测控技术2023年第42卷第7期模式识别与人工智能收稿日期:2022-10-14基金项目:国网新疆电力有限公司科技项目(5230BD220001)引用格式:陈剑波,唐锐,姚平,等.基于深度学习的架空线路关键部件典型缺陷识别研究[J].测控技术,2023,42(7):22-28.CHENJB,TANGR,YAOP,etal.ResearchonTypicalDefectsIdentificationofKeyComponentsofOverheadLinesBasedonDeepLearning[J].Measurement&ControlTechnology,2023,42(7):22-28.基于深度学习的架空线路关键部件典型缺陷识别研究陈剑波,唐锐,姚平,张赛飞,廖林(国网新疆电力有限公司巴州供电公司,新疆库尔勒841000)摘要:为了精确识别定位架空线路中关键部件的缺陷情况,提出了一种改进的YOLOv3检测方法,首先采用RepVGG模块替换骨干网络Darknet-53中的残差单元,加快推理速度;其次通过改进损失函数,引入SIoU使模型训练更快,精确率更高;最后通过改进检测头,采用不同的分支进行计算,提升检测效果。实验结果表明,改进方法与YOLOv3相比,精确率提高了3.4%,召回率提高了2.6%;性能相比于SSD、FasterR-CNN网络模型也具有一定优越性。关键词:YOLOv3;目标检测;RepVGG;SIoU;架空线路巡检中图分类号:TP391.4文献标志码:A文章编号:1000-8829(2023)07-0022-07doi:10.19708/j.ckjs.2023.07.004ResearchonTypicalDefectsIdentificationofKeyComponentsofOverheadLinesBasedonDeepLearningCHENJianbo,TANGRui,YAOPing,ZHANGSaifei,LIAOLin(StateGridXinjiangElectricPowerCo.Ltd.BazhouPowerSupplyCompany,Korla841000,China)Abstract:Inordertoaccuratelyidentifyandlocatethedefectsofkeycomponentsinoverheadlines,anim-provedYOLOv3detectionmethodisproposed.Firstly,theRepVGGmoduleisusedtoreplacetheresidualunitinthebackbonenetworkDarknet-53tospeedupinference.Secondly,byimprovingtheLossfunction,SIoUisintroducedtomakethemodeltrainingfasterandimprovetheaccuracy.Finally,byimprovingthedetectionheadandusingdifferentbranchesforcalculation,thedetectioneffectisimproved.TheexperimentalresultsshowthatcomparedwithYOLOv3,theimprovedmethodimprovestheprecisionrateby3.4%andtherecallrateby2.6%,italsohascertainadvantagescomparedwithSSDandFasterR-CNNnetworkmodelsinper-formance.Keywords:YOLOv3;targetdetection;RepVGG...