Artificial Intelligence

Future of Rail | Artificial intelligence helps Network Rail’s maintenance regime


The use of artificial intelligence in rail is growing. Here we look at the latest initiatives Network Rail and its engineering partners are undertaking as they seek to transform rail project delivery and asset management.

Up until now, Network Rail’s complex infrastructure has largely been delivered and managed by humans through, for example, manual calculations and site inspections.

But artificial intelligence (AI) is changing all of that. 

“There is massive potential for us to use AI not only to better understand issues around rail infrastructure, but also to create solutions to solve them,” says Network Rail head of maintenance Tim Flower.

“The key thing is to make sure we train those AI algorithms in the right way – making sure we’ve got the right insight in there.”

Plain line pattern recognition

Network Rail’s plain line pattern recognition (PLPR) system is the closest the track operator has come to fully embedding such technology. The system automatically films the railway, using lasers and cameras attached to trains. 

This data is then analysed using machine vision algorithms, which compare what the cameras see with an image of how the track should look. This identifies defects or issues that require further attention – such as missing Pandrol clips, which secure rails to the sleepers.

“Then a human sat in an office – as opposed to a human walking the track –  confirms if it’s a defect or not,” explains Flower.

A drone captures LiDAR data

This automatic identification of issues has the potential to increase efficiency.

“If we can automatically process data and present it to engineers, then they’re much better informed,” says Flower. “We can tell them what they need to know when they need to know it, which will then enable them to concentrate on staff capability and onsite supervision and really drive that quality result we’re striving for.”

The system was rolled out throughout the last control period and has been trained to ensure it does not miss anything, which has meant occasional false positives – where images have been misinterpreted – have been an issue.

Flower explains: “It’s based on machine vision, but we have started to do some machine learning to try and improve the accuracy.”

“The machine learning is about removing the false positives from the output, so the human inspectors are only looking at genuine things they need to review.”

Consultant Atkins has experienced similar false positive challenges in the development of the aerial survey
data work it has carried out with Network Rail.

The consultant maps the ground by capturing LiDAR (Light Detection and Ranging) data via light beams sent from scanners mounted on drones or helicopters. Height information is gathered for each point scanned and this data is classified into different categories, for example, ground or rail tracks. Analytical models are then applied to inform asset management decisions and to highlight which areas of track should be monitored.

However, it has again been necessary to limit the false positives. Along with gantries spanning rail tracks, motorway gantries and football goalposts could also be picked up if the machine learning algorithm is too aggressive.

“It’s finding out where that tolerance is for accuracy levels,” says Atkins technology enabled services senior analyst Jonny Corker. “What is and isn’t acceptable to miss.”

Earthworks and vegetation

Use of this aerial survey data has led to two different initiatives.

The ground points from the LiDAR data are used to build a digital terrain model (DTM). Terrain models from different years are compared to show change over time, and these changes are categorised using machine learning and cloud computing to detect and quantify potential geotechnical problems.

“We can understand where the ground has moved over that period of time and where there is potential instability – or where there have been works in the past and there might be more instability further down the line,” explains Atkins digital asset management engineer Fatema Walji.

The second initiative focuses on managing vegetation encroachment. Network Rail manages its vegetation using a specification that defines encroachment zones, which indicates the necessary track clearance to allow trains to pass safely. 

The rail operator has worked with Atkins to create these zones and then intersect them with vegetation LiDAR data to understand where on the network vegetation is encroaching. 

“Then managers can go into that specific area and carry out the management required,” says Walji.

These initiatives are part of Network Rail’s wider research and development portfolio, and the aim is to be able to eventually roll them out across the organisation to help end users make more informed asset management decisions. 

Leeds masonry viaduct

Meanwhile, a Cambridge Centre for Smart Infrastructure and Construction (CSIC) project for Network Rail has used an automated network of a Fibre Bragg Grating (FBG) sensing system, acoustic emission sensors, and high-sensitivity accelerometers to gain insight into a damaged masonry viaduct in Leeds. 

