Front of a train

More efficient railway technology thanks to neural networks

Digital support is needed to ensure that transport policies become more climate friendly and that rail transport can fully exploit its potential. This is where the innovation ecosystem RailCampus OWL comes in. Researchers at Bielefeld University are contributing their know-how to make rail transport more efficient and reliable.

The street named Pionierstraße in Minden lives up to its name: this is where they are paving the way for a sustainable railway system. On the premises of DB Systemtechnik GmbH, research, teaching, and work is being carried out on the topics of automation, innovative maintenance, and networked transport logistics. A research hall with several railway tracks is being built on the campus, and the sophisticated inspection and testing infrastructure of DB Systemtechnik can also be used. In addition to Bielefeld University, the initiators of the project are OWL University of Applied Sciences and Arts, Hochschule Bielefeld – University of Applied Sciences and Arts, and Paderborn University together with the city of Minden and the district of Minden-Lübbecke, regional companies such as Wago and Harting, and Deutsche Bahn with DB Systemtechnik und DB Cargo.

Teaching is already underway. Since the winter semester, the first students are being taught in the bachelor’s degree programme ‘Digital Railway Systems’. And although research work has not yet begun, it is already clear where the academics from Bielefeld University want to use their expertise: ‘Even if we have not yet applied it to the railway system, we have plenty of experience in the field of intelligent technical systems and machine learning,’ says Professor Franz Kummert from Bielefeld University’s Faculty of Technology. He represents Bielefeld University on the board of the RailCampus OWL association.

Professor Dr.-Ing. Franz Kummert
Professor Dr.-Ing. Franz Kummert from the Faculty of Technology represents Bielefeld University on the board of the RailCampus OWL association.

Using automated analysis to detect wear and tear at an early stage

Identifying weak points before they become defects—that is the challenge. One place where Bielefeld University is devoting itself to this task is in the laboratories of the Research Institute for Cognition and Robotics (CoR-Lab)—albeit in a somewhat different field. One recently completed project aimed to use artificial neural networks to detect stains on patterned laundry. ‘There is a difference between detecting holes and stains on patterned laundry and detecting the state of wear in a wheel,’ says Kummert. ‘But the methods that are applied are ultimately the same and can be transferred to the railway system.’

The principle of spot recognition is simple. When we humans look at a picture, certain areas catch our eye first. These are what are known as the salient areas: special objects that stand out from their surroundings such as a red object against a blue sky. ‘The idea was to use neural networks for saliency detection on patterned laundry as well, because after all, a stain on patterned laundry also catches the eye. It is a salient area,’ Kummert explains.

Time-consuming preparatory work is necessary before the neural networks are able to recognize the stains independently: ‘I need thousands of images containing stains on which I have to manually highlight the stain areas so that they can be recognized.’ Although training is time-consuming, the method of visually inspecting objects with the help of neural networks can also be used for train maintenance: for example, when it comes to checking the condition of a brake lining. Kummert sees so-called predictive maintenance as an important contribution to making the railway a more attractive means of transport: ‘The probability of errors decreases and makes the whole process more reliable and efficient.’

Using automation to counter staff shortages

Another major field of the research project is the automation of freight and passenger transport. The vision is to have fully autonomous rail transport, because if you want to get more trains on the tracks, you have to get a self-organizing rail system off the ground. ‘This requires, for example, environmental perception to ensure that the train is on the right track and that the track is free of obstacles.’ Here as well, neural networks come into play: academics at the CoR Lab use them to develop multisensory platforms for the automated recognition of people, objects, and the environment.

And not only that: they are researching adaptive robot systems that can be used cost-effectively for full or partial automation in assembly or intralogistics. ‘Transferred to the railway system, we can therefore also contribute our expertise to the loading and unloading of goods, for example with our research on robot arms.’  Kummert also sees a need for the Bielefeld academics in the area of scheduling: ‘More efficient planning is required regarding at which point which wagon is attached to which goods train so that the cargo arrives at the destination station as simply and quickly as possible.’ All these measures are not intended to replace human labour, ‘it is more the case that we want to counteract the limited availability of personnel.’