Tilting trains can travel at higher speeds without degrading the passenger comfort. However, these higher speeds lead to larger rail-wheel interaction forces. This larger forces have to be monitored in order to maintain general safety (e.g. to prevent derailment). The monitoring can be done with specially equipped passenger trains (On Board Monitoring „OBM“, as it is the case with the ICN types of trains), or with measurement trains like the VT612. These procedures are very cost-intense. Additionally, if limit violations are observed, it is often difficult to figure out what caused this effect.
This thesis tries to make a contribution to a cost efficient monitoring system. The goal is to learn the train characteristics with a deep learning network so that it can process track geometry data and predict the resulting forces. An optional task of the thesis is to learn abnormalities on the track geometry, in order to gain insights on the causes of limit violations.
- Develop deep learning algorithms for train characterization that predict the resulting rail-wheel forces based on the the track geometry
- Benchmark both predictive and computational performance of the developed algorithm
- Optional: Learn abnormalities of the track geometry
- Replacement of the cost intensive rail-wheel interaction measurements by the AI based system
- Indication of problematic areas of rarely observed tracks by the OBM (on board monitoring) passenger train. This is for example the case for tracks that are rarely run on both directions.
- Optional: Understand what causes limit violations. This would be a huge help for the maintenance of the tracks.
- Literature research and familiarization with topic
- Define metrics (Accuracy, F1 score, …)
- Datasource synchronization from the two trains (force measurements from the ICN with the track geometry from the „Diagnosefahrzeug“ DFZ)
- Dataset preparation (e.g. select data, normalize)
- Evaluation of different deep learning architectures
- Implementation of the most promising networks, and training on a server with GPU’s
- Iterative improvements
- Documentation of the work, including an 1 minute video summarizing the master thesis
The data comes from two different vehicles. The track geometry from the DFZ (Diagnosefahrzeug). The rail wheel forces are coming from the OBM ICN.
The model should take the speed of the train and the track geometry as an input variable and output the rail-wheel forces. In the next image, this dataflow is illustrated:
- Measured track geometry (Cant (Überhöhung), gauge (Spurweite), radius, twist (Verwindung), Schienenhöhenverschleiss, ….)
- Sampling Rate is 25cm on the tracks (@100km/h: ~100Hz)
- Data is available for the whole rail network of Switzerland and it is recorded twice a year
- Speed of the ICN OBM train
- Rail wheel forces
- The labeled training data will be rail-wheel forces from the ICN @1000Hz
The rail-wheel forces are the horizontal (Y) and vertical forces (Q) measured at each wheel. The index i represents the axle from 1 to 4:
We are open about the technology stack. Possibilities include:
- Python, Matlab, R
- TensorFlow, PyTorch, Keras, MXNet
Kind of Work
30% Deep Learning Theory
15% Theory Rail/Wheel Interaction
- Machine Learning & Deep Learning Basics
- Signal Processing
Time & Effort
Bachelor’s/Master’s Thesis, 1-2 Students