Keywords: railways, artificial intelligence, physical ai, predictive maintenance, condition monitoring, obstacle detection, human oversight, functional safety, edge computing.

Railway traffic is rising while the systems that keep trains apart and stop them in time rest on a safety base whose core logic was largely settled decades ago. Over the same period road transport has absorbed learning-based perception and driver assistance as a matter of course, yet the railway has been far slower to bring such methods into its safety-critical functions. This paper argues that the right response is neither to resist artificial intelligence nor to import it wholesale, but to adopt it wherever it strengthens human judgement and to avoid adopting it in ways that replace that judgement. We review where artificial intelligence already pays its way, chiefly in maintenance and inspection, and contrast this with the much harder problem of real-time safety and perception, where the margin for error vanishes. We then examine why current standards cannot yet certify learning-based components for safety-critical use, and why this favours a human-in-the-loop design rather than autonomy. An edge perception platform is presented as one practical embodiment of this principle, pairing a verifiable networking substrate with a learned layer that informs rather than overrules the driver. We conclude with the assurance steps still required.

download the full-text of this paper (PDF, 14 pages, 212 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description