30.01.18
Making the railway safer with big data
Peter Hughes, principal research fellow at the Institute for Railway Research, argues big data is revolutionising close call reporting on our railways.
Close call reporting systems are an important part of the GB railways safety management process, as they allow any worker on the railway to report safety hazards at any time. When Network Rail introduced its system in 2014 there were some 180 close calls being entered daily.
Now there are more than 650 each day. Every one of these reports can contain important safety information that can help prevent an accident on the railway.
Close calls can be reported by workers telephoning in a hazard report, sending an email, or using a mobile phone app. The systems allow workers to report any hazard in any way they like using their own words. Unlike many hazard reporting systems, users arent limited to drop-down boxes to describe hazards; they can describe them in detail and provide any information they need to. The close call system is a powerful proactive tool that allows hazards to be identified and controlled – before an accident happens.
This power of the close call systems brings problems: with so many hazards being reported it is not possible for anyone to read every single report. And entering text on a mobile phone screen is notoriously difficult: it’s easy for typos to occur, and they do occur very frequently. Close call systems contain a trove of valuable information, but the question is, how can that information be brought together in a way that allows the railway to be made safer? This challenge is being addressed by researchers at the University of Huddersfield.
The team at Huddersfield is applying advanced computer techniques to analyse the close calls. The computer can read through every single report and perform millions of calculations to sort the data and identify patterns in the text. But whilst computers are excellent at analysing text, they have no understanding of the real world: for example, a computer has never worked in an infrastructure gang and doesn't know the difference between an MOP (member of public) and an MOM (mobile operations manager).
Interacting with the data
It is at this stage that a safety analyst interacts with the computer and teaches it the words and phrases that are meaningful to safety management. The safety analyst builds an information model (called an ontology) based on the information provided in the close calls. The computer learns from the analyst’s actions and updates its search results based on what the analyst has told it. In this way the computer and the analyst work together to find new information and grow the understanding of the railway. Over time, the computer becomes more effective at finding the right data and more and more information can be uncovered.
Since the knowledge model is built from the information contained in the close call reports, the computer learns to speak in the language of railway workers and can also learn how to read typos. Standard spelling correction tools that you get in word processors can’t be used on close calls because as often as a spell checker gets the word right, it will also get it wrong. In the trials we found the spell checker will turn the word faulting into falling, or the word unexploded into unexplored. Instead the computer learns to identify hazards from whatever it is being told by workers.
An early trial of the work has been performed on accident reports from the Swiss railways, which are written in German, French and Italian. These trials showed that the computer could correctly identify safety information with more than 98% accuracy, regardless of which language the accident report was written in. These results make the researchers optimistic that the computer should be able to understand the intricacies of railway safety in Britain. A simplified version of the analysis technique has been applied in Network Rail, and a full implementation is now underway in RSSB which should be up and running by the end of 2018.
FOR MORE INFORMATION
W: research.hud.ac.uk/institutes-centres/irr