文章编号：1000-5641(2023)02-0132-11基于残差网络的配电柜设备元件状态识别张洋,赖叶静,黄定江（华东师范大学数据科学与工程学院,上海200062）摘要:随着工业智能巡检的不断发展,基于数字图像处理方法的设备元件状态识别系统被广泛应用.为提升配电室中配电柜设备元件状态识别的准确率,提出了一种基于残差网络(residualnetworks,ResNet)的设备元件状态识别方法.首先搭建数据采集系统,构建数据集;然后对配电柜图像,裁剪预设的设备元件目标区域,生成设备元件图像;对于设备元件图像,构建基于ResNet的元件状态识别模型并训练;使用训练完毕的模型识别元件的状态.以变电站配电室中配电柜设备元件数据集作为研究对象,对于特征复杂的元件采用单预测头的网络,对于特征简单的元件采用多预测头的网络;然后使用紧凑和剪枝的模型压缩方法在精度损失较小的情况下减少参数量和计算量;最后介绍巡检系统的架构设计,将JetSonNano边缘终端作为算法模块的运行硬件,以减少通信成本.关键词:智能巡检;残差网络;图像识别;模型压缩中图分类号:TP391文献标志码:ADOI:10.3969/j.issn.1000-5641.2023.02.014DevicecomponentstaterecognitionmethodofpowerdistributioncabinetbasedonaresidualnetworksZHANGYang,LAIYejing,HUANGDingjiang（SchoolofDataScienceandEngineering,EastChinaNormalUniversity,Shanghai200062,China）Abstract:Withthecontinuousdevelopmentofindustrialintelligentinspectiontechnology,theequipmentelementstaterecognitionsystembasedondigitalimageprocessingiswidelyused.Inordertoimprovetheaccuracyofpowerdistributioncabinet(PDC)equipmentelementstaterecognitioninadistributionroom,aResNet(residualnetworks)-basedequipmentelementstaterecognitionmethodisproposed.Firstly,thedataacquisitionsystemissetupandthedatasetisconstructed.Then,forthePDCimage,thepresetdevicecomponenttargetareaiscroppedtogeneratethedevicecomponentimage.Fordevicecomponentimages,aResNet-basedcomponentstaterecognitionmodelwasconstructedandtrained,andthetrainedmodelwasusedtoidentifycomponentstates.Takingthedatasetforpowerdistributioncabinetequipmentelementinsubstationdistributionroomsastheresearchobject,anetworkofsinglepredictionheadsisadoptedasthecomponentwithcomplexfeatures,andthenetworkofmultiplepredictionheadsisad...
