Smart railways: on track to a digital future?

The new breed of digital trains – not to be confused with the airport shuttles that we have been riding for years, or the Docklands Light railway in London which travels on a confined track – will share the tracks with conventional driver-operated trains, in a much more complex environment.

Trains are already an efficient method of moving people and goods around but can be 15 per cent more energy efficient when driven by an AI, be more punctual, and can carry 30-50 per cent more passengers or freight, by reducing spacing between trains, all on existing infrastructure. Rural public transport economics can also be transformed by making lines less expensive to operate and enhancing existing safety levels.

The last few months have seen a flurry of announcements that, at first sight, might appear to herald the imminent arrival of autonomy on the rails. This is not the case, but progress is being made. In November 2021, a Japanese bullet train operated by East Japan Railway Co travelled 5km and reached 110km/h travelling as an automatic or remotely controlled train running without driver intervention. The company proudly announced that it stopped 8cm from its intended stopping point, well within the allowable margin of 50cm.

Siemens Mobility and Deutsche Bahn tested the first passenger-carrying automatic train, which entered service on the S-Bahn in Hamburg, Germany, in December 2021, sharing the tracks with driver-operated trains. In the only truly autonomous part of the S-Bahn operation, the trains can shunt in their depot and turn around without human intervention.

Finnish company Proxion plans to test ‘autonomous’ freight trains this year. The freight concept train will initially be aimed at steel and forestry operations moving material from source to ports.

None of these are autonomous trains which would be making operational decisions for themselves but are merely piloted remotely from a control centre, a bit like a giant train set, but all these projects are part of the necessary research and development journey to true autonomy.

Two projects stand out in terms of real-world testing of autonomous trains and light rail systems. Siemens Mobility’s autonomous tram work in Potsdam, Germany, is one, while the other, which is currently running autonomous train tests in northern France, is a five-year Railenium-led project, begun in 2018 and involving a consortium that also includes SNCF (French railways), Thales, Bosch, SpirOps (a Paris-based AI company), and Alstom (formerly Bombardier). Railenium is France’s technology research institute for rail.

Matthias Hofmann leads the Assisted and Autonomous Light Rail programme at Siemens Mobility in Germany and has been teaching trams to drive for years. Siemens has had an experimental tram driving autonomously (but under supervision) since 2018, and the tram has clocked up 20,000km along a 13km section of the Potsdam tram network. The prototype tram is extremely heavily sensored-up for research purposes.

An array of 13 cameras on the front and sides of the experimental vehicle combine their data with three forward-facing radar detectors (which detect the metal in other trams and cars),while three lidar scanners calculate the precise speed, trajectory and distance of any object within the immediate environment – particularly important in the inner city where pedestrians, dogs and cyclists crossing the tracks may or may not obey the general rules of road safety.

Algorithms process all this real-world data to make decisions about whether to sound a warning bell, or brake the tram if necessary, but an attendant is always on hand to intervene. “The biggest goal is to always arrive at the correct assessment of action in the most complex of traffic situations in pedestrian zones or at confusing intersections,” says Hofmann. “Achieving GoA4 [full autonomy: see box] and dispensing with an attendant is many years away, but we have made significant progress toward autonomy in the low-speed, semi-closed, controlled environment of the tram depot.

“Once the driver has de-boarded, our trams can intelligently navigate to the washing or sanding bays, then self-park and shut down, starting up to order and arriving ready for the driver to board in the morning,” he adds. “Commuter lines – at least the sections outside the city centres – are expected to be truly autonomous by about 2030. Light rail or trams which operate in a complex inner-city environment are forecast to achieve autonomy by the middle of the next decade.”

One of the technical leads at the Franco-German consortium is Alain Le Marchand, an automation engineer who is Thales Ground Transportation Systems’ senior design authority. “We started our five-year project back in 2018 and have been running test trains during the school holidays when the timetable is quieter. We have achieved GoA2 – automatic train operation with a driver – and from spring 2022 will add the functionality to progressively reach GoA4,” he says.

The consortium has solved the problem of pinpointing the train’s exact location by combining data input from three different systems – navigation satellites, sensors on board and the train’s own map generated offline. The new positioning system replaces an expensive legacy system that requires an odometer mounted on the train axle, and a reader that determines accurate position as it passes over positional markers or balises set between the rails.

The system reads the all-important traffic control signals by combining input from one camera with three processing threads which separately identify colour, shape, and position of the lamp, combining to produce a level of certainty and safety. The team will replace the attendant on the train with cameras set up to monitor the passengers’ behaviour, for example to react in case of illness, resolving all the technical challenges of autonomous trains by 2023 ahead of the industrialisation phase.

The goal is to reach safety integrity level 4, at least equivalent to the safety level of the driver that it replaces. “We have obviously had to define that, to design the system. It was also important not to overdesign, which results in greatly increased levels of expense,” adds Le Marchand.

Present throughout the project’s life have been the French National Safety Authority for Rail and the national cyber-security agency ANSSI, to ensure that any design is certifiable. It would clearly be a lot easier to rip out all the existing infrastructure and start afresh, but the sheer quantity of legacy infrastructure built and installed over decades, and with decadal lifespans, mean that is not an option.

