基于卷积神经网络的变压器套管故障红外图像识别方法  

Infrared Image Recognition Method on Fault of Transformer Bushing Based on Convolutional Neutral Networks

作　　者：刘云鹏[1,2] 董王英 许自强[1] 夏彦卫 高树国 赵军 LIU Yunpeng;DONG Wangying;XUZiqiang;XIAYanwei;GAO Shuguo;ZHAO Jun(Key Laboratory of Safety Defense for Hebei Power Transmission&Transformation Equipment,North China Electric Power University,Hebei Baoding 071003,China;State Key Laboratory of New Energy Power System,North China Electric Power University,Beijing 102206,China;Electric Power Research Institute of SGCC Hebei Electric Power Co.,Ltd.,Shijiazhuang 050021,China)

机构地区：[1]华北电力大学河北省输变电设备安全防御重点实验室,河北保定071003 [2]华北电力大学新能源电力系统国家重点实验室,北京102206 [3]国网河北省电力有限公司电力科学研究院,石家庄050021

出　　处：《高压电器》2021年第10期134-140,共7页High Voltage Apparatus

基　　金：国家电网有限公司总部科技项目资助(5204DY170010)。

摘　　要：作为电力变压器的重要部件,套管的管理与维护对于设备的安全稳定运行起着至关重要的作用。为提升电力设备巡检的智能化水平,文中提出一种基于卷积神经网络的套管故障红外图像识别方法,该方法在特征提取方面具有显著的优势,避免了人为提取描述特征的低效和易误判问题。首先,建立了包含正常、缺油与局部过热3种状态类型的套管红外图像样本库;然后,将规范化处理后的红外图像作为卷积神经网络的输入,搭建了套管故障红外图像识别模型;最后,通过对网络超参数的选取进行实验分析,确定了激活函数种类、池化方法及卷积核数目。针对文中样本库,文中所提模型对套管3种状态类型的分类结果准确率达到96%,相较于SVM算法和BP神经网络算法分别提升约14%和15%,识别性能更为优异。As the important part of power transformer,the management and maintenance of bushing plays the vital role in the safe and stable operation of the equipment.For improving the intelligent level of patrol inspection of electrical equipment,a kind of infrared image recognition method on the fault of bushing based on convolutional neutral networks is proposed in this paper,which has significant advantages in terms of feature extraction and avoids such problem as low efficiency and prone misjudgment due to artificial extraction of descriptive features.Firstly,Infrared image sample database of the bushing at such three states as normal,lack of oil and local overheating is set up.Then,the normalized processed infrared image is used as the input of the convolutional neutral networks and the fault infrared image recognition model of the bushing is set up.Finally,the type of activation function,pooling method and number of convolutional kernel are defined by way of experimental analysis through selection of network hyper parameters.In view of the sample database in the paper,the accuracy rate of classification result of 3 types of status of the bushing by the model proposed in this paper is up to 96%which,compared to SVM algorithm and BP neutral network algorithm,increases by about 14%and 15%respectively and the recognition performance is more superior.

关 键 词：变压器套管 红外测温 卷积神经网络 图像识别 

分 类 号：TM41[电气工程—电器] TP391.41[自动化与计算机技术—计算机应用技术]