AI video analytics for railway: forecasting safe train operation with reinforcement learning

GNS Technology has been applying AI video analytics and machine learning to railway operations in Hong Kong since 2021. This case study explains one of our flagship projects: forecasting the allowable time of train operation using open weather data from the Hong Kong Observatory (HKO) and data collected from SACEM dataloggers.

The challenge

Railway operation control centres (OCC) need to know how environmental conditions affect safe train operation. Strong wind, for example, can limit how long and where trains may run safely. GNS was engaged in a 3-year consultancy to implement AI video analysis for the railway, and designed a system that predicts the time limit that trains can achieve under different environmental thresholds.

What we built

• Collect real-time railway operation status from SACEM dataloggers installed on 2XX trains, plus historical data since 2020
• Ingest open weather data from HKO, including real-time wind speed
• Find the threshold conditions for train operation against environmental conditions on different lines
• Compare collected data and use a reinforcement learning model to predict the achievable time limit
• Expose results through an API for third parties and a cloud-based web application for public enquiries

Outcomes

• The trial system can be integrated into the existing SACEM monitoring system used by MTR, for direct evaluation by the OCC and design teams
• The model targets 80–90% accuracy within ±15 minutes, giving OCC teams an additional reference for operation time and shutdown planning
• After internal testing, results can be published through a public cloud web application

Learn more

This project is registered on the EMSD I&T solution platform as reference S-1662: Forecasting allowable time of train operation using open weather data from HKO and collected data from SACEM datalogger using reinforcement learning model.

This case study was first published on the EMSD I&T platform in January 2024.