相关分析技术在直流电压下XLPE电缆局部放电信号特征提取与筛选中的应用

Application of Correlation Analysis Techniques in Feature Extraction and Selection for DC Partial Discharge Signals of XLPE Cables

目前,关于直流电压下局部放电信号特征提取技术的研究极少。用于表征连续放电间相关关系的特征散点图是常用的统计分析方法,但现阶段仅用于定性分析放电现象。引入互信息、最大信息系数（maximal information coefficient,MIC）、最大信息非参数扩展类（maximal information-based non-parametric exploration,MINE）等先进的非线性相关特征分析手段,提取该类散点图定量特征。基于互信息的MIC和MINE具有普适性、公平性和对称性等重要特性。最终共提取了36个相关特征参数,与22个传统统计算子一起组成特征指纹。之后,使用最大相关最小冗余（mR MR）算法选取最优特征指纹空间并使用MIC进行优化。利用XLPE单芯电缆制作了绝缘内部气隙、主绝缘表面划伤、高压端毛刺电晕、半导电层爬电4类典型绝缘缺陷模型,将文中方法应用于试验数据分析。最终确定了含有48个参数的最优特征指纹,使用人工神经网络等机器学习方法进行模式识别可获得91%的平均识别精度。该结果表明,使用文中方法提取的散点图非线性特征可以有效反映放电模式。

So far, studies on PD feature extraction technique at DC voltage are comparatively lagged. Scatter plot is a conventional statistical method used to describe relationship between continuous discharges qualitatively. In this paper, theories of mutual information（MI）, maximal information coefficient（MIC） and maximal-information-based non-parametric exploration（MINE） are introduced to reflect complicated nonlinear relationship in scatter plots. MIC and MINE have many advantages（especially the properties of generality, equitability and symmetry）. In total, 36 correlation characteristic parameters and 22 conventional operators are proposed to combine a fingerprint. Minimal redundancymaximal relevance（m RMR） method is improved with MIC and applied to feature selection. Four kinds of artificial defects are designed in XLPE cables： cavity inside XLPE, scratch at XLPE surface, metal needle inside XLPE and semiconductor fault. The method proposed in this paper is applied to experimental data. Finally, an optimal fingerprint with 48 selected features is built, and its average accuracy can achieve to 91% when using ANN to classify. Results show that non-linear feature of scatter plots can effectively reflect PD patterns.