基于深度卷积神经网络的GIS缺陷局部放电识别

Partial Discharge Recognition of GIS Defects Based on Deep Convolutional Neural Network

为了实现气体绝缘金属开关设备(gas insulated switchgear,GIS)缺陷类型的判别,在实验室以GIS设备为试验对象,设置了自由金属微粒、绝缘子表面微粒、母线金属毛刺、外壳金属毛刺、悬浮电位和绝缘子气泡六种缺陷,对GIS设备施加高压以产生局部放电,应用特高频法对局部放电信号进行检测,提取了缺陷脉冲相位分布模式(phase resolved defect pulse,PRDP)特征图谱。利用深度学习的开源框架Caffe搭建AlexNet卷积神经网络来实现PRDP模式识别,与传统的传统模式识别相比较,卷积神经网络简化了模式识别步骤,提高了计算效率,提升了GIS缺陷识别正确率。

In order to discriminate the types of gas insulated switchgear(GIS)equipment defects,the authors took GIS equipment in the laboratory as the test object,set up six types of deflects-free metal particles,particles with insulator surface,metal burrs on bushbars,shells with metal burrs,floating potential and insulator bubbles,produced partial discharge by imposing high pressure on GIS equipment,applied ultra high frequency(UHF)method to detect partial discharge signals,and extracted the characteristic map of phase resolved defect pulse(PRDP).The AlexNet convolutional neural network is built by Caffe,an open source framework of deep learning,to realize PRDP pattern recognition.Compared with the traditional pattern recognition,the convolutional neural network simplifies the steps of pattern recognition,increases the computational efficiency,and improves the accuracy of GIS defect recognition.