充油电气设备油色谱故障诊断ANN方法的性能优化  

Performance Optimization of ANN Method for Oil-filled Electrical Equipment Fault Diagnosis Based on DGA

作　　者：张宝全 马雅丽 关睿 白诗婷 李静 胡伟涛 ZHANG Baoquan;MA Yali;GUAN Rui;BAI Shiting;LI Jing;HU Weitao(State Grid Hebei Electric Power Co.,Ltd.Maintenance Branch,Shijiazhuang,Hebei 050070,China;Department of Electrical Engineering,North China Electric Power University,Baoding,Hebei 071003,China)

机构地区：[1]国网河北省电力有限公司检修分公司,河北石家庄050070 [2]华北电力大学电力工程系,河北保定071003

出　　处：《广东电力》2021年第6期39-47,共9页Guangdong Electric Power

基　　金：国网河北省电力有限公司科技项目(kj2018-58)。

摘　　要：为了提高基于人工神经网络(artificial neural network,ANN)方法的充油电气设备油色谱故障诊断的准确性,对油色谱故障诊断中ANN的参数优化问题进行系统研究。基于搜集得到的大量故障特征气体数据,构建多层前馈ANN系统,研究训练算法、隐层神经元数量、训练目标、隐层和输出层神经元激活函数对训练性能和诊断准确率的影响。结果表明:训练目标一致时,8种不同算法训练得到的ANN具有相近的故障诊断准确率,建议选择其中速度最快的Levenberg-Marquardt算法来训练ANN;隐层神经元数量在一定范围内对故障诊断准确率影响不大;随训练目标(均方误差)的减小,训练时间增加,诊断准确率先增大后减小,建议均方误差选择0.01;隐层神经元激活函数选择线性函数时网络训练不易收敛,建议选择sigmoid函数。To improve the accuracy in fault diagnosis of the artificial neural network(ANN)method for the oil-filled electrical equipment using dissolved gas-in-oil analysis(DGA),this paper studies the problem of parameter optimization for ANN used in fault diagnosis based on DGA.According to a large number of collected DGA samples,it constructs a multi-layer feedforward ANN,and studies the influence of the training algorithms,the numbers of hidden layer neurons,the training goals,and the activation functions of the hidden layer and output layer neurons on training performance and diagnosis accuracy.The results indicate that the ANNs with the same training goal trained by eight algorithms including the back propagation(BP)algorithm,the BP algorithm with momentum term,the BP algorithm with variable learning rate,the conjugate gradient method with Powell-Beale restarts and Polak-Ribiére updates,the resilient propagation algorithm,the quasi-Newton method and the Levenberg-Marquardt algorithm have similar accuracy in fault diagnosis.The paper suggests to adopt the Levenberg-Marquardt algorithm with the fastest speed to train the ANN.In addition,the numbers of hidden layer neurons have little effect on the accuracy in fault diagnosis,and with the decrease of training goals,the training time increases and the fault diagnosis accuracy increases firstly and then decreases.It suggests to set the root mean squared value of the ANN as 0.01.When the activation function of the hidden layer neurons is the linear function,the network is hard to converge in training,the paper suggests to choose the sigmoid function.

关 键 词：充油电气设备 油中溶解气体分析 人工神经网络 故障诊断 性能优化 

分 类 号：TM855.1[电气工程—高电压与绝缘技术] TP183[自动化与计算机技术—控制理论与控制工程]