Machine learning brings more safety and availability to rail traffic.

In order to be able to recognize the wear of railway switches – the most critical elements of the railway infrastructure – in advance, intelligent forecasts are carried out with the help of large amounts of sensor data. In the research project “Intelligent Turnout Performance Prognosis 4.0 (iTPP4.0),” methods close to the basics were developed to enable an intelligent switch for railway traffic. This should be able to reliably predict future wear, tear or defects of the points at any time. To this end, machine-learning algorithms derive decisions on planned maintenance from railway infrastructure sensors in a self-learning manner.

Switches: Critical elements of the railway infrastructure

An essential component of the railway infrastructure is the switch. Reliable switches make a significant contribution to the availability of the entire network. There are more than 10,000 of them in the Austrian network. The turnout service life is around 25 years, with a turnout having to be serviced several times during its service life in order to function trouble-free. If a turnout does not reach the safe and proper position in the specified time after the signal box has issued a setting command, for example, due to insufficient pressure in the turnout setting device, this turnout has to be maintained in an unscheduled manner. In the event of a malfunction or unplanned standstill, trains can only pass the points at low speed or not at all. Delays can occur due to the necessary “slow speed” points. The availability of the network is restricted.

The intelligent switch

Railway switches are already equipped with a wide variety of force, displacement and pressure sensors. To date, a reduction in accidents of approx. 30 percent has been possible by evaluating the sensor data. In order to determine even earlier when a turnout causes a malfunction, atypical sensors have now been used for the railway infrastructure. These include ultrasonic sensors, microphones, optical sensors for detecting vibrations and vibration sensors.

External influences such as the environment, climate and weather are also taken into account. The resulting large amount of data is collected centrally and processed into maintenance forecasts using mathematical algorithms and machine learning approaches. Thus, wear, tear or faults can be predicted ahead of time and malfunctions can be prevented.

Machine learning

The challenge lies in the recognition of relevant patterns of the data sources in order to determine the state of wear of a switch. Training data from existing points were used for this purpose and the fault cases detected by the existing point diagnostic systems were analysed and categorized. This data served as a basis for the development of intelligent forecasts.

RISC Software GmbH worked on the project “iTPP 4.0” in close cooperation with voestalpine SIGNALING Zeltweg GmbH. It was funded within the framework of the FFG project with the FFG number 855345 and by the strategic economic and research program “Innovative Upper Austria 2020,” the economic strategy of the State of Styria 2020 and the research strategy of Styria.

Diesen Beitrag mit Ihrem Netzwerk teilen: