Extracting Reliability and Maintenance Knowledge from Maintenance Reports of Freight Transport Trains: A Methodology for Annotation based on Ontology and SpERT

We consider the problem of extracting information from repositories of maintenance reports of freight transport trains, aiming to identify factors influencing malfunctions and failures, and assess the effectiveness of maintenance activities. We propose a methodology for automatically annotating maintenance reports, which involves assigning semantic labels to the words of the reports and identifying the relations between them. The conciseness of the texts and the extensive use of technical language pose significant challenges, which are overcome by combining an industrial maintenance ontology with the Span-based Entity and Relation Transformer (SpERT) method. Specifically, SpERT is fine-tuned in two stages: initially on a large dataset of maintenance reports from other industrial sectors, and, then, on a limited number of manually annotated maintenance reports of electrical freight transport trains. The obtained results show that the proposed methodology successfully identifies entities and relations in maintenance reports of freight transport trains.

Keywords: Maintenance reports, Ontology, Annotation, Entity and relation extraction, Natural language processing.