A fault knowledge graph construction method for train-controlled on-board equipment

In order to extract effective information from the complex fault data of high-speed railway train-controlled on-board equipment, as well as to improve the efficiency of processing high-speed railway accidents and querying fault knowledge, a fault knowledge graph construction method for train-controlled on-board equipment is proposed. Firstly, the deep learning model BERT-BIGRU-CRF is introduced, which is used for entity recognition. Secondly, a rule-based method is used for relationship extraction. Then, the cosine similarity algorithm is used for entity alignment operation for entities with similar textual expressions. Finally, the ternary composed of entities and relationships is stored in the Neo4j database. The results show that for entity recognition, comparing with the mainstream model BERT-BILSTM-CRF with the proposed model under the same dataset, the precision P, recall R and F1 value are improved by 2.58, 1.7 and 2.14 percentage points respectively; and there are rich fault data in the final constructed graph, including there are 946 entities and 3 527 relations; and in addition, the fault knowledge queries can be efficiently performed by using the Cypher statement of the Neo4j.