MACHINE LEARNING FOR REAL-TIME TRAIN ARRIVAL DETECTION IN RESOURCE-CONSTRAINED DEVICES

Abstract

Early train detection is an essential component to provide both safety and operational efficiency within railway systems. As railway infrastructure continues to expand, the demand for affordable and portable train arrival detection and notification (TADN) solutions is rising rapidly. There is an increasing need for a proactive, budget-friendly, and easily deployable plug-and-play system that can function independently, deliver real-time alerts, and require minimal maintenance. This study introduces an intelligent real-time TADN system designed as a sensor network that leverages railway vibrations and Machine Learning (ML) to predict approaching trains and alert workers. Specifically, the study explores the performance of an ML prediction model within a vibration sensing device operating under resource constraints. Various well-known supervised ML estimators, including curve fitting regressors, decision trees, random forests, gradient boosting, and sequence-based models, were evaluated on a dataset collected from an actual railway track. In addition to maintaining a low error in train arrival time prediction, the system achieved a rapid response time of less than 3 seconds, including 2 second data buffering, using the available resource-constrained off-the-shelf equipment.