Towards Data-driven Metro Rail Maintenance Following the MLOps Paradigm

Abstract: Railway maintenance, particularly in curved sections, is a complex and costly operation requiring effective solutions to minimise wear. Machine Learning (ML) models were developed to monitor curved tracks using accelerometer data for lubrication level prediction, sensor fault detection, and outlier analysis. To support this, a monitoring system designed within a Machine Learning Operations (MLOps) framework was implemented. While vibration alone proved inconclusive in predicting lubrication levels, a significant increase in outliers during lubrication system inactivity indicated potential rail issues. As a result, the solution offers near real-time insights, helping end users make data-driven maintenance decisions by presenting model outcomes.