Artificial Intelligence and people in rail: maximising the benefits and managing the risks

Artificial intelligence (AI) capabilities have rapidly accelerated in recent years. Historically, the rail industry has been slow to adopt new ways of working, but several railway companies have started to explore and implement AI for some activities. Today, all railway companies have the potential, and the opportunity, to harness the power of rapidly evolving AI technologies to improve the planning and delivery of services.

A recent seminar organised by the IMechE and IRSE explored the impacts that AI is having in the rail industry through the perspectives of human performance, design acceptance, and change management. The seminar focused on the potential benefits of using AI in decision making, failure prevention, and infrastructure monitoring, as well as the risks that need to be considered in rail’s highly regulated industry.

There are some who question if AI should be used in the rail industry and how the industry will adopt new intelligent technologies, such as AI and Machine Learning (ML). ML is a part of AI which uses algorithms to enable systems to learn from and make decisions based on data. Unlike traditional programming, where explicit instructions are given, ML allows the system to learn and make predictions from the data, without explicitly being programmed for each task.

The point was made early in the day that AI and ML are already involved in many applications, and are being deployed in the industry. The next generation of engineers is also already working with the technology at university or college. Like it or not, it will be coming to the rail industry. Furthermore, it offers huge benefits but with risks that can be managed. This was demonstrated and discussed throughout the day.

AI fundamentals

Professor Rod Muttram, junior vice president and a Council member of the IRSE, presented the first session of the day by explaining the fundamentals of AI deployed in a safety function. Rod confirmed that AI and ML offer the rail sector significant benefits but, like all new technologies, there is a need to take a well-structured and disciplined approach to gain acceptance, and avoid reliability and safety problems.

Rod explained the difference between automation and AI. Automation has been used for decades, using computers and conventional engineering techniques with behaviours programmed and validated by human engineers. There is a need to define carefully what an AI system is, as some systems today referred to as being AI are actually ‘automation based’ with limited intelligence.

Alan Turing (1912 – 1954) introduced what has become known as the ‘Turing Test’ in his paper ‘Computing Machinery and Intelligence’.  The Turing Test involves a human evaluator interacting with both a human and a machine without visual cues (i.e. using textual messages). If the evaluator cannot reliably distinguish between the two based on their responses, the machine is deemed to have passed the test.

The Turing Test still serves as a benchmark for assessing a machine’s intelligence. It gauges the machine’s capacity to emulate human-like cognitive abilities. True AI devices are machines that behave like a human brain, using multi-layer or deep neural networks, which are then taught or allowed to learn what to do. The environment and the learning programme for these systems are as important as the engineering.

AI and ML techniques are already being used in the road transport sector, for example, and there have already been some high-profile incidents. The comparative safety ‘benchmark’ for roads is much less challenging than rail and the accident rate for human drivers on roads makes it relatively easy to make improvements.

However, the safety benchmark for rail is typically much higher compared to road, and even a single rail incident caused by AI would result in a public outcry. So, using AI / ML in any safety or operationally critical area will have to be very carefully considered, and is probably years away.

Another issue with very complex intelligent systems is the need for continual OEM support and ‘supplier lock-in’, which many railways want to escape from. Avoiding this may require the confinement of AI / ML to certain modules, rather than allowing extensive control. Rail suppliers are also unlikely to take on the long-term support liabilities required for intelligent systems without significant rewards. But could AI systems be made intelligent enough to design their own replacement?  

Rod warned that a true advanced AI may develop human traits such as boredom and making errors. He also recommended that rail does not use AI when conventional technology will do. Where AI is used the following must be addressed:

• Transparency – especially for any safety function.
• Algorithmic bias – systematic and repeatable errors that create ‘unfair’ or unsafe outcomes different from the intended function of the algorithm.
• Traceability – important and essential for safety and reliability.
• Ethics – AI needs to learn ethically, which was a key part of the 2024 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS).

