EP-BERTGCN: A Simple but Effective Power Equipment Fault Recognition Method

Abstract

With the advancement of China’s State Grid in recent years, text-based power equipment fault recognition has become an essential tool for power equipment maintenance. The task suffers from the domain gap that exists between the electric power domain and the general natural language processing domain. To improve the recognition performance, we proposed a method that combines pre-trained Bidirectional Encoder Representations from Transformers (BERT) and Graph Convolutional Network (GCN), i.e., Electric Power -BERTGCN. Our EP-BERTGCN first builds the graph among documents and words within documents based on pre-trained BERT. Then, the two softmax outputs with pre-trained BERT and GCNs are combined for final classification results. Extensive experiments show that our method outperforms previous baselines.