“Machine Learning is at the heart of every ’smart’ adaptive system we may think of, and already delivers huge tangible industrial benefits. It will be key for delivering the railway of the future, and it can help the industry meet some of its current challenges here and now. The main difficulties are likely to be around ensuring data access and securing the right partnerships with data experts. Ultimately, it is our responsibility as an industry to identify the challenges where Machine Learning can help us.” Olivier Marteaux, Principal Horizon Scanning.
Machine Learning may furthermore provide the keys to solving complex socio-economic problems. This view seems to find confirmation from looking at the list of the Alan Turing Institute’s research projects, which cover, among other topics, healthcare diagnosis, urban analytics, air pollution in London, immigration and labour market analysis, cybersecurity, nuclear and offshore platform safety, resilient networks, and large transport systems.
What can Machine Learning specifically bring to the rail industry?
Machine Learning is an enabler of many of the capabilities underpinning the Rail Technical Strategy, but it is still at a low readiness level for the rail industry, with few solutions truly deployed on the rail network and delivering benefit to the rail customers. How can rail start to reap the benefits of this emerging technology? The first step is to have a high-level understanding of the nature, the scope, the techniques, the potential and the limits of computer learning, for they sometimes remain arcane or misunderstood.
It is important to understand that Machine Learning does not try to imitate human behaviour, but rather uses the statistical strengths and the data crunching abilities of computers to complement human intelligence. As a complement and enhancer of human expertise, Machine Learning applications should focus on its comparative advantages. Machine Learning is particularly suited for tasks too complex to program that go beyond human ability, such as the analysis of the multidimensional relationships within and between large and complex data sets: real-time aircraft engine digital twins and sensor data, astronomical data, network traffic, meteorological records, medical archives, behavioural patterns, consumer data. Machine learning also shines for tasks that that humans can do, but that defy traditional programming because they require strategic flexibility and adaptive behaviour (driving decisions, speech recognition, real time optimisation, etc).
We can often hear critiques, doubts and fears regarding Machine Learning and AI in general, that fail to see that Machine Learning is mostly Data Science: it consists of advanced analytics packages and algorithms to recognise or discover patterns in data, classify events and objects, compute probabilities, forecast values and predict outcomes based on past data or on simulations. As such, the “black box” criticism – ie lack of transparency on how the system reaches its conclusions - only applies to a small portion of Machine Learning systems (mostly neural networks), as the majority of the techniques used by Machine Learning are underpinned by statistics, probability theory and linear algebra. The system’s learning path and decision-making is therefore traceable and under human control.
Perhaps the true challenges with Machine Learning have more to do with the data than with the algorithms: how to give the system access to a sufficient volume of data for it to be trained in the first instance and ongoing data for it to continue to improve, and how to select the right statistical techniques for the data at hand, should be our primary focus. Expertise of both data scientists and subject matter experts is needed for the success of Machine Learning initiatives.
To help the industry on its Machine Learning adoption journey, this blog series will explore how it can bring value to:
- engineering with Machine Learning at the core of automating inspection and predictive maintenance
- operational performance, where Machine Learning can play a role in train delay prediction and mitigation
- safety with the automated classification and analysis of safety-related records
- customer experience where the opportunities and expectations raised by progress with Natural Language Processing and automated real-time passenger interaction are significant.
We will also reflect on some “do”s and “don’t”s, and give some insight for a strategic business approach. Machine Learning is not necessarily an expensive investment, however it requires thoughtful planning, and, as hinted above, it relies on ensuring adequate data access, securing the right partnerships with data experts, and knowing what industry challenges can benefit from Machine Learning more than others. Ultimately, it is our responsibility as an industry to identify the challenges where Machine Learning can help us.
In the next article, we will look at different categories of Machine Learning techniques and systems to ensure that in the rest of this series we use a common language, before we move on to look at this in the context of rail.