利用小波神经网络的电力变压器故障诊断方法

Fault Diagnostic Mothed Using Wavelet Neural Network for Power Transformers

为提高变压器传统油中溶解气体分析(DGA)的故障诊断能力,提出了一种利用小波神经网络(WNN)的变压器故障诊断方法。WNN隐含层采用离散仿射小波函数,仿照前馈BP神经网络算法构造WNN,引入学习率和动量系数来训练网络。实验结果表明:相同条件下,较之传统比值法与BP神经网络,WNN的故障模式识别准确率更高,对照BP神经网络,所提出的WNN变压器故障诊断方法在稳定性和收敛时间方面表现更优。

Building a fault diagnostic model is the key point of the power transformer. In order to enhance the fault diagnostic condition-based maintenance （CBM） technology for ability of conventional dissolved gas analysis （DGA） in power transformer, this paper proposes an incipient fault diagnostic method using wavelet neural network {WNN）. The discrete affine wavelet function is employed as an activation function of the hidden layers, and the learning ratio and the momentum coefficient were used to train the feed-forward BP algorithm. The weight value and bias value of WNN were adjusted by this method, so the computation quantity of WNN was reduced, the convergence speed of WNN was increased, and learning capability and fault diagnosis accuracy of WNN were improved. Experimental results based on 500 actual gas records of civil power transformers on the same condition demonstrate that the WNN model provides the highest diagnostic efficiency in faults pattern cognition which can reached over 87% compared with conventional ratio methods （under 65%） and BP neural network （81M）, while WNN requires less convergence time and shows better stability than BP neural network. So the method proposed can satisfy the needs of CBM technology. This paper also draws a conclusion that choosing the proper hidden layer node numbers is another key point to fault diagnosis based WNN and ANN, in addition, the fault diagnostic efficiency will be improved with the increase of hidden layer nodes, however the efficiency will slightly be lowered and reach saturation when the number of hidden layer nodes reaches a certain level.