞行学院航空电子电气学院,四川德阳618307

出　　处：《航空发动机》2025年第1期133-139,共7页Aeroengine

基　　金：中央高校基本科研业务费(J2021-056);四川省科技厅重点研发项目(2022YFG0356);西藏科技厅重点研发计划(XZ202101ZY0017G);中科院西部青年学者项目;中国民用航空飞行学院科研基金(J2020-040,CJ2020-01)资助。

摘　　要：针对目前航空发动机叶片损伤检测精度低的问题,提出了一种基于改进YOLOv7的发动机叶片损伤检测模型YOLOv7-CC。对发动机叶片缺损图像进行损伤标注,构建航空发动机叶片损伤数据集,并且采用二分K-means算法对标记框进行聚类,获取与该数据集最匹配的锚框(anchor)。在模型中Backbone网络输出之后采用坐标注意力机制,分别捕获长距离依赖关系和保留精确的位置信息,提高对损伤目标的检测能力,并在特征重组过程中采用轻量级上采样算子(CARAFE),同时保留了语义信息以及位置信息,通过更大的感受野来完成上采样,提高了网络对特征的提取能力。结果表明:所提出的基于YOLOv7-CC算法的损伤检测的平均精度达到了83.53%,相较于基准网络提升了7.4%,能够对航空发动机叶片3种常见的损伤类型实现高效检测。Aiming at the current problem of low accuracy of aeroengine blade damage detection,an engine blade damage detection model YOLOv7-CC based on improved YOLOv7 was proposed.The engine blade defect images were labeled with damage to construct an aeroengine blade damage dataset,and the labeled frames were clustered using the bifurcated K-means algorithm to obtain the anchors that best match this dataset.After the output of Backbone network in the model,the coordinate attention mechanism was used to capture the long-distance dependency and retain the accurate position information respectively,to improve the detection ability of the damage target,and the CARAFE lightweight up-sampling algorithm was used during the feature reorganization process,retaining the semantic information as well as the positional information at the same time;the up-sampling was completed through the larger sensory field,improving the feature extraction ability of the network.The results show that the proposed YOLOv7-CC algorithm for damage detection achieves an average accuracy of 83.53%,which is a 7.4%improvement compared to the baseline network,and is able to realize highly efficient detection of the three common damage types of aeroengine blades.

关 键 词：损伤检测 深度学习 YOLOv7模型 注意力机制 航空发动机 

分 类 号：V232.4[航空宇航科学与技术—航空宇航推进理论与工程]