基于红外图像数据域转化和Inception-CNN网络的变压器套管在线异常识别方法  

Online Identification Method for Transformer Bushing Anomaly Based on Data Domain Transformation of Thermal Images and Inception-CNN Network

作　　者：王姿尧 梁惠施 王薇 贡晓旭 周奎 詹阳 WANG Ziyao;LIANG Huishi;WANG Wei;GONG Xiaoxu;ZHOU Kui;ZHAN Yang(Tsinghua Sichuan Energy Internet Research Institute,Chengdu 610213,China;School of Electrical and Electronic Engineering,North China Electric Power University,Beijing 102206,China)

机构地区：[1]清华四川能源互联网研究院,成都610213 [2]华北电力大学电气与电子工程学院,北京102206

出　　处：《高电压技术》2023年第8期3425-3436,共12页High Voltage Engineering

摘　　要：套管故障近年来已成为引起变压器故障的主要原因,采用红外检测可在线识别套管故障。然而,在实际工程中,变压器套管异常状态数据极少或缺失,给变压器套管在线智能异常监测模型的训练带来了困难。为解决这一问题,该文提出了一种基于红外图像数据域转化和Inception卷积神经网络(Inception-convotionalneural networks,Inception-CNN)的变压器套管在线异常识别方法,通过仿真模拟套管多种典型异常情况下的温度分布获得异常红外热图像样本,然后基于循环生成对抗网络(cycle generative adversarial networks, CycleGAN)完成数据在仿真图像域与实测红外图像域之间的映射,实现对异常数据的增强。最后建立基于Inception-CNN的套管异常分类网络,实现对套管异常状态的在线识别。实验结果表明,所提方法对各相套管异常分类预测的平均F1值均在97%以上,与未经过数据域转化的消融实验相比精度更高,说明该方法能够有效识别变压器套管的正常及异常运行状态,并且为异常数据的增强提供了新思路。Bushing faults have become the main cause of transformer failures in recent years,which can be detected online by using thermal images.However,the abnormal states data of thermal images of transformer bushing are rare in practice,which brings difficulties to the model training of online intelligent anomaly identification for transformer bush-ings.To solve this problem,this paper proposes an online identification method for transformer bushing anomaly based on data domain transformation of thermal images and Inception-CNN network.Firstly,abnormal thermal image samples are obtained by simulating the temperature distribution of bushings under a variety of typical abnormal conditions.Then,we complete the mapping between the simulated image domain and the thermal image domain based on CycleGAN to re-alize the enhancement of the abnormal data.Finally,an Inception-CNN-based bushing abnormality classification network is established to identify the bushing abnormal states.The experimental results show that the F1 value of the proposed method exceeds 97%for each phase of bushing abnormality classification,outperforming the results in ablation experiment without data domain transformation.It is revealed that the proposed method can be adopted to effectively identify the normal and abnormal operating states of transformer bushings and provide a new idea for the enhancement of abnormal data.

关 键 词：变压器套管 红外图像域 仿真图像域 数据增强 异常状态识别 数据域映射 

分 类 号：TM407[电气工程—电器] TP391.41[自动化与计算机技术—计算机应用技术]