Train Localization with Magnetic Field Distortions and Bayesian Methods

Shifting transportation of persons and goods from the road to the rail has the potential to significantly reduce the greenhouse gas emissions of the transportation sector. For the shift from road to rail it is crucial that the capacity of the track networks is increased to serve the increased demand. In order to increase the capacity either new infrastructure has to be built or the existing infrastructure has to be used more efficiently. When comparing the two options it seems to be preferable to increase the capacity of the existing infrastructure since new tracks require a large amount of space, money, and time for the planning and building phase. In contrast, increasing the capacity of the existing tracks can be achieved by reducing the safety distances between consecutive trains from the absolute braking distance, that can be up to kilometers, to the relative braking distance. To enable the capacity increase, the degree of automation in railway traffic has to be increased and the required technologies have to be developed. One key technology for automation is absolute train localization that is essential to control the traffic in the network and the maneuvers of individual trains. ... mehrThe key role of the absolute position for automation of railway traffic is in contrast to its role in the automation of road transportation, in which the absolute position is not essential to maintain a safe operation of the vehicles. The importance of the absolute position for railways is a result of the long braking distances, which exceed the measurement range of sensors such as radars, laser scanners, and cameras. In contrast to road transportation, were the braking distances are much smaller, for railways it is thus not sufficient for safe operations to just rely on a supervision of the surrounding environment with the before mentioned sensors. In this thesis, the focus is therefore on the development and the evaluation of a novel train localization system to enable automated train operations.

The developed localization system is based on position dependent distortions of the magnetic field along a railway track. The distortions are a result of the interaction between the Earth magnetic field and the magnetic material in the railway infrastructure or buildings close to the track. For localization first, a map of the magnetic distortions is created that relates to each position on a track the corresponding magnetic vector at that position. Second, a train moving within a mapped track can estimate it absolute position by comparing multiple measurements of an on-board magnetometer to the magnetic vectors in the map.

In this thesis, general properties of the magnetic field in the railway domain are investigated and based on the findings suitable localization algorithm are derived. Concretely, recursive Bayesian filters are developed and evaluated that enable magnetic field-based train localization in real-time. The developed filters explicitly account for disturbances in the magnetometer measurements and address the problem of magnetometer calibration. Magnetometer calibration plays a central role when it comes to the applicability of magnetic field localization in a practical setting, because without calibration the comparison of the magnetometer measurements to a prior created magnetic field map becomes more difficult or even impossible. The calibration not only depends on properties of the sensor but also on the train in which the sensor is mounted. The dependency on the train is a result of the interaction of the magnetic material in the train and the external magnetic field. Without calibration it might be necessary to create an individual map for each train that we wish to localize. To avoid this, a novel calibration method was developed that, in contrast to commonly used methods, is also applicable to trains. Common methods require a rotation of the magnetometer in space while mounted on the target platform. For calibration, the rotation has to be performed in a homogeneous field and has the goal to ensure the observability of all parameters. In the concrete example of magnetic train localization, this means the sensor has to be rotated while mounted on the train. Due to the large extend and the high weight of trains this would require a considerable effort. In this thesis, therefore a Bayesian filter was developed that estimates the calibration parameters and the train position simultaneously without the need of rotating the train. The proposed filter uses that the observability of the parameters can be ensured also by moving through a inhomogeneous magnetic field rather than a rotation in a homogeneous field.

In addition to the development of concrete filter algorithms, a lower bound for the achievable localization accuracy was derived during the course of the thesis. The bound is based on the Bayesian version of the well-known Cramér-Rao lower bound and can be used, e.g., as a benchmark for the filter algorithms. The challenge in the calculation of the bound for magnetic localization lies in the missing analytical model of the magnetic field. The magnetic field is only known from measurements and the derivation of an analytical model is practically impossible. Even with perfect knowledge of the environment and its material properties, the interaction between the Earth magnetic field and the environment is too complex to find a closed form model. To close the gap between theory and the real-world, we propose in this thesis to fit a Gaussian process to the measurement data to obtain a probabilistic model of the magnetic field that fits nicely into the Bayesian framework of the lower bound.

All developed methods are tested and evaluated based on measurement data. The measurement data for the evaluation was collected with three different trains driving in two different track networks. The evaluation shows, that the methods developed for magnetic train localization enable meter-level accuracy and that magnetic train localization can potentially contribute towards a higher degree of automation in railway traffic.