AI-Driven Predictive Maintenance for Fare Collection Systems

This technical report presents the development and implementation of PartLlama, a predictive maintenance model designed to automate the identification of replacement parts for fare collection gates. The report outlines the operational challenges faced by transportation systems, particularly the inefficiencies in manual incident triage processes that lead to delays and increased operational costs. It details the methodology employed, including the creation of a robust data preprocessing pipeline and the application of Low-Rank Adaptation (LoRA) for fine-tuning a large language model (LLM). The report documents the model's performance, achieving high classification accuracy across various input styles, and emphasizes the potential for reducing on-site engineer visits by up to 70%. Furthermore, it discusses the implications of transitioning from reactive to proactive maintenance strategies, leveraging AI to enhance operational efficiency in public transportation systems. The findings suggest a significant advancement in the automation of maintenance workflows, contributing to improved service reliability and passenger satisfaction.