基于CycleGAN与深度残差网络的局放数据增强与模式识别方法  

Partial Discharge Data Enhancement and Pattern Identification Method Based on CycleGAN and Deep Residual Network

作　　者：刘兆宸 谢庆[1] 王春鑫 张雨桐 李靖航 谢军[1] 戴珍 侯佳萱 LIU Zhaochen;XIE Qing;WANG Chunxin;ZHANG Yutong;LI Jinghang;XIE Jun;DAI Zhen;HOU Jiaxuan(State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources,North China Electric Power University,Hebei Baoding 071003,China;Digital Grid Research Institute of China Southern Power Grid,Guangzhou 510670,China)

机构地区：[1]华北电力大学新能源电力系统国家重点实验室,河北保定071003 [2]南方电网数字电网研究院有限公司,广州510670

出　　处：《高压电器》2022年第11期106-113,共8页High Voltage Apparatus

基　　金：国家重点研发计划(2020YFB0906000,2020YFB0906005)。

摘　　要：为了解决局放模式识别的准确性受不平衡样本与神经网络深度结构制约的问题,提出了一种基于CycleGAN与深度残差网络的局放数据增强与模式识别方法。首先对稀疏表示去噪及脉冲提取得到的局放脉冲信号进行S变换得到局放时频谱图作为训练样本。然后利用CycleGAN实现对局放时频谱图的重构增强,同时引入对抗损失函数、循环一致性损失函数,以保证局放数据的高质量生成,扩充后的局放样本库丰富度更高。最后利用增强后的局放数据集训练深度残差网络,利用残差块的恒等映射结构自适应调节网络深度,解决了深度网络不易收敛的问题,同时实现对局放信号的精准辨识。实验结果表明,经数据增强的深度残差网络模式识别准确率达到98%,较增强前提高了6.8%。For solving the problem that the accuracy of PD pattern identification is restricted by unbalanced samples and the depth structure of neural network,a kind of method of PD data enhancement and pattern identification based on CycleGAN and depth residual network is proposed.First,PD pulse signal obtained by sparse representation de⁃noising and pulse extraction is transformed by S⁃transform to obtain the partial discharge time spectrum as the train⁃ing sample.Then,CycleGAN is used to achieve reconstruction enhancement of spectrum of partial discharge.At the same time,confrontation loss function and cyclic consistency loss function are introduced to ensure the high⁃quality generation of partial discharge and the richness of the expanded partial discharge database is stronger.Finally,the en⁃hanced partial discharge data set is used to train the depth residual network,and the identity mapping structure of the residual block is used to adaptively adjust the network depth,which has solved the problem that the depth network is not easy to converge and,at the same time,achieved the accurate identification of the partial discharge signal.The experimental results show that the pattern identification accuracy of the depth residual network enhanced by Cycl⁃GAN is up to 98%,which is 6.8%higher than that before enhancement.

关 键 词：局部放电 生成对抗网络 深度残差网络 数据增强 模式识别 

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