基于AdaBoost的局部放电综合特征决策树识别方法  

Pattern recognition based on AdaBoost decision tree for partial discharge

作　　者：姚林朋[1] 郑文栋[1] 钱勇[1] 王辉[1] 黄成军[1] 江秀臣[1] 

机构地区：[1]上海交通大学电气工程系,上海200240

出　　处：《电力系统保护与控制》2011年第21期104-109,114,共7页Power System Protection and Control

摘　　要：在GIS局部放电模式识别研究中,为解决传统决策树方法中只针对单一特征及有限模式进行学习而导致决策树结构复杂、预测准确率不高、对噪声数据的抗干扰能力差等问题,提出综合多类特征的AdaBoost决策树识别方法。设计实验并通过超高频方法采集GIS中高压导体毛刺放电、悬浮电极放电、气隙放电、微粒放电及手机、灯光干扰信号,从p-q-n图谱的统计分布、q-t图谱的矩分布、q-n图谱的Weibull分布三个不同角度提取特征,研究单一及综合形式的特征对C4.5决策树及AdaBoost决策树的识别效果的影响。实验及现场检测的识别结果表明综合三类不同特性的特征并通过AdaBoost方法生成决策树,能有效优化决策树的识别性能,提高决策树的时间和空间效率。In the research of pattern recognition on partial discharge （PD） in GIS, the traditional decision tree method faces problems of complex structure, low recognition rate and vulnerability to noise data due to the single features and limited training pattern modes. In this paper, a method of using AdaBoost decision tree integrating with composited features is presented. Features are extracted from three aspects including statistical distribution of p-q-n diagram, moment distribution of q-t diagram and Weibull distribution parameters of q-n diagram and samples are collected from the typical discharges from high voltage needle, floating electrode, void, free particle in GIS and interferences from mobile phone and light. The influence of single features and composited features on the recognition effects of C4.5 decision tree and AdaBoost decision tree is studied. Recognition results of laboratory test and field test show that AdaBoost decision tree made with features composited with three aspects can effectively optimize the recognition rate find improve the efficiency of its time and space use.

关 键 词：气体绝缘组合电器 超高频 局部放电 决策树 ADABOOST C4.5 模式识别 

分 类 号：TM595[电气工程—电器]