Trains already have a driver-assistance system and artificial intelligence tools on board, in particular for speed control, and lines are also equipped with many sensors. What are the different systems on board the autonomous train, and what paradigm shift do they hail?
Today, SNCF trains are equipped with what is called an Automatic Train Protection (ATP) system, which makes it possible to ensure that a driver remains within a speed range that is compatible with security. If this speed is exceeded, the ATP makes the train make an emergency stop. But it is up to the driver to adapt the speed to the train ahead: the ATP simply calculates the train’s deceleration curve when it has to stop at a red light. It’s an electronic system equipped with a speed curve calculator, not an artificial intelligence – at least, not in the technological sense of deep learning. Depending on the generation of train, this equipment is more or less sophisticated. For its part, the signalling system, coming in the shape of signal lights and enabling the train to position itself according to other trains and to the network’s danger spots such as switches, is ever more efficient and digitalised.
Tomorrow, with the autonomous train, the train will drive itself, without any in situ human assistance: the decision to accelerate and brake will be made by the train – which doesn’t stop us keeping the ATP, our protection and control system. The aim is to progressively operate trains equipped with different levels of automation (see video below). Sensors and driving intelligences will essentially be placed on board the trains and not on the ground (tracks or railway surroundings), so as to facilitate the transition: the change from one model to another will take place over several decades, a period of time over which we do not wish to profoundly change the infrastructure. Why? Firstly, because the marginal cost of our first autonomous train would, in this case, be much too high. Furthermore, during this long transition period, on this one infrastructure we wish to run both manually driven trains and autonomous trains of automation levels 2, 3, or 4, one after the other and at random. Our autonomous trains must therefore be able to run whilst adapting to infrastructures (railway track, signalling equipment, etc.) built for manually-driven trains.
Geolocation and obstacle detection play an important part in the autonomation of trains. What are the technical challenges still to be met in these areas?
Over the last few years, the use of autonomous trains in a closed circuit has become widespread, like on line 14 of the Paris Metro. However, the running of trains in an open environment remains a challenge. In a closed setting, we consider the environment to be safe, with no obstacle on the infrastructure. On the other hand, in an open setting, the outside environment requires a high level of monitoring, which makes tasks a lot more complex.
If we want autonomous trains to run without having to change the existing infrastructures, the trains must know how to position themselves precisely on the tracks or the platform. Today, train location is communicated in two ways: through the empirical knowledge the drivers have of their line and through blocks, divisions of tracks ranging from 500 m to 40 km. In this geolocation system using track circuit, the train shunts the circuit when travelling, and from there, the signalling system is capable of detecting the block in which it is positioned, with a precision of one kilometre.
Tomorrow, to anticipate an upcoming red light for example, the autonomous train will need to read signals and place itself precisely within its block, which requires geolocation to within the metre. We thus change technology completely by merging the sensors of an inertial measurement unit with those of a GPS. Accelerometers will be able to record variations in position, and will be recalibrated regularly with a GPS position for checking.
So, the first challenge would be signal detection…
Exactly. The very latest systems installed on signals already send digitalised information on board the train, which simplifies automatic driving. But when we are on an older system with signal lights, it is essential to take into account what they are indicating. This is an ambitious challenge that we are taking up, notably thanks to a camera that can extract, analyse, and interpret the traffic light image within the scenery.
By means of an artificial intelligence?
For lineside signals, as for obstacle detection, we require such a high level of security that we cannot call on deep learning technology: to rely on a statistical model would require too many trials. However, artificial intelligence is used for monitoring the train’s environment. The overhead line for example, which provides electricity to the train, can be monitored by an artificial intelligence, because in the case of an error on its part, there would be limited damage: in effect, the train would simply tear away the overhead line and stop without causing an accident.
And what about obstacle detection?
We have two major issues: medium distance and long distance. Medium distance, around 200 m, corresponds to the obstacle detection distance for vehicles: therefore, we took inspiration from the technology used in the automotive sector. However, a train brakes much more slowly than a car so we need a second detection at a greater distance (800 m or 1 km), which is also a technological challenge to be taken up. Finally, we are researching a better definition of the surface of obstacles so the train doesn’t stop, for example, for a dead leaf on the line. In fine, the removal of uncertainty will be provided by the train supervisor (who will be in a control centre on the ground).
We’ve heard of the remotely-driven train as a step towards the autonomous train. What about that?
As for the remotely-driven train, that we also call a drone-train, we should have a first demonstrator by the summer. The benefits of this experimentation are twofold: at first it will be a degraded mode of the autonomous train, which means that a broken-down autonomous train will be taken control of remotely to be driven to the next station. Remote driving will also be used for technical routes, between the warehouse and the station for example.
How does one go from a train that can drive itself to smart management of the entire network?
There are two different systems: autonomous driving, which we have just talked about, and autonomous supervision – bearing in mind that it is quite possible to have one without the other. Autonomous supervision is equipment that makes it possible to make choices in train sequencing. For example, if two trains that should follow one another arrive at a switch at the same time because one of them is late, supervision asks the question of which train to send through first. Today, this is very well managed by people based on their experience, but we know that we can optimise this task because the depth of computer calculation gives an idea of the more long-term consequences of decisions taken.
How will the autonomous train fit into the other SNCF innovation programmes, and to develop what vision of mobility for tomorrow?
The autonomous train is one of the building blocks of the third rail revolution, which prefigures the train of the future: the sum of an autonomous driving system, an autonomous supervision system, a digitalised train signalling system, a resource management optimisation system, etc. and finally it is the entire traveller interface that is based on digital tools to make the customer experience smoother and more efficient. This is going to revolutionise the way in which trains are operated, by making it possible to put many more trains on the same infrastructures, thus reducing network congestion in core areas and offering much more flexibility in less congested areas. Digitalisation also enables us to work on all the approaching journeys, from home to the station for example, so as to make using the train simpler for our customers. The aim is to once again extend the relevance of rail in the coming decades.