Fault diagnosis of power transformers using graph convolutional network
Wenlong Liao, Dechang Yang, Yusen Wang, Xiang Ren
Abstract
Existing methods for transformer fault diagnosis either train a classifier to fit the relationship between dissolved gas and fault type or find some similar cases with unknown samples by calculating the similarity metrics. Their accuracy is limited, since they are hard to learn from other algorithms to improve their own performance. To improve the accuracy of transformer fault diagnosis, a novel method for transformer fault diagnosis based on graph convolutional network (GCN) is proposed. The proposed method has the advantages of two kinds of existing methods. Specifically, the adjacency matrix of GCN is utilized to fully represent the similarity metrics between unknown samples and labeled samples. Furthermore, the graph convolutional layers with strong feature extraction ability are used as a classifier to find the complex nonlinear relationship between dissolved gas and fault type. The back propagation algorithm is used to complete the training process of GCN. The simulation results show that the performance of GCN is better than that of the existing methods such as convolutional neural network, multi-layer perceptron, support vector machine, extreme gradient boosting tree, k-nearest neighbors and Siamese network in different input features and data volumes, which can effectively meet the needs of diagnostic accuracy.