Driving the AI train of thought

Today, AI is hailed as our era’s greatest breakthrough, the all-purpose solution to social challenges. No wonder: its potential to unlock hidden efficiencies, integrate disparate sources of data, and reveal deep insights can help our sector to surmount obstacles and deliver on its promises. Fuelled by a need to adapt to digital transformation, climate change, Net Zero, limited public funds and shifting customer demands, the rail industry can no longer rely upon traditional ways of working. Emerging technology is upending established patterns of travel and if our sector is slow to adapt, we may find ourselves increasingly irrelevant. The need for evolution is urgent.  

To harness AI, it must be applied to tangible problems, serving the needs of the rail industry and consumers. We must ensure the technology is applied in safe, effective, and sustainable ways - lest it increase risk and add complexity. This represents a significant opportunity for organisations that have a deep understanding of rail, technological innovation, and integration alike. 

As with all breakthrough innovations, the hype around AI can misrepresent its value. The challenge is applying AI from LLMs to areas where they can be of service to rail and by extension, communities. Without careful identification of specific challenges, it’s hard to develop a strong use case - and without a use case, the rail sector will struggle to adopt AI in a meaningful way.  

Benefits-focussed innovation 

Asset management and operational decision-making are two areas that have long been ripe for greater application of AI, based on our experience with automation and machine learning. Both are beset by generic standards, labour-intensive processes, and reactive practices - all of which could yield great efficiencies. Asset degradation is a prime candidate for AI innovation. AtkinsRéalis is using machine learning to detect deterioration in asset health. Traditionally, this laborious process requires surveyors to manually inspect or review vast amounts of remote video footage. Alleviating this time-consuming task and delivering efficiency were prime motivations for our teams in seeking machine learning solutions.  

We applied the AI software deployed on NASA’s Mars rover to asset management in the water industry for exactly this purpose, using historical video data guided by our specialist engineers to train machine learning in defect detection to target maintenance activities. The model learned to identify cracks underground, automatically sifting through thousands of hours of footage and flagging seams for surveyors to review. By focusing on a specific task, building on an existing model, and harnessing data, we deployed a machine learning solution that liberated capacity and increased efficiency on an important problem. This highlights the value of applying AI with experienced engineering SMEs to solve tangible problems, rather than trying to reorganise our sector around the hypothetical benefits AI might yield. This approach delivers real benefits faster - which in turn makes it easier to win backing for further innovation. This ‘time to value’ represents where best to deploy limited resources.  

We took this proven approach to improve efficiency and accuracy of rail gauging, demonstrating how machine learning can be deployed in rail to solve critical challenges while yielding rapid time to value. For trains to be put on a path, they must be assured that the train fits safely through the infrastructure space envelope - this industry process is called gauging. Previously a manual process involving in-person surveying of trackside hazards, we developed an industry-first automated gauging solution. It applies machine learning to Network Rail’s point-cloud data, providing accurate location and classification information for trackside features. This enables accurate gauging clearance processing, by checking and validating the data against the current national gauging database. Tangible problems, existing data, models trained by our engineering experts and agile modular deployment: that’s how we deploy AI with reduced time to value, unlocking efficiencies for our clients without compromising safety. 

Driven by data 

Ensuring AI serves our industry ultimately matters insofar as it serves our users. Innovation is changing these relationships, opening new possibilities for us to facilitate user needs and exposing the limitations of our existing systems. In a post-Covid world, where mature rail markets are facing reduced commuter demand, understanding travel characteristics is critical to driving ongoing usage. This depends on data, and as data increases in volume, diversity, and complexity, machine learning models will be needed to accelerate the processing and evaluation of it to enable timely decision making.  

Generic models can be replaced with specific portraits, not just with assets but with passengers too. Data is revealing that younger generations are less likely to drive or to aspire to vehicle ownership1. Train stations with disproportionately youthful users can use these insights to better anticipate demand and support the business case for investing in supplementary travel options for first and last mile, such as scooters and e-bikes. Evidence of demographic shifts can help owners and operations increase provision for older passengers. 

As well as anticipating demand, better usage of data - and the machine learning models analysing it - can help to shape it, harmonising transport modes with people’s behaviour, and vice versa. ‘Nudge’ theories, popularised in the 2010s, are only just getting started. Prompting people to travel at different times to reduce peak pressure is the most obvious application. Many trains now ask passengers questions about the train (how’s the cleanliness of your seat?) or nudge them towards choices that could improve experience and lead to greater likelihood of repeat usage.  

Yet in order to be impactful, such insights must be integrated. The more connected rail is to other modes, the more flexibility our industry will have to adapt to particular circumstances. Improved integration with the bus network can prevent passengers shifting to their cars on busy event days. Just-in-time alerts can inform passengers of problems prior to their arrival at the station, easing congestion and reducing complaints or risk of injury. Partnering with mobility-as-a-service providers can enable operators to orchestrate other modes, from taxis to e-bikes. Different countries are in different stages of digitisation, autonomous systems and trains. But already, expectations are shifting, and what was good enough a few years ago is quickly becoming outmoded.  

Safety and regulation 

If hype is one side of the AI innovation coin; the other is fear. Connecting transport infrastructure to the internet also increases exposure to cyber risks. Digital systems mean trains are no longer just trains anymore. Digital-driven autonomous decision making affects efficiency, passengers, and security. Balancing these will be critical. 

Already, high-profile cases of AI ‘hallucinations’ - where otherwise-cogent machine learning models confidently state fictitious statements as fact - have exposed AI’s present limits, and the potential for confusion. Errors that are amusing in a personal context have the potential to create real world problems when combined with critical national infrastructure. Even if our doubts are only doubts, they may still prevent these technologies from winning the widespread backing they need to improve. Ultimately, unless we are assured and completely confident in the safety of machine learning models, we will be unable to deploy them and realise the benefits, still less at the necessary speed. Rail is highly-regulated and safety is paramount. It takes time and expertise to build trust, which can be lost quickly. We cannot risk operations or safety by overeagerness to adopt innovation without safe frameworks for deployment.  

Much depends on testing. For example, in the UK, AtkinsRéalis operates the National ETCS Test Laboratory, independently assuring and verifying new safety critical equipment from multiple OEMs against European and British standards. The service encompasses ‘white hat’ hackers - employed to stress-test the security of their digital systems. Testing AI’s fitness for purpose in the R&D phase of the lifecycle, alongside cyber-resilience, is the logical next step, and as important for the industry as the technology itself.  

Integrate to accelerate 

AI’s potential is as vast as the hype claims it to be. And for rail, the possibilities are significant. Reducing downtime, increasing capacity, improving safety, more effective timetable planning and improving the passenger experience: all are within reach. However, we can only unlock these benefits if we deploy AI wisely. That demands careful consideration of the problem to which it is being applied, improving the time to value and increasing the chances of delivering return on investment.  

It also demands thoughtful integration with existing standards, tools, and technologies. Rail organisations must find partners capable of connecting industry needs with technological capabilities, understanding the safety and operational context so that AI works for us. Thanks to the insights afforded by data, we can now map innovation onto real human needs, rather than speculative technology. We can enhance the communities and economies which depend upon transport, restoring both rail and transport to the forefront of human ingenuity.