基于卷积神经网络的变压器故障诊断方法

变压器是电力系统中的重要设备,其安全与稳定直接影响着国民经济的健康发展。油中溶解气体分析(Dissolved Gas Analysis,DGA)是分析变压器故障类别的重要手段。卷积神经网络是深度学习的一种模型,广泛应用于图像识别、语音处理等领域,具有非常好的分类能力。文章选取了变压器的五种油中溶解气体含量作为模型输入量,在借鉴传统浅层BP神经网络油中气体分析方法的基础上,针对BP神经网络表达能力不足以及容易过拟合的缺点,将卷积神经网络应用于变压器故障诊断,并与BP神经网络的分类效果进行了对比,通过算例研究证明了卷积神经网络的效果更优。文章也对卷积神经网络的卷积核个数、卷积核大小以及采样宽度对分类效果的影响进行了探讨。

Fault diagnosis method of transformer based on convolutional neural network

Transformer is an important equipment in power system, its security and stability directly affect the healthy development of the national economy.Dissolved gas analysis (DGA) is a key method of transformer fault analysis.The convolutional neural network, as an important model of deep learning, has strong classification ability, which is widely used in image recognition, speech processing, and so on.The content of five kinds of dissolved gases is selected as the input of the model in this paper.On the basic of analysis method of dissolved gases by using BP neural network, according to the shortcomings that BP neural network is insufficient in expression ability and easy to over-fitting, the application of convolutional neural network is proposed to diagnose transformer fault in this paper.Moreover, its simulation proves that the proposed method has a better performance compared with BP neural network.Additionally, the effect of convolution kernel number, kernel size and sampling width of convolutional neural network on the classification results is discussed in this paper.