Predictive Alarm Detection
SBB ZKE measures the bearing and wheel temperatures of passing trains on the open track. SCS has trained an AI model that can predict applied brakes based on measurements from several consecutive measurement arrays. This makes it possible to predict an alarm before it occurs.
-
Initial situation
The SBB operates measuring systems on which brakes that have not been released (fixed brakes) are detected. In the case of the Gotthard base tunnel, it is advantageous if an imminent fixed brake can be predicted before the tunnel is entered, as a malfunction in the tunnel would be very inconvenient from an operational point of view.
-
Solution SCS
SCS has trained a machine learning model that can predict fixed brakes. On the basis of measurements from several successive measuring systems, the beginnings of a malfunction can be detected at an early stage.
-
Added value
The study has shown that the probability of successful early detection of fixed brakes is high. Affected trains could thus be pulled out of traffic before entering critical sections of the line. This prevents the blocking of the track section.
Project insights
Within the framework of the ZKE (train control equipment), the bearing and wheel temperatures of the passing trains are measured on the open track with the help of HFO systems (hot box and fixed brake detection). If excessive temperatures are measured, an alarm is triggered and, if necessary, intervention is initiated (reduction of speed, stop at an intervention station, inspection of the vehicle by the locomotive driver, rectification of the problem). Brakes that are applied (fixed brakes) or wheel bearings that are too hot (hot runners) cause damage to the rolling stock and can lead to axle or wheel breakages and thus to derailments.

In certain situations, it is advantageous if it is known before limit values are exceeded that a critical condition is imminent. This is particularly important on sections of the rail network where it is not possible to stop the train for a longer period of time due to the network topology (e.g. in the Gotthard base tunnel). If an exceeding of a limit value is detected on such a section, the train is only allowed to run at reduced speed and thus blocks the track for following trains.
When predicting an alarm, a challenge that should not be underestimated is that every axis is examined with the algorithm. The algorithm must therefore be very sure of an alarm in order to reduce possible false alarms. If an alarm is predicted for an axis without a problem, the resulting unnecessary intervention is associated with high costs. It follows that the proportion of axes that lead to an alarm even though there is no alarm must be around one millionth by orders of magnitude. But nevertheless, the real alarms should be detected.

A study conducted by SCS has delivered promising results for the prediction of fixed brakes. With the help of an AI model(machine learning), it is predicted whether an alarm caused by a fixed brakeman will occur on the next HFO system based on the measurements of previously used HFO systems. Such an approach is made possible by the fact that SBB stores the measurements of the HFO installations and their assessment. Over the years, this has created a database that is suitable for training machine learning approaches.
The close cooperation between SBB and SCS has allowed the knowledge of SBB's technical experts to be incorporated into the design of the early detection system. A good understanding of the system helps to develop meaningful parameters, which are ultimately fundamental to the success of a machine learning approach.
Do you have questions or would you also like to look into the future through your data? Get in touch with Florentin Marty!
