The TU Berlin is participating in the project »Berlin Digital Railway Operation« (BerDiBa). Within the next four years, technologies for automated rail traffic are to be developed and tested. Artificial intelligence is playing an important role.
First and foremost, the Daimler Center for Automotive IT (DCAITI) at the Technical University of Berlin is creating algorithms for the automated vehicle's perception of its surroundings. The Fraunhofer Institute FOKUS is also involved. In addition, the Department of Electronic Measurement and Diagnostics at the TU is developing an automatic planning tool for train maintenance appointments in collaboration with the Zuse Institute.
Digital rail operations encompass three aspects: driverless trains, automatically scheduled maintenance, and the ability to control remotely in situations where autonomous travel is not possible due to weather, for example. The aim is to make rail transport safer and processes more efficient.
In the project, researchers are training deep-learning algorithms using specially prepared video films. In this way, the trains can learn to use sensors such as radar and lidar – to correctly detect their surroundings. In the process, the trains must correctly recognize both static objects such as signaling systems, tracks and stations, and dynamic objects such as people, animals, vehicles and other trains.
Long-term, continuous observation of the route is particularly important: Are there branches hanging over the track that could soon break off? Has a hole suddenly appeared in a fence through which animals or children could get onto the tracks? Up to now, train drivers have been instructed to watch out for such things.
Prof. Dr. Clemens Gühmann from the Department of Electronic Measurement and Diagnostic Technology at TU Berlin is to work on the so-called clearance gauge of a railroad line – in other words, the space that must be kept free of vegetation and other objects. Using camera data over the seasons, neural networks predict the further development of plant growth here. This can be used to efficiently plan when and where to trim back.
»We also use deep neural networks to obtain information about the condition of train components from data from sensors on the train,« explains Gühmann. For example, he adds, the power required by a door's actuator motor can change over time. »From a lot of individual data like this, we estimate when components need to be replaced,« Gühmann says. In collaboration with the Zuse Institute in Berlin, this is being used to create a planning tool that automatically and efficiently sets the train's maintenance dates. On the one hand, this allows repairs to be combined and costs to be saved – and on the other, it increases the reliability and thus the punctuality of the trains.
Technische Universität Berlin