Nguyen-Widrow法用于变压器油气分析故障诊断

为解决神经网络在油中溶解气体分析变压器故障诊断的应用中存在收敛速度慢的问题,将Nguyen-Wid-row方法用于神经网络可变参数的初始化。该法通过调整隐层神经元的权值和阈值使隐层各个神经元输入的线性区间相等,从而减少网络训练时权值和阈值的调整量,加快网络收敛速度,增强网络学习能力,提高故障诊断的精度。对393个样本用5-80-6的结构网络分别不使用和使用该法训练,其样本误差平方和分别为37·10和15·21,训练所得网络对155个变压器的诊断结果准确率分别为80·97%和84·19%。表明该法的确能提高神经网络的学习能力和对变压器的故障诊断能力。

Application of Nguyen-widrow Method in Transformer Dissolved Gas-in-oil Analysis Fault Diagnosis

For solving the problem of slow convergence speed of ANN used in transformer fault diagnosis based on dissolved gas-in-oil analysis (DGA), Nguyen-Widrow method is used to initialize the variable parameters of ANN. The weight value and bias value of ANN are adjusted by the method to make neurons of hidden layer have equal linear interval, so the computation quantity of ANN is reduced, and the convergence speed of ANN is increased, learn capacity and fault diagnosis accuracy of ANN is improved. A neural network take 5-80-6 as architecture, and 393 samples are used to train, and 155 samples are used to verify effect of training is build. The method of Nguyen-Widrow makes the training error sum of square of ANN from 37.1 down to 15.21, and make the diagnosis accuracy of ANN from 80.97% improved to 84.49%. The effect of the method is verified by the training result and diagnosis reuslt of ANN for transformer fault diagnosis.