How SNCF uses AI to better anticipate train breakdowns

Thanks to thousands of sensors placed on 500 trains, a colossal mass of data is transmitted every day to a small team of engineers in order to generate models.

Attached to the SNCF maintenance center in Saint-Pierre-des-Corps, on the outskirts of Tours, a small, anonymous building houses around twenty statisticians whose mission is both simple and ambitious: to anticipate train breakdowns using artificial intelligence.

SNCF calls this “predictive maintenance”. Thanks to thousands of sensors placed on each train, a colossal mass of data is transmitted every day to the small team of engineers to shed light on the train’s behavior.

In Saint-Pierre-des Corps, the technical center is responsible for the maintenance of a large part of the trains in the Paris region (RER and Transilien) as well as TERs from different regions.

Thousands of data

“Our job is to process colossal data and extract relevant information from it,” explains Audrey Nze-Eyoun, engineer of the western engineering cluster (CIO) of the SNCF.

This engineer supervises a very specific series of trains deployed in France, the Regio 2N, of which 500 units are in circulation and equip TER, Transilien and RER D lines in the Paris region.

Each carriage of these trains, which will celebrate its 10th anniversary in October, transmits 7,000 codes every day to the Saint-Pierre-des-Corps team.

Temperature in the passenger information screens, performance of the air conditioning systems, size of the gap between the doors when opened… A colossal mass of data, which concerns all the components of the train, is transmitted every minute to detect and anticipate the slightest failure.

In front of the giant screen that compiles and formats this data in her office, Audrey Nze-Eyoun can, for example, see in real time the position of a train on the network as well as the number of people on board, including the percentage margin of error.

Graphs and tables that are cryptic to the uninitiated eye convey various signals, which must then be deciphered.

Zero breakdown objective

Thanks to this work of prediction and anticipation, “we avoid breakdowns and reduce maintenance costs”, says Guillaume Branger, head of remote diagnostic engineering at Saint-Pierre-des-Corps.

“For example, we can tell a driver ‘your door has a fault, but you can continue driving for three days, because it’s not too serious’. And we know that we have time to notify the technical center, order the necessary part, etc.”, he continues.

Thus, the trains undergo maintenance when necessary and not necessarily at regular intervals, with the risk of intervention being too late or, on the contrary, unnecessary because premature.

“On a fleet of Parisian trains, we have overall 30% more availability (of trains)” thanks to the work implemented since 2017 at Saint-Pierre-des-Corps, underlines Mr. Branger.

SNCF CEO Jean-Pierre Farandou has even set an ambitious goal: “zero technical breakdowns” within a decade, thanks to the use of artificial intelligence.

Eco-driving

The fact remains that only recent trains are equipped with sensors allowing this type of maintenance. The TGVs in circulation have far fewer, and are only installed by the SNCF after their manufacture.

But the arrival of the new generation TGV M from 2025 should make it possible to extend this program to high speed.

On September 4, the head of the railway group participated in the launch ceremony of a teaching and research chair at the École Polytechnique entitled “AI and optimization for mobility”, supported by the SNCF.

The group and its CEO believe a lot in artificial intelligence for predictive maintenance, but not only. Optimizing work schedules, improving the prediction of delays or giving driving instructions to drivers to limit consumption are some of the areas that SNCF wants to improve thanks to AI. For example, driving instructions have made it possible to save 7% to 12% of electricity, according to SNCF.