From Inspection to Action: Probabilistic LLM-Based Planning for Resource-Aware Railway Maintenance

High-speed rail is becoming the main transport option for medium-distance travel, thanks to its efficiency and convenience. However, the increased frequency of operations causes significant infrastructure wear, raising both the requirements and costs of maintenance. For these reasons, the AIRIS (Autonomous Intelligent Railway Inspection System) project, jointly developed by COPASA and the University of Vigo, aims to automate inspections using onboard optical sensors and artificial intelligence tools for the autonomous characterisation of railway infrastructure. The system includes LiDAR sensors as well as global-shutter cameras for the automatic detection of track defects, such as damaged concrete sleepers, deformations in the ballast bed, and issues with fastener elements. All this defect information is integrated into a Large Language Model (LLM) decision-making framework, which, based on pathology analysis data from the sensors, generates maintenance plans that account for factors such as resource availability, damage severity, and uncertainties in defect detection. The output is a user-friendly tool that provides maintenance operators with an inspection plan and recommended maintenance actions. This work presents the main results of the AIRIS project, focusing on the system's outcomes and uncertainties, as well as a quantitative analysis of the proposed decision-making framework, evaluating the reliability of LLM-based solutions for planning and decision-making in railway maintenance.

Keywords: Railway Maintenance, Artificial Intelligence, Operation Planning, Large Language Model (LLM).