Transformer Diagnostic Entity Recognition and Relationship Extraction based on Dilated Convolution and Bidirectional Graph Convolution Networks

Abstract

As the core equipment of the power system, the stable operation of power transformers is crucial for ensuring the safety and reliability of the entire power grid. The normal operation of transformers is an important guarantee for the safety of power transportation. Transformer operation and maintenance involve a large number of technical and operational guidelines. These data sources are extensive and the format is not uniform. Knowledge graph can integrate dispersed knowledge into graph data in the form of triplets, achieving unified and standardized management of data. By constructing a transformer knowledge graph, intelligent operation and maintenance work can be achieved, reducing labor costs and improving the efficiency and quality of operation and maintenance. In recent years, there have been some achievements in the construction of knowledge graphs for power equipment. Researchers have used natural language processing technology for entity recognition and relationship extraction to construct knowledge graphs. Describing concepts, entities, events, and their relationships in a structured manner in the power system to improve the efficiency of knowledge queries and assist decision-making. However, previous transformer knowledge graph construction methods did not consider entity recognition and relationship extraction of discontinuous entities, resulting in limited semantic representation of the knowledge graph. For example, " No abnormalities in voltage regulation, medium voltage and low voltage windings". The discontinuous entities "voltage regulation winding" and "medium voltage winding" in this sentence cannot be processed by previous transformer knowledge graph construction methods. However, previous research on discontinuous entities has only focused on entity recognition tasks, and previous methods have obvious deficiencies in the recognition ability of entity boundaries. In order to solve the recognition problem of entity boundaries of discontinuous entities, this paper adopts the learnable spacing dilated convolution to adaptively capture text interval information at different distances. Secondly, for the problem of extracting relationships between discontinuous entities with incomplete entity intervals, this paper uses a bidirectional graph convolutional neural network to obtain relationship extraction feature grids from text information and entity information. In response to the characteristics of discontinuous entities with incomplete entity intervals, the relationship extraction feature grid is obtained by adding dimension functions and repetition functions to obtain the relationship extraction tensor. This article inputs the relation extraction tensor into the linear layer to predict whether a triplet holds true. Experimental results have shown that our method achieves an F1 score of 89.01% on entity recognition tasks with discontinuous entities, which is superior to previous methods; And the F1 score of the accuracy of extracting relationships between discontinuous entities with incomplete entity intervals in this article is 85.22%, filling the research gap in the task of extracting relationships between discontinuous entities.