Information Extraction from Aircraft Maintenance Records Using Large Language Models

This paper analyzes the characteristics of aircraft maintenance records and develops an information extraction Agent using general large language models (LLMs). The Agent demonstrates effective extraction performance across maintenance records from different aircraft models and airlines, supporting various LLMs of different scales and capabilities. Utilizing this Agent, we extracted key information from 17,544 maintenance records of COMAC ARJ21, forming a dataset used for relationship finding and completion tasks in the ARJ21 maintenance knowledge graph. Despite its effectiveness, challenges remain in extracting key information from complex troubleshooting measures involving multi-step isolation procedures and dealing with non-standard abbreviations and multiple names for the same component. Future work will focus on refining the Agent, further subdividing processing steps to handle more complex maintenance records, and providing editable prompt templates and a visualization interface.