Data-Driven Approach for Predictive Maintenance of Metro System

Predictive maintenance is a method used to analyze the real-time and historical data from the various sensors deployed in critical areas of a machinery to detect the failures and their impact. This paper proposes a study to measure the effectiveness of supervised and unsupervised machine learning models for accurate and efficient detection of failures by using techniques such as anomaly detection, clustering, bagging, and ensemble methods. The study analyzes multivariate time series data of various sensors deployed on the Air Production Unit (APU) of the Metro Do Porto train system. The primary goal is to develop a flexible pipeline to be implemented in rail transportation industry with ease, The data-driven approach improves the maintenance practices and minimize inoperability time and save costs. LSTM was found to be the best performer with an accuracy of 99.8% accuracy and recall of 95%.