Abstract: Dissolved Gas Analysis (DGA) is one of the important methods for determining transformer internal faults.Aiming at the shortcomings of the traditional shallow⁃based machine learning method in transformer fault diagnosis in feature extraction and generalization ability, a transformer fault diagnosis method based on Convolutional Neural Network (CNN) was proposed. The convolution layer in the network was used for feature conversion of dissolved gas in oil, and fault sensitive features were extracted by combining with the ability of pooling layer to strengthen important features. The effects of the number of convolution kernels, the size of convolution kernels, pooling layer and network depth on the diagnostic performance of the model were studied experimentally. The models of Convolutional Neural Network, Support Vector Machine (SVM) and Back Propagation Neural Network (BPNN) were compared and analyzed by confounding matrix, ROC curve and PR curve. The experimental results show that the Convolutional Neural Network model has better diagnostic performance.

Key words: Dissolved gas in oil, Transformer, Fault diagnosis, Convolutional neural network

摘要： 油中溶解气体分析法（Dissolved Gas Analysis，DGA）是判断变压器内部故障的重要方法之一。针对传统基于浅层的机器学习方法在变压器故障诊断中存在的特征提取和泛化能力方面的不足，提出了一种基于卷积神经网络的变压器故障诊断方法。利用网络中的卷积层对油中溶解气体进行特征转换，结合池化层强化重要特征的能力，对故障敏感特征进行提取。通过实验研究了卷积核数目、卷积核大小、池化层、网络深度对模型诊断性能的影响。通过混淆矩阵、ROC曲线和PR曲线对比分析了卷积神经网络模型、支持向量机(Support Vector Machine，SVM) 模型、BP神经网络(Back Propagation Neural Network，BPNN)模型。实验结果表明，卷积神经网络模型的诊断性能更为优秀。

关键词: 油中溶解气体, 变压器, 故障诊断, 卷积神经网络