Artificial intelligence won’t be replacing train drivers any time soon – but it may be making their working lives, and the lives of others, far safer.
Artificial intelligence is set to play a crucial role in the world of rail. But that doesn’t mean human beings will be squeezed out. Rather, new technology will be helping ensure work environments are both safer and more efficient.
Derel Wust is the managing director of 4Tel Pty Ltd, a software and hardware engineering business based in Newcastle, NSW.
4Tel has developed an AI solution environment called HORUS, after the falcon-headed all-seeing god of Ancient Egypt. 4Tel is now commercialising this work in a specialised company called “AI System Pty Ltd” which will separate 4Tel’s AI work from its general software development work. Wust talks of the benefits of AI technology which can learn from examples of success via machine learning processes based on neural network processing. Machine Learning is a very powerful analytical technique for processing uncertainty in stochastic processes where there is uncertainty or randomness in the outcomes. The 4Tel work on AI, and now within AI Systems, has a focus on artificial intelligence and particularly in the context of rail. The business deploys a team of military command, control and communications engineers building military-grade networks for rail. 4Tel is also a platinum sponsor of University of Newcastle Robotics Laboratory for development of AI and machine learning techniques for rail applications.
“There’s this idea with artificial intelligence it will be about getting rid of the driver, but for at least the foreseeable future that’s not the main game. Safety is the main issue,” Wust said.
“For some rail networks automating will be important, but on many rail networks, particularly passenger networks, the approach is going to be much more conservative.”
Rather, AI can play a crucial role in the context of safety and improved performance. AI can help on a variety of levels, for example with track modelling, signalling and even driver fatigue assessment.
“The real strength of AI is it is another technology to help us in normal day-to-day operations to improve safety,” Wust said.
“No-one goes to work to have an accident – most people go to work, they lose situational awareness, they make a mistake and they have an accident. AI is very effective at helping drivers, controllers and track workers get more accurate information to make better decisions and better decisions will lead to safer outcomes.”
Wust talks of “doing boring better” or improving the basics.
“If we can do boring a lot better, we can save people from getting distracted, losing situational awareness and making a mistake,” he said.
For example, improving location assurance enables the checking of the separation between all track users.
“Improving situational awareness helps by being better informed and minimising out-of-authority events like Signals-Past-At-Danger (SPAD) and speeding because the computers can alert the driver when they are approaching a signal or they have inadvertently started to speed,” Wust said.
“A lot of poor decisions can be prevented just by giving people better information in time.”
AI sensors also can be deployed to detect hazards, both day and night and in underground tunnels.
“They don’t get tired, they don’t get drowsy, computers and sensors just keep doing what they are doing,” Wust said.
AI may also contribute to preventing self-harm tragedies.
“With AI we have the capability to detect people on or near the track who may be in an unauthorised area or acting in an agitated way.
“This may allow the driver to brake or at least warn other rail staff to intervene for help in time,” Wust said.
“While this may not be the answer to the tragedy of self-harm, anything that can reduce its frequency or save one life is a good thing.”
AI can help in the safety process by way of fatigue management in train drivers, with long-haul routes a particular challenge. It has a role in managing driver fatigue also.
“AI processes can help monitor for that – to help the driver not to catch them out and get them into trouble,” Wust said.
“Via AI, we can do many other boring but vital tasks.”
A vital part of 4Tel’s operations are two concepts, Stochastic processes and Deep Learning, which Wust describes.
The real world and safety assessments do not work with zero risk of things going wrong, or safety measures failing. Many real-world examples exist whereby extremely unlikely accidents have occurred because they weren’t stopped by an appropriate safety process because a decision somewhere in the project’s history concluded the risk of such an event was extremely unlikely – but then it happened.
“Safety assessments inherently try to manage risks in the terms of risk identification, likelihood of occurrence, consequences and risk class matrixes which is an inherently stochastic process,” Wust said.
