基于CWGAN-div和Mi-CNN的GIS局部放电图谱识别  

Recognition of partial discharge patterns of GIS based on CWGAN-div and Mi-CNN

作　　者：刘航斌 林厚飞 褚静 叶静 林权威 LIU Hangbin;LIN Houfei;CHU Jing;YE Jing;LIN Quanwei(College of Electrical Engineering and Automation,Fuzhou University,Fuzhou 350000,China;State Grid Pingyang Power Supply Company,Wenzhou,Zhejiang 325401,China)

机构地区：[1]福州大学电气工程与自动化学院,福州350000 [2]国网浙江省电力有限公司平阳县供电公司,浙江温州325401

出　　处：《浙江电力》2023年第8期75-83,共9页Zhejiang Electric Power

摘　　要：为解决GIS(气体绝缘封闭组合电器)局部放电图谱识别任务中样本数量少且分布不均衡对深度学习模型性能的制约,提出一种CWGAN-div(带条件约束的Wasserstein生成对抗网络)模型以指导多类别局部放电图谱的生成,克服了原始GAN(生成式对抗网络)训练不稳定的问题,使样本数据得到增强并将平均不平衡率由11.01降至3.03。然后,使用5种分类器进行样本增强前后的对比实验,各分类器的F_(1mean)值在样本增强后均得到3.7个百分点以上的提升。实验中,Mi-CNN(多输入卷积神经网络)模型因能够同时利用特高频法和超声波法的PRPD(局部放电相位分布)图谱而表现最优,其F_(1mean)值达到95.8%。In order to solve the constraints of the limited number and uneven distribution of samples on the performance of the deep learning model in the identification of partial discharge patterns of GIS(gas-insulated switchgear),a CWGAN-div(conditional Wassertein generative adversarial network-divergence) model is proposed to guide the generation of multi-class partial discharge patterns,which overcomes the instability of the original GAN(generative adversarial network) training,enhances the sample data,and reduces the average imbalance rate from 11.01 to 3.03.Then after using five kinds of classifiers for the comparative experiments before and after sample enhancement,the F1mean value of each classifier has been improved by more than 3.7% after sample enhancement.In the experiment,the Mi-CNN(multi-input-convolutional neural networks) model proposed in this paper can use the PRPD(phase resolved partial discharge) spectrum of ultra-high frequency method and ultrasonic method at the same time,and its final F_(1mean) value reaches 95.8%.

关 键 词：气体绝缘封闭组合电器 局部放电 生成式对抗网络 卷积神经网络 

分 类 号：TM855[电气工程—高电压与绝缘技术] TM595[自动化与计算机技术—控制理论与控制工程] TP183[自动化与计算机技术—控制科学与工程]