Software safety analyses require knowledge of their algorithms, while Machine Learning (ML) models remain software ‘black boxes’. However, there is an approach that makes it possible to ensure safety requirements are met while benefiting from advances in Artificial Intelligence.
The principle: combine an ML model with an independent safety controller. ML proposes a solution (for example, detecting that a railway track is free based on images), and the controller verifies its compliance with safety rules (geometry of the path, gauge, parallelism, curvature, etc.).
Safety thus relies solely on the controller, avoiding the need to justify the internal functioning of the Machine Learning model. A proof of concept based on a U-NET model demonstrates the relevance and reliability of this approach.
This combination paves the way for the use of Artificial Intelligence in critical railway functions, while complying with the highest safety standards.
Mr Frédéric HENON (UIC) France presented this work at the conference: ‘12th UIC WORLD CONGRESS ON HIGH-SPEED RAIL’ from 8-11 July 2025 – Beijing, China.
This study was conducted by CLEARSY under the direction of the Operations and Safety Department of the International Union of Railways UIC.
Presentation here : 12WCHSR_NMSD_FH_final_20