摘要
近年来,数据驱动的人工智能模型在气体绝缘组合电器(GIS)绝缘缺陷诊断上取得了一定突破。然而,这些以海量实验数据构建的模型难以部署到现场复杂工况和小样本条件下,导致现有诊断方法现场应用困难。为了解决现场制约传统诊断方法应用的数据匮乏难题和现有诊断模型现场应用困难的问题,该文提出了一种新颖的域对抗迁移卷积神经网络用于小样本下的GIS绝缘缺陷智能诊断。首先,以自动寻优构建的卷积神经网络从缺陷样本中学习可迁移绝缘缺陷表征特征,自动寻优构建方法在减少网络构建过程人为干预的同时,有效提升了网络精度等多方面性能。然后,引入域对抗迁移学习,实现海量数据(源域)下训练模型到复杂工况和小样本(目标域)下的迁移,以提升诊断准确率。通过对抗训练方法学习类边界表征特征和域空间表征特征,实现了诊断知识的迁移。在域对抗训练中引入两个领域分类器来进行决策边界域空间的对齐,获得了更合适的特征匹配。在实验室和现场实验验证中,所提方法在目标域下分别达到了99.35%和90.35%的诊断准确率。结果表明,该方法可以有效学习可迁移特征,实现小样本GIS绝缘缺陷的高精度、鲁棒性诊断。
In recent years,the artificial intelligence diagnosis method driven by massive data has made a certain breakthrough in the diagnosis of gas-insulated switchgear(GIS)insulation defects.However,these methods built with massive laboratory data are difficult to deploy to complex working conditions and small sample conditions on site,which leads to difficulties in field application of existing diagnostic methods.In order to solve the problem of data scarcity that restricts the application of traditional diagnostic methods in the field and the problem of difficult field application of existing diagnostic models,this paper proposes a novel domain adversarial transfer convolutional neural network for small samples GIS insulation defect diagnosis.First,the convolutional neural network(CNN)constructed with automatic optimization learns transferable features from defective samples.The automatic optimization construction method reduces human intervention in the network construction process,and effectively improves the network accuracy and other aspects of performance.Then,domain adversarial transfer learning is introduced to realize the migration of the trained model under massive data(source domain)to small sample complex working conditions(target domain)to realize the reliable diagnosis under complex small samples.Through the adversarial training,the class boundary representation feature and the domain space representation feature are learned to realize the transfer of diagnostic knowledge.And two domain classifiers are introduced to align the domain space of the decision boundary,and a more suitable feature matching is realized.The experimental verification was carried out on the laboratory and on-site GIS,which verified that the proposed method achieved 99.35%and 90.35%diagnostic accuracy in the target domains,respectively.The results show that this method can effectively learn transferable features,and achieve high-precision and robust diagnosis of insulation defects in small samples of GIS.
作者
王艳新
闫静
王建华
耿英三
刘志远
Wang Yanxin;Yan Jing;Wang Jianhua;Geng Yingsan;Liu Zhiyuan(State Key Laboratory of Electrical Insulation for Power Equipment Xi’an Jiaotong University,Xi’an 710049 China)
出处
《电工技术学报》
EI
CSCD
北大核心
2022年第9期2150-2160,共11页
Transactions of China Electrotechnical Society
基金
国家电网有限公司科技项目资助(5500-202199527A-0-5-ZN)。
关键词
气体绝缘组合电器
域对抗迁移学习
卷积神经网络
小样本
智能诊断
Gas-insulated switchgear(GIS)
domain adversarial transfer learning
convolutional neural network
small sample
intelligence diagnosis
