核主成分分析与随机森林相结合的变压器故障诊断方法
Transformer Fault Diagnosis Method Using Random Forests and Kernel Principle Component Analysis
胡青 1孙才新 1杜林 1李剑1
作者信息
- 1. 重庆大学输配电装备及系统安全与新技术国家重点实验室,重庆,400030
- 折叠
摘要
油中溶解气体分析(dissolved gas analysis,DGA)是变压器故障诊断的重要方法.变压器故障诊断研究大多采用人工智能方法学习建立单个分类器,与单个分类器相比,分类器群能够更全面地学习样本集特性,达到更好的诊断效果.分类器间的差异性是影响群体性能的主要因素,针对DGA特征量较少训练得到的分类器差异不大的问题,提出将核主成分分析(kernel principle component analysis,KPCA)与随机森林方法相结合,KPCA将样本从低维的状态空间非线性地映射到高维的核空间,在核空间用随机森林方法训练得到分类器群.对DGA故障样本以及加噪样本的诊断实验结果表明,KPCA能够有效地提取故障特征,用核特征量建模的诊断效果优于直接采用DGA特征量,分类器群的诊断效果以及抗干扰能力均高于单个分类器.
Abstract
DGA(dissovled gas analysis) is one of the most important approaches in transformer fault diagnosis. Most diagnosis model is to construct a single classifier with some Al algorithms,but compared to single classifier,classifier ensemble can achieve better performance. To the performance of classifier ensemble,keeping the differences a-mong classifiers is vital,however,the differences are small among classifiers built with same features when there're few features available. To apply random forests to transformer fault diagnosis,KPCA(kernel principle component a-nalysis) was introduced to increase the number of features,then random forests were constructed in high dimensional kernel space. The experimental results show that,KPCA can effectively extract fault characteristics,models constructed with kernel features have better performance than those with DGA features,and the ensemble achieves higher accuracy than single classifier.
关键词
电力变压器/故障诊断/溶解气体分析/分类器群/随机森林/核主成分分析Key words
power transformer/fault diagnosis/dissolved gas analysis/classifier ensemble/random forests/kernel principle component analysis引用本文复制引用
基金项目
国家重点基础研究发展规划(973计划)(2009CB724506)
出版年
2010
国家科技期刊平台
NSTL
维普
万方数据