基于改进YOLOv7的航空发动机叶片表面缺陷检测

打开文本图片集

中图分类号：TN911.73-34；TP391.4；V260.6 文献标识码：A 文章编号：1004-373X（2025）15-0135-09

Aeroengine bladesurface defect detection based on improved YOLOv7

WURenkangl，CHENGZhijiang²，WUDongbo，WANGHui4，LIANGJiawei5

（1.CollegeofElectricalEngineering，XinjiangUniversity，Urumqi 83oo17，China; 2.Schoolof IntelligenceScienceand Technology，XinjiangUniversity，Urumqi83oo17，China; 3.InstituteforAeroEngine，Tsinghua University，Beijing1Ooo86，China; 4.Research InstituteofAero-Engine，BeihangUniversity，Beijing1O22O6，China; 5.Department of Mechanical Engineering，TsinghuaUniversity，Beijing1Ooo86，China）

Abstract：Thedetection of defectsarising during the manufactureand processing ofaero-engine blades relies heavily on manualvisualinspection.AnenhancedYOLOv7basedapproachforprecisebladedefectdetectionisintroducedtogetidof the inconsistentoutcomesduetohumansubjectivityandlimitationsandineficiency.Inviewof thefourkindsofcommontypical defects inproductionand procesing，thedatasetofaero-enginebladesurfacedefectsisconstructed.Themodelisenhanced by integrating SKNet（selectivekernelnetworks）intotheELAN-W（effcientlayeragregationnetworkswide）oftheYOLOv7's featurefusionnetwork，enablingittoobtainanadaptivereceptivefieldtoimproveitsabilityofextracting networkfeatures.The introductionofDyhead（Dynamichead）intheheadnetworkfurtherelevatesthemodel'scategoryrecognitioncapabilityand detectionperformance.Moreover，the MPDIoU（IoUwithminimumpointsdistance）lossfunctionisadopted toreplacetheoriginal CIoU（completeintersectionoverunion）lossfunction，soastoachieve more precise boundingboxregresson.The proposed methodimprovesthedetectionperformanceofthemodelwhilemaintainingahighrecallrate.Theprecisionof themodelis improved by 5.3% ，its recall rate is improved by 2.2% ，and its mAP@0.5 is improved by 3.7% . The detection time for asingle bladeis 4.93s .Thisresearch provides a novel approach for automated blade defect detection.

Keywords：computervision;defectdetection;aeroengineblade;improvedYOLOv7;deep learning;MPDIoUlossfunction

0 引言

在航空发动机叶片的生产过程中进行质量检测尤为重要。（剩余13506字）