Fault detection and failure prediction for railway networks

Staff Mentor:

prof.dr.ir. B. De Schutter (Bart)

Other Mentor(s):

K. Verbert, R. Babuska


Railway networks; Machine learning


High reliability and availability are important requirements for railway assets such as switches and track circuits. To satisfy these requirements, adequate maintenance is of paramount importance. To realize both high availability and low costs, an adequate maintenance schedule is crucial. For optimal maintenance planning, knowledge about the degradation of assets needs to be timely available, so that this knowledge can be used to organize and prioritize maintenance. To get more insight into the system degradation, various monitoring devices have been developed and installed. However, the measurement data are still not effectively used for maintenance planning. The aim of this research is to develop and validate methods to predict failures in railway switches or track circuits. Based on system knowledge and measurement data, we aim to predict the moment that a system fault leads to a functional failure. This information can then be used to optimize maintenance planning. Besides the time-to-failure, it is important to know the type of fault that leads to functional failure. This knowledge provides additional information about what actually is broken and what type of maintenance is required.

Railway track circuits:

In the Netherlands, track circuits are used for the detection of trains, i.e. to determine whether a certain part of the track is occupied or not. This information is used to control relevant signals (e.g., red signs, position of switches). Therefore, the track is divided into different sections, which are electrically separated. One side of the section is connected to a transmitter that sends a current. When the section is free, the current flows via the rails to the other side of the track where it is recognized by a receiver, indicating that the track is free. In the case that the section is occupied, the wheels and axles of the train short-circuit the section, and the current will flow via the axles and wheels back to the transmitter and no current is measured at the receiver, indicating that the track is occupied. This is how it works in
the ideal situation. In practice, the behavior can be deviating for different reasons, for example weather circumstances, rail or insulated joint defects, specific train problems.

Railway switches:

A railroad switch, is an electromechanical installation enabling trains to switch from one track to another at a railway junction. A switch consists of a number of components, of which the most important ones are:
Stock rail: The non-movable, diverging outer rails of the switch, against which the point blades operate.
Point blades: The movable parts of the rail, guiding the wheels of the train towards one of the directions
Switch motor: Motor driving the switch from one position to the other.

In the ideal situation, the point blades move smoothly to the desired position. In practice, it can happen that the movement is non-smooth, that the end position is not reached, or that the actual position is not correctly communicated to the signaling system.

Possible assignment:

The raw measurement data generally contains a lot of irrelevant and redundant information. Therefore, the original data should be transformed to another space. This space should be as compact as possible, but such that it contains all relevant information. Because the final goals are the identification of faults and the prediction of future failures, this space should be chosen such that the retained information is optimally suited for this purpose.

Overview of a track circuit (left) and a switch.

© Copyright Delft Center for Systems and Control, Delft University of Technology, 2017.