“Into this context of assessing risk statistically, and all projects always have residual risks of some type, we recognise that both projects and risk assessments are stochastic as there is uncertainty or randomness involved in the outcomes.
“Unfortunately, our current conventional hardware and software solutions are programmed to make decisions on a ‘IF…THEN…’ basis and do not handle randomness well. An answer can be calculated to be perfectly right as far as the computer process is concerned, but an embedded error not found by testing processes may have caused a ‘wrong’ answer from a risk management perspective.”
Wust said they could forever add more and more “IF…THEN…” statements to improve conventional coding and decision making, but unknown project and calculation error risks would remain.
“Or we can improve our decision making by actively using measures of uncertainty of key safety calculations and assessments,” he said.
“Into this situation comes the new technology of AI that will offer us an alternative way forward because AI uses statistics and weightings in its processes.”
A particular AI processes whereby an image or data is assessed for identifying a pattern of interest, based on machine learning, is inherently based on stochastic processes, but this doesn’t make AI processes unsuitable for safety work if a solution has measures to manage residual uncertainty.
“AI actually allows engineers to design safety to a required standard by adding more and more orthogonal stochastic processes into situational assessments,” Wust said.
“Safety will significantly improve as AI and machine learning techniques become better understood, because AI processes will manage safety processes better than conventional techniques in many situations.” 4Tel is now well developed in investigating advanced statistical techniques on its path to better protecting people and assets by applying holistic safety assessment processes based on the better assessment all available data.
Deep learning is a more recent specialised development of machine learning and uses computerised neural networks to solve cognitive problems. DL using neural networks especially suitable for image, language and speech data where all data is inherently stochastic. Whereas most conventional software programs run on common computer processing units (CPUs), deep learning techniques work fastest on Graphical Processing Units (GPUs) where neural networks can be emulated. Modern GPUs often have specific features for deep learning and are themselves evolving into a new type of neural processing unit (NPU) where neural techniques are specialised.
The general process of neural processing is the key to improving image and video processing in real time for rail applications.
“It is these advances in deep learning that empowered the automotive industry into AI techniques economically, and the same is now happening in rail,” Wust said.
“AI techniques based on deep learning technology will provide the technical basis for improving safety, because more and more data, even with statistical characteristics that was otherwise unusable for processing by conventional computers, can be assessed in real time to assess developing risks.”
4Tel identified the importance of AI to the rail industry back in 2016.
Since then, they have sought to learn deep learning techniques and consequently develop their own specialised rail deep learning mathematics and processes with the assistance of the Robotics Laboratory of the University of Newcastle under the guidance Associate Professor Stephan Chalup. 4Tel subsequently recruited its own team of PhD computer scientists to work cooperatively with the university team.
“4Tel, with AI Systems, now has highly specialised staff who have developed mature techniques ready for proof-of-concept deployments across transport sectors, as evidenced by the recent trial of our work announced by Rio Tinto at Heavy Haul 2021 in Perth this year,” Wust said. “4Tel’s mathematics and AI algorithms are developed in Australia, supported in Australia, by Australians, for Australian conditions and are demonstrably globally competitive for Rio Tinto to have selected us to participate.”
“4Tel has been attracting considerable overseas interest in its work for some time now,” Wust said.
“It is now apparent 4Tel is not only an Australian leader in rail AI thinking, but also is globally competitive in applying AI to improving the efficiency and safety of rail applications”.
Unfortunately, the current lack of overseas travel is severely reducing opportunities to interact personally with the international industry to explain a very interesting but complex topic.
“Zoom/Teams/webinars allow us to keep opportunities moving at some level, but obviously Australian opportunities are most important at the moment because we need to apply our AI work in real applications.”
Wust said AI would take effect incrementally.
“The rail industry has a reputation for being conservative, albeit it is conservative for a reason – because rail accidents can have a devastating impact,” he said.
“What we are doing is bringing about progress and change by applying new techniques to age-old problems, but in a way that people can understand and directly contributes to safer and more efficient outcomes for the rail workforce.”