Reliable AI for the measuring wheel set

Sensors on the wheel

The ICN runs along the southern foot of the Jura - a beautiful, but also winding route. To ensure the safety of the route, the forces between the wheel and the rail are measured regularly. SBB uses an ICN equipped with sensors for this purpose. The sensors are attached to one of the bogies. On this bogie, measuring wheelsets on the rotating wheels measure the forces on the wheel. Further sensors are fitted to the bogie itself.

The sensors on the wheels are high-maintenance. It would be nice if the forces could only be measured or calculated from the sensors on the bogie. However, the physical relationship between the forces on the wheels and the accelerations on the bogie is complex. This cannot be easily converted analytically.

Machine learning algorithm

One option would be to use a machine learning algorithm. AI can deal with complex systems. It is trained with the available data and the physical relationships do not need to be known. However, the algorithm is a black box. We do not know exactly how the results are calculated. However, it is important for SBB's measurement runs that the results are correct at all times. So how can an AI algorithm be used on a measuring wheelset?

The AI is like a pilot

The AI can be compared to a pilot flying an airplane: The pilot can also make mistakes. Nevertheless, flying is considered safe. This is because the pilot is well trained, conveys trust through communication and adheres to standards and processes. In addition, the risk is kept as low as possible through both technology and procedures. All of this is regularly checked and documented in a comprehensible manner.

Trustworthy AI

Applied to a machine learning algorithm, this means that AI is trustworthy if the algorithm has been trained correctly, the results are explainable, standards and processes are adhered to and the risk is minimized. All of this must be traceable at all times.

For the measuring wheelset, the algorithm must therefore be trained specifically in those areas that occur in everyday life on the SBB track network. These are curves with radii of 250 m and larger, as well as lateral accelerations of up to 1.8 ^m/s2.

The calculated values must be considered plausible by SBB experts, i.e. they must be explainable. In addition, the data used to train the algorithm and which data model was trained with which codes is documented at all times.

Minimize risk

To minimize the risk, a risk analysis is carried out and, for example, the plausibility of the sensor data is checked. It could be that the sensor provides data, but the wrong values - the sensor drifts. As the train always travels the same route, this can be checked quite easily by comparing the values with previous journeys. If the values are consistently too low, as in the example below, the sensor probably needs to be calibrated or zeroed.

Trustworthy AI in the tilting train

The process described, which can be traced at any time, makes it possible to use AI algorithms to calculate the wheel-rail forces using acceleration sensors. The solution was developed for SBB together with the Bern University of Applied Sciences. The University of Applied Sciences developed the original algorithm, while the developers at SCS modularized the code, cleaned it up and integrated it into the automated data processing platform RCM-PP. The AI algorithm is currently being tested for reliability in a test phase. The measuring wheelsets - the sensors on the rotating wheels - are still being used occasionally for validation.