Auto-reverse

Rod does not recommend using AI for any safety function above Safety Integrity Level (SIL) 1. For example an ‘auto reverse’ function is used at Westbourne Park on Crossrail, for turning the 14 trains per hour in the reversing sidings. The driver selects ‘auto reverse’ and walks back through the empty train.

By the time the train gets back to Paddington (about a mile away) the driver should be in the other cab ready to form the next eastbound departure. However, the signalling control is achieved using a conventional high safety integrity SIL 4 system, with AI providing surveillance of the track at the turnback to ensure no staff or trespassers are harmed by the moving train. This does not require more than SIL 1.

Other AI uses envisaged by Rod included identifying and highlighting incidents on platforms (for example) in real time from AI surveillance of many parallel video channels. High resolution CCTV is relatively inexpensive and extensive throughout rail, but human analysis of the video images is expensive, boring, and prone to errors. Using AI to search accurately and quickly through the video information, and ‘flag up’ incidents for human decision making and action could deliver huge benefits.

Trust me

Dr Nick Reed, chief road safety advisor at National Highways and founder of Reed Mobility, gave a presentation titled ‘Trust in me? How emerging technologies might keep us safe and why they may not…’. Nick explained that many road incidents are caused by tired drivers who lose concentration. Surely an AI controlled road vehicle would not get tired and be safer?

Nick then discussed an incident where an AI controlled road vehicle hit a pedestrian who had already been knocked over by a ‘normal’ car. After hitting the pedestrian, the AI vehicle tried to continue its journey by reversing and going around the pedestrian, but this resulted in the pedestrian’s clothing getting caught by the vehicle and they were dragged up the road.

Incidents like this with AI vehicles could be corrected, but how could the behaviour of the AI vehicle be verified? One solution suggested was by using a data version of a driver’s spoken thought commentary, which could be shared with the regulator.

AI road vehicles were discussed a number of times throughout the day with the observation that society accepts that human drivers will make some errors but is less likely to accept AI making any errors.

David Golightly, lecturer in human systems integration at Newcastle University, covered ‘A human-centric view on AI in railways’. He explained that AI needs to be looked on as a tool and examined how AI could help human perception, interpretation, decision making, and action taking?

If designed and implemented correctly, AI could help with the perception and interpretation of huge amounts of data, leaving the decision making and actioning to humans. The scale of data generated in modern systems far exceeds the human capacity to understand and analyse it, but AI algorithms can help process higher volumes of complex data, turning it into information which can be used more easily by decision makers.

To achieve this, AI must achieve the trust of humans and reduce the human workload. It must help not hinder. However, gaining confidence is not easy as it’s not possible to see how neural networks work.

Risk perception

Heather Taylor, of Fraser-Nash Consultancy, presented on ‘Risk Perception and Management: how does AI change perception of risk?’ She explained that the key is finding the balance between human oversight and AI autonomy, to ensure that the AI technology supports rather than replaces human decision making.  

Regular checks and audits of any AI decisions will be required to ensure that risk management, safety and critical judgements are maintained as if they are made by a competent human. Training, collaboration and ethics must all be addressed, so that humans and AI are ‘teammates’.

To assure the safety of an AI enabled system-of-systems the key is to build trust in the system, and achieve ‘calibrated trust’, which is a balance of user trust and system trustworthiness. If it is not balanced, then there is either too much trust in the system, leading to ‘over trust’, or ‘under trust’, where the human does not believe the system. 

Calibrated trust as a concept can apply to all system-of-systems, including human centred ones. Frazer-Nash and University of Bristol have used a systems-engineering approach to design a set of ‘trustworthiness categories’, which can be prioritised for a system based on the concept of operations and risk analysis.

Applications of AI in rail

AI and autonomous ML intelligence systems are already being used in rail to deliver infrastructure management solutions.