The fibreoptic sensing in the FBGs provides dynamic strain data at multiple locations simultaneously, identifying the extension and contraction of the bridge when
trains cross it. 

In addition to this, acoustic emission sensing detects high frequency waves that indicate cracking. These high frequency waves can be easily distinguished from typical vibration frequencies of bridges.

Vegetation encroachment detection survey used by Atkins and Network Rail

CSIC co-investigator Matthew DeJong explains that this makes it possible to “detect which cracks are active and propagating and which have been there for 30 years but aren’t a problem”.

This sensor data has allowed CSIC to research new approaches to asset management, in collaboration with the Alan Turing Institute.

For example, algorithms developed for the bridge have separated seasonal variation from other long-term data trends, identifying areas on the viaduct where deterioration is actively taking place.  

“Because of changes in temperature and humidity, the part of the bridge that moves when a train goes across changes with the time of year,” says DeJong, who is also assistant professor of structural engineering at the University of California, Berkeley. 

“We’ve been able to unlock some of those things and we already see locations where continued damage is occurring.”

Another Network Rail development project is exploring the correlation between weather data and asset failure. 

Network Rail R&D manager Rob Forder explains: “A human being can say when it rains, I can see the trend of this type of asset failing. We’re currently working on a predictive model that can say: ‘this weather’s coming, these are the hot spots in the areas you’re likely to see failure’.”

Training 

Once a model understands the variable factors, it will learn. 

“So, you have a period of a weather type, you’ll have an asset failure and it will go back into the model to learn something you haven’t taught it,” says Forder. 

“You give it some basic training, and it starts training itself.”

When the model has been trained in this way, the AI is repeatable – and once it gets to the stage of being reliably right the first time every time, standards rise. 

“It’s a bit like training a human being,” says Forder. “The more someone does a task, the more they recognise the subtleties in it. With machine learning, the more information you put through it, the better the algorithm becomes.” 

Scale and proactivity

These methods also allow larger amounts of data to be processed at faster speeds than previously possible. 

“The ability to crunch huge data sets increases the value of those data sets,” says Corker. “You can be more proactive. You have new tools in engineers’ toolkits.”

From a Network Rail perspective, Flower highlights the importance of this proactivity. 

“We should be able to much better inform the managers of the asset,” he says. “We’ll start to see when things are trending towards an alarm, as opposed to waiting for an alarm.” 

DeJong also emphasises this. “Before it was reactive. Now, we detect things much earlier and act on them when it’s most appropriate.”

He adds that looking for trends in large amounts of high dimensional or multi-parameter data is challenging manually, but that AI allows teams “to look for trends in the data that we otherwise might not have been able to capture”.

As such, AI has the potential to eliminate much of the groundwork and calculations that stops design teams focusing on what they are trained to do. 

HackPartners and Network Rail are for example, currently building technology which will use AI to automate station design work. 

The NextStation tool would allow planners to enter specific parameters relating to the number of passengers entering or leaving a station, then AI would design the station to suit those.

“It would speed up the design process of stations as well as make it much cheaper,” explains HackPartners chief executive River Tamoor Baig.

When it comes to operations, Baig believes “we will have missed a trick if in a decade we still have humans solely making operational decisions”. 

Overall, Forder believes it is not about “removing the need for people”, but about “removing the repeatable tasks that people find quite mundane”.

Corker agrees. “It will help people make decisions, not make decisions for them,” he says. 

“An algorithm is very difficult to hold responsible for its decisions. Everybody understands that. It’s your name that’s going to be going on the bottom of the report.” 

It is this collaboration that will enable AI to effectively support the rail management of the future.

Like what you’ve read? To receive New Civil Engineer’s daily and weekly newsletters click here.



READ SOURCE

This website uses cookies. By continuing to use this site, you accept our use of cookies.