将具有可信度的BP神经网络应用于变压器故障诊断

Application of BP ANN with Credence for Transformer Fault Diagnosis

BP神经网络具有自学习和自适应能力,比较适合于基于油中溶解气体分析的变压器故障诊断。分析了常规BP算法的不足,指出了加动量项BP算法的原理和优点。该方法减轻了网络训练过程中的振荡,加速了网络的收敛。针对常规的BP神经网络不能给出诊断结果的可信度问题,利用多个网络共同诊断,根据多个诊断结果的标准差获得诊断结果的可信度,多个结果的均值作为最终诊断结果,从而提高了准确性。构造了适合于变压器油中溶解气体分析故障诊断的神经网络,并将训练所得的多个神经网络用于变压器故障诊断。结果表明了该方法的有效性。

BP neural network has learning and adaptive capacity, so it is suitable to transformer fault diagnosis based on DGA （Dissolved Gas-in-oil Analysis）. But the ordinary BP algorithm has some shortcomings, and BP algorithm with momentum can smooth out the oscillations in the network train trajectory. The modified algorithms smooth out the oscillations and speed up convergence of the network training. In the meantime, the diagnosis of ordinary BP ANN has not credence, then variance of multiply ANNs diagnosis result can used to given credence, and the mean ofthern is the last diagnosis result, and which can decrease error. The artificial neural network for transformer fault diagnosis is constructed, and trained by the modified algorithm; cases verified the effect of the algorithm.