基于PCA与GA-LM-BP神经网络的变压器故障诊断
Transformer fault diagnosis based on PCA and GA-LM-BP neural network
禹建丽 1潘笑天 1陈洪根1
作者信息
- 1. 郑州航空工业管理学院 管理工程学院, 河南 郑州 450046
- 折叠
摘要
研究基于油中溶解气体的变压器故障诊断问题.采用主成分分析与数据归一化方法,对变压器故障样本数据进行规范化处理,使其更具有代表性.对比主成分规范化前后的样本故障诊断结果,主成分分析能够消除特征气体样本数据间的相关性,使输入层样本数据更加符合神经网络工作机理.实验可得主成分规范化后的样本故障诊断结果优于未经过主成分分析规范化的故障诊断结果.在主成分分析对数据规范化的基础上,进一步改进BP神经网络算法,建立基于Levenberg-Marquardt算法的LM-BP神经网络故障诊断模型,改善了BP神经网络模型诊断精度不高,网络收敛困难以及易陷入局部极小值等问题.利用遗传算法对LM-BP神经网络的权值和阈值进行优化,然后再进行第2次神经网络训练,克服了LM-BP神经网络性能受初始权值和阈值限制的问题,使故障诊断正确率提高了6.16%.通过对441组样本数据中随机选取的376组训练样本和65组检验样本进行故障诊断实验,诊断正确率达到83%,表明所构建的基于PCA与GA-LM-BP神经网络的故障诊断方法是一种有效的变压器故障诊断方法.
Abstract
The research on transformer fault diagnosis is based on dissolved gases analysis. The principal component analysis and data normalization method are used to normalize the sample data to make it more representative. Compare the fault diagnosis results of before and after the PCA normalized. The PCA can eliminate the relativity among the characteris-tic gas sample data, therefore the data is more in line with the neural network work mechanism, and has better fault diag-nosis results than that without PCA normalized. Based on PCA, to make further improvement of the BP neural network al-gorithm, construct a LM-BP neural network model of transformer fault diagnosis based on Levenberg-Marquardt algorithm, which helps to solve the problems such as diagnosis inaccuracy of BP neural network model, network converge difficulty and easily trapped in local minima points. In addition, using genetic algorithm to optimize the weights and thresholds of LM-BP neural network, and carry on the second neural network training can help solve the problem of LM-BP neural net-work limited by initial weights and thresholds. As such, the fault diagnosis accuracy is increased by 6.16%. Based on 376 training samples and 65 test samples randomly selected from 441 groups of sample data, the fault diagnosis accuracy can be increased to 83%. Therefore, the fault diagnosis method based on PCA and GA-LM-BP neural network is proved to be an effective fault diagnosis method for transformer.
关键词
主成分分析/神经网络/遗传算法/变压器/故障诊断Key words
principal component analysis/neural network/genetic algorithm/transformer/fault diagnosis引用本文复制引用
基金项目
国家自然科学基金(U1404702)
航空科学基金(2014ZG55021)
河南省科技攻关计划(162102210083)
郑州航院大学生科技创新基金(Y2016L09)
郑州航院研究生教育创新计划基金(2017CX014)
出版年
2017
NETL
NSTL
维普
万方数据