The digital railway switch- Digitalized railway switches for the future (DigiSwitch)

Researchers: Jan Lundberg (PL), Taoufik Najeh

Sponsor: Formas, Bombardier, Trafikverket, Infranord

Duration: 2019-2021

Goal:

The overall aim of this project is was increase the transport capacity in the long term and to indirectly reduce the environmental impact in Sweden by reducing operational disruptions at railway switches and thereby increasing the punctuality of rail traffic. This provides a better utilization of environmental transport infrastructure in Sweden, with a focus on increased accessibility for the interchanges as a whole.

The priority switch components are the ones that cause the most interference, which is the switch blades, crossing and other parts. The goal is to develop new knowledge and skills that can later be developed into a demonstrator in the form of an auxiliary equipment inside the point machine for switches in real traffic. This auxiliary equipment that in real-life traffic should have the ability to, using IoT, AI and smart algorithms with pattern recognition, provide continuous information about the state of the switch to the Swedish Transport Administration and its maintenance contractors (for example Infranord). This means that maintenance can be planned and carried out in a better way to increase the availability of switches and thus reduce traffic disruptions.

The project has developed an AI demonstrator that shows that it is fully possible for pattern-recognition neural networks to correlate measured vibrations in the point machine, against actual wear and tear and damage in different positions in our test switch with test wagon, which also forms part of the demonstrator. This opens up for full-scale trials in real traffic in the future. If this future project is also successful, the last remaining step is that this new technology can be product-adapted by companies to become commercially viable for the Swedish Transport Administration's switches. If this happens, the Swedish Transport Administration and related maintenance contractors will have access to a completely new technology that can effectively reduce the disruptions in train traffic by considerably facilitating the planning of preventive maintenance of the switches. The demonstrator is patent applied.

Project status and results:

The project is completed and some of the most important results are shown below:

• Squats: Within a distance of 14 m approximately symmetrically around the point machine, it was quite possible to detect squats with a diameter of 50 mm and with a depth of 0-4 mm, as well as 1-4 mm, over the entire switch.
• Crossing: With a speed of 0.05 m/s in the switch, it was not possible to detect damage in the crossing, either by means of acceleration amplitudes or an FFT. However, with a speed of 0.6 m/s, it was entirely possible to detect and separate levels of wear of the order of 1-7 mm at a distance of 63 mm, 153 mm and 243 mm from the crossing tip. With a speed of 0.6 m/s, it was not possible with an FFT to detect and separate such levels of wear at a distance of 63 mm, 153 mm and 243 mm from the tip.
• Crossing: LSTM algorithm with 7 vibration parameters show 82 % prediction accuracy for 60 % unseen data.
• Middle rail: Levels of side wear of the order of 0.5-8 mm were possible to detect and separate from each other over the entire rails.
• Middle rail: LSTM algorithm with 7 vibration parameters show 89 % prediction accuracy for 50 % unseen data
• Switch blade: Levels of side wear G of the order of 3.5-4.3 mm at a position 200 mm from the tip of the tongue were possible to detect and separate from each other.
• Switch blade: Levels of side wear - H of the order of 3.14–0.48 mm between a position 200 mm from the tip of the tongue and a position 228 cm after the centre of the point machine were possible to detect and separate from each other based on the maximum amplitudes and the number of amplitudes over 0.1 g. The results point to a relatively strong nonlinear dependence.
• Switch blade: Levels of side wear + H of the order of 0.80-3 mm at a position 228 cm after the centre of the point machine were possible to detect and separate from each other. Moreover, levels of heavy rail height wear of the order of 4 mm were possible to detect and separate from other.
• LSTM algorithm with 7 vibration parameters show 85 % prediction accuracy for 60 % unseen data
• Support rail: Levels of rail height wear of the order of 3 mm on support rails were possible to detect and separate from each other.
• Wheel plates: Wheel damage of the order of 12-60 mm was detectable. The results point to an approximately linear dependence.
• Trailed switch: The risk of a trailed switch occurring was possible to detect for all combinations of carriage position and position in the switch blade.
• Tamping error: Tamping errors at the deflection device could be detected by measuring the number of amplitudes above 0.15 g when the relative difference in the height position between rails was of the order of 10 mm.
• The AI demonstrator based on LSTM is capable of predicting wear and damages in the S &,  C with probability 41-82 %, dependent on the type of damages and the position of the damages in relation to the point machine.