Correcting Maintenance Reports Using Large Language Models

Maintenance reports are a rich source of information for reliability engineering and asset management. They typically contain unstructured text and categorical information, which can be used to estimate quantities of interest for maintenance decisions, such as Mean Time Between Failures (MTBF) and Mean Time to Repair (MTTR) of system components. However, the extraction of knowledge from these reports is challenged by linguistic variability, technical jargon and frequent human errors committed when filling the categorical fields. To address these issues, the present work proposes a novel, two-step method based on Large Language Models (LLMs) to automatically identify the units, subunits and components involved in maintenance interventions. First, an LLMbased model ( L L M₁ ) identifies possible quality issues in the unstructured text, flagging reports that suffer from brevity, lack of context, excessive use of technical terms, internal incoherence and inconsistencies with the categorical fields. Then, based on an equipment taxonomy, a second LLM-based model ( L L M₂ ) systematically identifies the units, subunits and components involved in the maintenance interventions from reports of low quality. The overall method is validated considering a real-world case study of maintenance reports from a mining haul truck fleet. The obtained results demonstrate that: i) L L M₁ successfully identifies quality issues in a significant fraction (36%) of the reports; ii) L L M₂ effectively corrects erroneous categorical assignments of the units, subunits and components involved in the maintenance intervention in low quality reports.

Keywords: Reliability engineering, Maintenance, Knowledge extraction, Safety assessment, Decision support, LLM.