Representation and Extraction of Diesel Engine Maintenance Knowledge Graph with Bidirectional Relations Based on BERT and the Bi-LSTM-CRF Model

The effective and efficient maintenance of diesel engines in a power plant highly relies upon the knowledge accumulated in previous maintenance and overhaul practices. A large part of this knowledge exists in maintenance reports in the form of engineering know-what, know-how and know-why with complex elements and relationships. As such, the need of supporting automatic extractions of such knowledge using state-of-the-art machine learning and NLP techniques is raised. This paper proposes a framework for the presentation and extraction of the knowledge graph from unstructured text in maintenance reports. Different from previous work, this framework supports extractions of bidirectional relations through a novel combination of reports prepossessing, the BERT model and the Bi-LSTM-CRF model. Specifically, the BERT model is applied to get character vectors and the Bi-LSTM-CRF model is applied to realize automatic extractions of entities and relations. We use an application named Neo4j to construct and storage knowledge graphs, and an application named Protégé to construct bidirectional relations. All the results from users' search is ranked by a scoring formula developed by Yahya et al. Conducted experiments over both public and private data set show that the proposed framework achieves good performance and hence provide a promising solution.