Emily Kent, of One Big Circle presented ‘Putting the AI in rAIl’ and explained how the company is assisting railway engineers by using AI to analyse huge amounts of infrastructure data, covering for example: video and heat monitoring of overhead lines, troughing routes, vegetation, IBJs, conductor rail, and switches and crossings. She emphasised the importance of identifying what information engineers actually need to harness the data and ‘power up’ intelligence with tailored AI. This involved talking to engineers for ‘hundreds of hours before designing the AI system. 

Ansia Mamaniyat, director Chrome Angel Solutions and Jordon Langfield, fleet engineer at Angel trains, gave a talk titled ‘Train Engine Health Monitoring and Failure Prediction using AI’ which explained how ML techniques are being used to carry out deep analysis of Class 180 train data to predict engine health scores. Funded by Innovate UK, the scores are used to determine the priority of which trains to send into service, and tests have shown that 26% of Schedule 8 (unplanned service disruption) costs could be saved.

They emphasised that AI will not replace humans and jobs but will enable better targeted interventions to significantly reduce train cancellations, save costs, and deliver better reliability.

It’s important to gain user acceptability, collaboration, and trust in AI by addressing and providing:

• Transparency
• Privacy
• Safety assurance
• Unbiased outcomes
• Accountability
• Reliability
• Compliance
• Explainability

Rebecca Sellick, business development director UK and Europe, Cordel discussed “AI for railway infrastructure engineering monitoring – panacea for people and processes.” She said it was important to find out what the customer’s ‘pain points’ are when designing AI systems. For example, in Network Rail structure gauging is a pain!

Gauging is a complex issue involving the dynamic behaviour of a train and there are many limited clearance locations on the rail network. The lack of gauging information has caused problems with introducing new trains, and for example can prevent rail freight operators moving containers over a new route. Routes may be classified as not clear, but the reason and location why they are not clear may not be apparent.

It can take hours to manually gauge a structure, and even with accurate data the manual processing required is laborious and takes too much time. The Cordel Workbench is an AI-powered automated survey data processor that can be mounted on in service trains, and analyse LiDAR (laser point clouds), track position and other data gathered from third-party sources and technologies. The data (110 Gbit per sensor, per day) is translated using AI into accurate and comprehensive high-quality, survey-grade gauge profiles for entry into the National Gauging Database.

Nick Kotsis, chief data scientist at Network Rail, explained ‘Network Rail’s vision for AI – what can railway engineers do with AI for us’, and covered Robotics, Automation and Artificial Intelligence (RAAI). This offers a potential step change in the way asset data is managed and how infrastructure inspection and maintenance activities are performed.

The vision is that autonomous systems will monitor the network, providing Al systems the data to analyse and develop trends of asset-risk. This will enable decision support tools to schedule the most effective inspection and maintenance regime with minimal disruption. It is also expected that robotics and automation could be one answer in enabling more productivity in possessions and reduce the number of staff working red zone and other workforce safety benefits.

Benefits of AI in rail

The seminar provided a very comprehensive view of the benefits of AI in rail and the risks involved. The main line rail industry is very complex and open, which will limit the use of AI, so operationally AI may be deployed easier and sooner in more closed metro systems. Although as this article has shown AI and ML is already being deployed in UK rail.  

A report by the International Union of Railways (UIC) and McKinsey titled ‘A journey to building AI-enabled railway companies’ identifies the AI use cases that have been deployed, or have potential to be deployed. The report says there are more than 100 potential AI use cases for rail.

Use cases are at different stages of maturity, but the most mature use cases focus on shift planning and rolling stock predictive maintenance, while other uses cases include: energy efficiency; service scheduling; autonomous trains; real-time disruption management; predictive maintenance; capacity planning; real-time traffic management; inventory management; maintenance co-pilots; and network infrastructure digital twins.

The report says that AI could potentially unlock £10 billion to £16 billion a year, globally, for railway companies. However, as the IMechE / IRSE seminar demonstrated, good implementation is also key for realising the value that AI can offer to rail.

Image credit: iStockphoto.com