基金项目:山西省重点研发计划项目(201803D31041);国网山西省电力公司科技项目(5205B020000V)收稿日期:2021-04-16修回日期:2021-04-22第40卷第2期计算机仿真2023年2月文章编号:1006-9348(2023)02-0132-07多源信息融合的设备热缺陷智能实时检测方法赵锐,周雪枫(国网大同供电公司,山西大同037008)摘要:针对现有算法对多类变电站设备热缺陷诊断效率不高,难以满足边缘端实时检测需求等问题,提出了一种基于多源信息融合的设备热缺陷智能实时检测方法。首先利用粒子群优化的SIFT描述子实现多源图像的配准。其次提出改进的YOLOv4算法实现设备检测,将特征提取网络CspDarket53替换为轻量级网络GhostNet,并将特征融合模块的普通卷积层替换为深度可分离卷积,使模型轻量化;将三尺度检测扩充为四尺度,加强对遮挡目标的检测;在特征融合部分嵌入ASFF模块,提升设备检测精度。最后完成设备的热缺陷诊断以及缺陷等级判定。实验结果表明,文中方法达到93.56%的mAP值,推理速度达到35FPS,可用于变电站设备热缺陷的实时监测。关键词:多源信息融合;热缺陷诊断;实时检测中图分类号:TP391文献标识码:BIntelligentReal-TimeDetectionMethodforEquipmentThermalDefectsBasedonMulti-SourceInformationFusionZHAORui,ZHOUXue-feng(StateGridDatongPowerSupplyCompany,DatongShanxi037008,China)ABSTRACT:Aimingattheproblemthatexistingalgorithmsarenotefficientindiagnosingthermaldefectsofmultipletypesofsubstationequipment,anditisdifficulttomeettheproblemofreal-timedetectionofedge-endequipment,andaredifficulttomeetthereal-timedetectionrequirementsoftheedge,anintelligentreal-timedetectionmethodforthermaldefectsofequipmentbasedonmulti-sourceinformationfusionisproposed.anintelligentreal-timedetec-tionmethodofequipmentthermaldefectsbasedonmulti-sourceinformationfusionisproposed.First,theSIFTde-scriptoroptimizedbyparticleswarmwaisusedtorealizetheregistrationofmulti-sourceimages.Secondly,theYOLOv4algorithmiwasimproved,thefeatureextractionnetworkCspDarket53waisreplacedwithalightweightnet-workGhostNet,andtheordinaryconvolutionlayerofthefeaturefusionmodulewaisreplacedwithadeepseparableconvolutiontocompletemodellightweightprocessing;forForsubstationswithcomplexan...