基于最优小波包变换与核主分量分析的局部放电信号特征提取

Feature Extraction for Partial Discharge Signals Based on the Optimal Wavelet Packet Basis Transform and Kernel Principal Component Analysis

UHF法作为GIS设备PD检测的有效方法已得到广泛应用,但GIS内UHFPD信号的特征提取一直是研究的难点问题。作者从小波包对UHFPD信号分解过程入手,根据已建立的GIS内4种典型缺陷UHFPD数学模型,分别采用熵最小原则选取最优小波包基,利用所得到的最优小波包基对UHFPD信号进行分解得到的小波包系数,计算信号在各频带投影序列的能量、在各个尺度下的模极大值和绝对平均值,构造出能完整描述UHFPD信号的特征空间,并用KPCA法将高维特征空间降到低维特征空间,解决了维数危机,消除了类内散度矩阵的奇异性,并最大限度地保持原有信号的特性。由此作为模式识别的特征量能够较好地应用于UHFPD信号模式识别。

Ultra-high frequency （UHF） method has been widely used for partial discharge （PD） detection in gas insulated switchgear （GIS）,but the feature extraction for UHF PD signals is a difficult issue all the while. In this paper,a method using wavelet packet transform （WPT） is proposed to decompose the UHF PD signals,and the best basis is selected using minimum entropy criterion based on UHF PD mathematical model of four typical defects in GIS,then the energy in each frequency range,maximal values of module and absolute average values in each scale are computed according to WP coefficients,and the features space is constructed integrally; Kernel principal component analysis （KPCA） is also proposed for reducing dimension of features,and dimension crisis is resolved well,and the divergence matrix strangeness in every class is eliminated. At the same time,the characteristics of signals are retained the farthest. The classification results show that the features used in this paper are quite well for UHF PD defect identification.