Artificial intelligence – bringing us the future
Each day we better understand the immense potential of artificial intelligence for transport. Like many sectors, the rail industry is investing in this powerful driver of innovation. Artificial intelligence will thus be applied much more widely across the SNCF Group in the years to come.
Occupations, processes, tools of production, and commercial relations will no longer be the same. Maguelonne Chandesris, head of the Data, Mobility and Regions cluster in Innovation & Research sheds light on the changes artificial intelligence is bringing.
What exactly is covered by the concept of artificial intelligence?
One of its creators, Marvin Lee Minsky, defines it as the construction of computer programmes to carry out tasks that for the moment are performed more satisfactorily by human beings because they require high-level mental processes such as perceptual learning, organisation of memories, and critical reasoning. Artificial intelligence, which was first conceived in the 1950s, is enjoying a real comeback thanks to the huge growth in the production and storage of data as well as the processing capacities of computers. The digitalisation of virtually every aspect of our lives, from photos to music, along with the development of the Internet, contributes to the production of massive quantities of data, which are the raw material of artificial intelligence.
To work with these data, we use algorithms. These do not of course enable our present-day machines to imagine and invent. But they are capable of deep learning. That is how Google’s AlphaGo beat a champion Go player in 2015. Some researchers are now working on networked games, which comprise strategy, alliances among players, and complex tasks… a bit like a rail network!
How can full advantage be taken of these new possibilities?
SNCF is already using some forms of artificial intelligence every day, but we can find more uses and rethink how we operate in various areas, for example, in the relations with our passenger customers. The robot Pepper, which was tested in stations in late 2015, holds a tablet, winks its eye, and speaks. Its ability to perceive its environment and interact with humans relies on very complex programs. It must understand what passengers are saying using automatic language processing. In addition to verbal interaction, there are other types such as gestural interaction, involving the analysis of the movements and facial expressions of humans, and the proposal of complex services. Artificial intelligence can also help our agents with diagnoses during maintenance and repair operations. Work is also being done on artificial intelligence in SNCF’s technological renewal programme TECH4RAIL
What are some specific AI subjects that Innovation & Research is focusing on?
Besides the two examples I have just given you, we have begun a programme called Base Artificial Intelligence. It includes exploratory work related to operations management. Some of this work is being done with partners, major actors like the Institut de Valorisation des Données (IVADO) in Montreal and the École des Mines in Saint-Étienne. We also have a post-doctoral researcher from the École nationale supérieure Paris- Saclay working on an innovative subject, the adaptation of epidemiological mathematical models to train delays. These models are developed not only to understand how infectious diseases spread, but also to determine who should be vaccinated as a preventive measure and who should be given antibiotics. They could help us better understand delays on train lines and how to limit their impact. Adopting this type of multidisciplinary approach turns out to be very rewarding for innovation. I also believe it is essential to address the ethical issues raised by artificial intelligence. Artificial intelligence allows us to automate certain intellectual processes, not just physical ones, and that raises questions about the role of humans. Can algorithms be used to hand down judicial rulings, for example? With autonomous trains, what can be delegated to the machine with compromising safety? Scientists create algorithms, but they are not the only ones who should answer these questions.
The view of Dominique Cardon, sociologist, professor at the Médialab of Sciences Po, and author of À quoi rêvent les algorithms? (What Do Algorithms Dream About?)
“The expression ‘artificial intelligence’ suggests that machines could acquire the autonomy that we grant to humans. It makes no sense to separate humans from artefacts. Since the dawn of humanity, we have been coupled with technical systems that we have create to enhance our capacity for action and for the comprehension and intellection of the world. Rather than “artificial intelligence”, I prefer to speak of “augmented intelligence”. The third springtime that it is enjoying today represents a new and original coupling of man and machine. Today’s machines make statistical calculations using enormous quantities of data to build models that are no longer deterministic, but probabilistic. And these models are capable of learning. However, whether they are used for image recognition or translation, their intelligence remains hyperspecialised, in contrast to human intelligence. The real debate is not about the autonomy of machines. It is to know whether we are properly educating the algorithms, which propose actions or make decisions based on the indications we give them. Are we giving them the right data and the right models? Their current usefulness is fairly individualistic: providing the best service to each user, with the risk of being enclosed in the bubble of their own rules of behaviour. They could augment our capacities, help us to think better, to be more curious, and to cooperate better.”
A closer look at two projects
When a train slows down a little, it can have a big impact on a line, particularly in a densely populated area. By better understanding how this impact can spread and findings ways to reduce it, traffic flows on the network can be optimized and journey times shortened.
