基于深度迁移学习的输电线路涉鸟故障危害鸟种图像识别  

Image Recognition of Harmful Bird Species Related to Transmission Line Outages Based on Deep Transfer Learning

作　　者：邱志斌 石大寨 况燕军 廖才波 朱轩 QIU Zhibin;SHI Dazhai;KUANG Yanjun;LIAO Caibo;ZHU Xuan(Department of Energy and Electrical Engineering,Nanchang University,Nanchang 330031,China;Electric Power Research Institute of State Grid Jiangxi Electric Power Co.,Ltd.,Nanchang 330096,China)

机构地区：[1]南昌大学能源与电气工程系,南昌330031 [2]国网江西省电力有限公司电力科学研究院,南昌330096

出　　处：《高电压技术》2021年第11期3785-3794,共10页High Voltage Engineering

基　　金：国网江西省电力有限公司科技项目(52182018000W);江西省青年科学基金(20192BAB216028);江西省重点研发计划(20201BBE51019);江西省研究生创新专项资金项目(YC2020-S096)。

摘　　要：为了实现输电线路渉鸟故障的差异化防治,提出一种基于深度迁移学习的危害鸟种图像识别方法。根据历史渉鸟故障的鸟种信息及输电走廊周边鸟种调查结果,建立88种相关鸟类图像数据集,采用类激活映射进行图像预处理,滤除复杂背景噪声。基于迁移学习的思想,首先利用AlexNet、VGG16、ResNet50、Inception V3这4种深度卷积网络架构建立学习模型,采用ImageNet图像数据集对其进行预训练,通过对微调预训练后的网络结构进行模型迁移,使其匹配鸟种图像识别任务。然后,利用鸟种图像样本集对迁移学习模型进行训练与测试,对比4种网络模型的识别准确率。最后,借鉴Delphi法的思想,建立一种融合多卷积神经网络的渉鸟故障危害鸟种识别模型。算例验证结果表明,该模型对88种危害鸟种的识别准确率可达91.21%,能够有效实现架空输电线路巡检图像中的鸟种识别,进而为渉鸟故障防治提供参考。In order to achieve differential prevention of bird-related outages happened on transmission lines,this paper proposes an image recognition method of harmful bird species based on deep transfer learning.According to the information of historical bird-related outages and the investigation results of bird species in the vicinity of transmission corridors,an image dataset including 88 relevant bird species was constructed.The bird images were preprocessed by the class activation mapping(CAM)method,thus to eliminate the complex background noise.Based on the idea of transfer learning,four learning models were firstly established by different deep convolutional networks including AlexNet,VGG16,ResNet50,and Inception V3.These learning models were pre-trained by ImageNet data set,and the network structures after pre-training were fine adjusted for model transfer,so as to match the bird image recognition task.Then,the bird image sample set was used for training and test of the transfer learning model.The recognition accuracy of the four networks was compared.Finally,combined multiple convolution neural networks,a recognition model of harmful bird species related to transmission line outages was established based on the idea of Delphi method.The result of a verification case indicates that the recognition accuracy of this model reaches 91.21%for 88 kinds of harmful bird species,which can recognize the bird species in overhead transmission line inspection images effectively and thus to provide reference for preventing bird-related outages.

关 键 词：输电线路 渉鸟故障 深度迁移学习 类激活映射 卷积神经网络 鸟种图像识别 

分 类 号：TM75[电气工程—电力系统及自动化] TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]