top of page

SBB/CFF/FFS adopts AI to improve railway maintenance

LatticeFlow collaborates with SBB, Siemens, and ETH to enable safe, efficient, and cost-effective AI-based railway maintenance.

Swiss Federal Railways (SBB) has one of the world’s most dense railway networks spanning over 7,500 km of tracks with 300 tunnels and servicing over 1.25 million passengers every day. The intensive usage requires high investments into operational maintenance to ensure track security and safety.

Track safety is traditionally done by manually inspecting the rails, a task requiring significant human labor and people deployed in dangerous environments that are sometimes difficult to access. SBB has set out to revolutionize railway maintenance by automating inspections using the latest generation of Artificial Intelligence (AI). A specially dedicated train instrumented with cameras will collect high-quality image data of the rails, which will then be processed by AI to identify any rail defects. This will make railway inspections efficient, safe, and cost-effective.

Dr. Ilir Fetai and Dr. Andre Roger, who lead the Center of Competence in Machine Perception at SBB, are dedicated to making these plans a reality, reinforcing SBB’s position as the innovation leader in the railway domain. Dr. Fetai said:

“Artificial Intelligence (AI) is one of the core topics at SBB, as we see a huge potential in its application for an improved, intelligent, and automated monitoring of our railway infrastructure.”

Railway inspection is, however, a safety-critical task. How can we ensure that the new AI-based system is reliable? AI models often work well in lab environments, but perform poorly when deployed in the wild. This has thus far limited their use in such safety-critical tasks. Before SBB hands off the inspection task to the “eyes” of the new AI system, it needs to ensure that the trained AI models correctly identify rail defects under different environmental conditions. Lens scratches, raindrops or snow on the rails, and other variables that affect the image quality must not undermine the AI system’s reliability.


bottom of page