What Innovation & Research is doing
Approaches are being developed, tested, and compared in several projects to improve:
Our understanding of train delay propagation
One approach being explored involves machine learning. Traffic data are analysed in order to predict statistically the propagation of delays. A second approach consists in modelling the way various rail system components function and simulating their functions so as to predict the delay that will occur and which train it will affect. The third approach uses epidemiological models.
Regulation of delays
Two algorithms have already been developed. The first recommends the best sequencing of trains coming from different lines but that must converge on the same track. The second recommends to the driver an optimal standing time for his train in a station according to downstream traffic. Innovation & Research is also assessing epidemiological simulation models as a means of evaluating decision-support strategies for traffic regulation.
The two algorithms were tested in early 2017. They do indeed allow delays to be managed in real time. A test of their combined use will be done the autumn of 2017.
A chatbot is a robot (bot) capable of conversing (chatting) with a human being whose natural language it understands and to whom it can reply in an appropriate way. The best-performing chatbots are self-learning systems. Their emergence has been driven by the Web giants, who have developed “personal assistants” (Apple’s Siri, Google Home, Amazon Echo, etc.).
What Innovation & Research is doing
Our teams worked on the first dialogue terminal back in 1994. Their subsequent projects, all on automated dialogue systems, contributed to the development of the SNCF information number 3635, which included voice recognition, and then to the bot Léa of Voyages-Sncf.com. In 2014, Innovation & Research developed a new system, this one to guide travellers in stations. Trials were conducted two years later.
Innovation & Research is going to do state-of-the-art and state-of-the-market surveys before comparing existing solutions. This will be help entities in the Group like Gares & Connexions, Transilien, and Voyages- Sncf.com that are developing chatbots independently. To encourage exchanges and pooling, Innovation & Research (along with the Digital department of e-sncf) plans to contribute to the creation of a community for these entities and provide them with its expertise in certain technical aspects such as language processing.
The foundations of a disruptive technology
After the high-speed train revolution, SNCF intends to maintain its lead by successfully applying artificial intelligence in several areas. Here are three.
1. To spot defective pantographs
The aim of SNCF Réseau’s CAFEINE (CAaméraFErroviaire à Intelligence NEuronale ») project is to automatically detect pantograph degradation on trains in motion that could lead to a broken catenary.
The idea is to develop a unit consisting of cameras hooked up to a local network of artificial neurons. “Like a computer terminal endowed with local intelligence, the unit analyses the images and transmits only the relevant information. It is a self-learning system”, explains Alain Rivero, the IT specialist at SNCF Réseau who initiated the project. Three prototypes have been produced with the company GlobalSensing Technologies since 2015. The latest has only two cameras the size of tennis balls, unique characteristics in the market, as is the way in which they work. Eight units are being tested now. Installed on gantries over the tracks, they can detect pantograph defects on trains travelling at 160 km/h. Eurotunnel and Deutsche Bahn are already interested in testing them. “We are targeting industrialisation at strategic locations in the Paris region in 2018”, says Rivero.
2. Improve rail system performance and responsiveness
The project Boost Operations with Artificial Intelligence is part of the TECH4RAILprogramme. “The aim is to make our operations more agile. To do that, we have to manage in real time not only train traffic, but also the flows of passengers and the rotation of rolling stock and human resources”, points out the project’s leader, Fabienne Réveillac.
By using artificial intelligence, it should be possible not only to determine conditions in the system in real time, but also to use projection algorithms to predict future conditions and take appropriate action. “SNCF has already developed a few decision-support algorithms. The aim of our project is to speed up their integration in operations, notably through the development of a test platform”, says Réveillac. With this modular platform, it will be possible for the operators to test the uses of the various technological building blocks. Our objective is to have a first prototype ready by the end of the year.” Among the technologies that could be brought into play is deep learning as well as visualisation, modelling, and voice recognition.
3. Autonomous train: from dream to reality in five years?
SNCF firmly believes that automation and artificial intelligence will be the keystones of the future rail system. Its objective? Put a first autonomoustrain on the tracks early in the next decade.
SNCF has made the development of an autonomous train by 2023 one of the priorities of its technological renewal programme TECH4RAIL launched in 2016. With the transport industry evolving rapidly and becoming more competitive, the goal is not just to operate trains without a driver, but also to provide the rail system with more capacity, flexibility, and safety. Automation offers solutions to the problems of congestion on TGV routes and lines in densely populated areas like the Paris region, where the number of trains in simultaneous operation is almost at its maximum already. This technology will optimise train speeds, thus providing smoother traffic flows and increased capacity on the routes. Trains on the same line are all programmed to run in the same way, which will increase frequencies and improve punctuality. Calculating speeds will also reduce energy consumption.