Le Marchand laments that small countries like Denmark and Norway will remove all their wayside signals and equip all their trains with an onboard controller, but that simply is not possible in France, Germany or the UK, where the rail networks are much larger, and with greater diversity of rolling stock. He sees a future where most of the infrastructure will have been removed from the tracks and migrated on to the smart trains, and eventually all that remains trackside will be critical points-changing gear and level crossings – incidentally, the number one source of accidents on the railways.

Another benefit of having autonomous trains running along a track is the data gathered during normal operations. Constant monitoring of the trackside environment is compared with the system’s own map of the trackside ecosystem. Changes detected by the train’s sensors may give an early warning of landslips, defects in bridges or an impending tree fall, all before the human eye could spot such changes.

The fact that a train runs on tracks does not render the achievement of full automation a simple objective. Aside from the need to monitor the status of the track, and the position of other trains using sensors, positioning systems and cameras, the physical integrity of the train itself must be monitored – for example the links between engines and carriages, and between carriages.

The distance needed to bring a train to a full stop will also vary enormously. The stopping distance of a train is a function of its body weight, load (passengers or freight), and speed. A train running at 100km/h with ten carriages might need 500m to stop. A high-speed French TGV travelling at 320km/h can stop in an emergency in 3.39km, but in practice the train will decelerate slowly over 5 minutes and 10km, to protect its passengers from an abrupt stop. Because of this, autonomous trains or light rail systems like trams need to use sensors that look much further ahead than the average semi-autonomous car (for context, a car travelling at 90km/h would have a braking distance of approximately 80m). Any autonomous rail system must consider signalling and obstacles that far out.

Signalling is also extremely diverse, certainly compared to traffic lights on the road, and there may be many different types of passenger and freight trains running on the same infrastructure. Combinations of all these must be learned by any system. Multiple types of level crossing, track-changing systems, and other infrastructure elements dating back decades add complexity. In addition, there is a critical necessity for extremely high standards of safety given the numbers of humans that could be impacted given a serious accident involving a single train.

Less headline-grabbing than the path to train autonomous operations, but just as important, is the digitisation of infrastructure. The UK’s rail infrastructure has had 150 years of continuous operation and is the oldest rail network in the world. It comprises millions of interconnected assets – trains, carriages, 20,000 miles of track, signalling systems, bridges, sidings, depots – all must be monitored, maintained, and updated to ensure ongoing safe operations. Some of these are antiquated legacy systems dating back so far that they are becoming increasingly difficult or impossible to maintain. It’s a huge task, and an expensive one. Across Europe, rail operators spend €15- 20bn a year on maintenance and necessary renovation. There are huge dividends to be had in cost savings, safety, and reliability, to come from integrating the technological monitoring of these assets.

Data inputs from track and train-borne sensors, engineers on the ground, footage from drones, inspection trains, and helicopters can be combined and analysed to predict impending faults, help operators get smarter about monitoring and managing those assets in real time, and preventing points of failure. The traditionally reactive timetable of system shutdowns inconveniences passengers, and failures that require urgent responses in the form of unscheduled maintenance cause timetable chaos.

Simon Hosking of Thales UK is head of the group’s Digital Competence Centre based in Manchester. He says: “In our smart maintenance software, TIRIS, our mantra is to discover, predict and advise – analysing the flow of data from sensors on assets using proprietary deep-learning algorithms that track multiple variables from, say, a signalling system, and turn that data flow into an action plan for maintenance engineers. Faults can be fixed before they become failures that impact passengers, and because engineers know what they are going to encounter, only the appropriate tools and parts need to be carried to site.”

Network Rail’s Anglia route was notorious for delays of hundreds of minutes a month due to failure of critical trackside equipment. Sensors on 300 axle-counter sections (which allow signalling systems to know when a train has entered and left a given area of track), now allow engineers to monitor their condition digitally and remotely in real time, and this has been rolled out over a larger part of the network.

Digitisation can also bring with it new types of data. With Thales’s new fibre-optic axle counter, trains are not only detected when they pass by, but also analysed for weight so that the number of passengers in each carriage can be calculated in real time, and can be fed back to boarding passengers via an app.

An ongoing project in Hungary loads data to the cloud from data loggers at the points, which switch trains from one track to another. These record power voltages and signals running between the points motor and the points control system – a safety-critical system.

“Our algorithms may have seen a particular pattern of power consumption before that indicates a stone stuck, or a power problem,” says Hosking. “It is critical to include a customer feedback loop to continually improve the quality of the deep learning and its ability to indicate maintenance requirements or predict impending failure issues. Users help us to link the data flow back to reality, and because the whole system is cloud-based it can be accessed from any browser.”

While digitisation of the railway infrastructure is proceeding apace, the decadal, or even longer, lifecycles of the rolling stock will prevent a speedy progress to 100 per cent automation on the tracks. “It’s not going to happen overnight, but we are preparing for the future. The autonomous train will arrive before the autonomous car – the tracks make the problems to be solved simpler. I expect to see autonomous trains in operation before the end of this decade,” says Thales’s Le Marchand confidently.