Different levels of automation
SNCF plans to develop smart trains with varying degrees of automation, ranging from assisted driving to total automation. In this way, the solutions can be adapted to different types of trains (TGV, freight, or TER commuter trains), to the future European signalling standards, and to existing signalling. In the Paris region, the NExTEO project will provide a response to the need for more capacity with the westward extension of the RER E line. The first step consists in developing by 2019 a train-drone prototype for freight, i.e. a train remotely controlled by a driver. The first two prototypes, one for freight and one for regional passenger service, will be equipped with an autonomous, entirely automated system in late 2022. The future TGV equipped with an automated driver assistance system is expected to enter into service in 2022 or 2023.
Innovation through cooperation
To tackle this technological challenge, SNCF is collaborating with multiple partners. The Technological Research Institute (IRT) Raileniumhas been asked to do studies on the “drone” train and artificial intelligence. Last spring, a partnership was also formed with Alstom and the IRT SystemXto create an automatic environment perception system for a train. It will be based on a combination of sensors, including cameras and lidars, to detect obstacles along the tracks and signalling. A test train has been carrying out trials since February. SNCF has also launched a partnership initiative to build prototypes of autonomous passenger and freight trains. And at the European level, SNCF is working closely with Deutsche Bahn. While both companies have their own automation projects, they regularly share the results of their studies and tests. The objective is to advance in a coordinated manner to define a European approach to autonomous trains.
In the Autonomous Train project, which is part of the TECH4RAILprogramme, a range of possibilities, from partially automated driving to full autonomy, are being explored. Hugues Cheritel, a member of the project team working on the Artificial Intelligence package, sums up the hoped-for benefits: “Increased network capacity on commuter and regional lines as well as high-speed lines like Paris-Lyon, improved punctuality, and reduced energy consumption.”.
Two key objectives are targeted. First, autonomous train systems adaptable to the entire SNCF fleet. Preceding this, prototypes suited to various types of operations will be developed. Second, development of the necessary technological building blocks, like the environment perception systems developed in the partnership with the IRT SystemX. Another partnership, this one with Railenium, is dealing specifically with artificial intelligence. “We want to specify possible needs in this area for the autonomous train, identify the useful technologies, and determine the state of the art in the sector”, says Cheritel. This exploratory work will contribute to the development of various levels of automation for regional (TER) and freight trains.
Luc Laroche, director of the Autonomous Train project
“Trains in the future will be packed with automated systems, just like other vehicles. This has been the case for some metros for a long time, but only in a closed environment and over short distances. With autonomous trains, we will be going long distances in an open environment. We face new challenges: detection of obstacles and observation of the environment, geolocation, telecommunication, and on-board surveillance, among others. These are major human and technological issues. We are mobilising the collective intelligence of SNCF, its experience, its competencies, and its know-how. We are enlisting the entire rail industry in this effort, and inviting partners from other industrial sectors, notably the automotive industry, to join us.”
Point of view
Two questions for Nicolas Vayatis, director of the Centre de mathématiques et de leurs applications at the École normale supérieure Paris-Saclay.
What uses can be imagined for artificial intelligence in industry?
There is of course predictive maintenance, where it can help target and anticipate operations based on data collected by networks of sensors. However, current tools remain fairly descriptive. They are much more mature in the Internet realm, where there are already a lot of applications like automatic translation, targeted advertising, recommendation engines, and search engines. It is possible to imagine, for example, specialised recommendation algorithms used in industry as decision-support tools. Little organizational and cultural revolutions might take place at manufacturing firms. Exhaustively digitalising what happens in a company and uniting the databases of its activities in a common digital environment would result in a more transparent, less segmented, and more immediate vision that could totally modify how information circulates and therefore how decisions are made. Artificial intelligence is going to radically transform the human resources activity as well as legal services. It is already possible to draw up contracts automatically.
Where are the obstacles to these transformations located?
The real challenge is not technological; it is to achieve a dialogue between digital technologies and business expertise. In some sectors, there is a natural inclination to establish this dialogue, and things are moving forward quickly. Other obstacles are development costs and the human impact. Few people are able today to determine the return on investment of an activity’s automation, for example. Under these circumstances, it is hard to plunge in! Some companies go ahead anyway and occupy terrain that might otherwise attract other players. This risk of intermediation is enough to trigger huge investments, for example, in the banking sector.