Minding the Gap Automating infrastructure monitoring with artificial intelligence

Keeping our railway safe depends on reliable, up-to-date data about the assets adjacent to it. Network Rail must survey track-side assets to gather this data – but traditional methods are labour-intensive, expensive, and difficult. That’s why Network Rail enlisted Atkins to develop a solution: a platform capable of automating the output of surveys harnessing cloud point data to help it meet its regulatory duties and ensure that the tracks are safe – without the strain.

It’s a big responsibility, maintaining over 10,000km of railway track. The challenges, duties, and possibilities are as lengthy as the rail itself. Providing a safe experience – not just for passengers, but for all stakeholders – is a complex, never-ending process. One that requires continuous innovation and rigorous oversight.

One of the key elements of safety is rail infrastructure gauging. This ensures that vehicles can travel along a route whilst remaining a safe distance from lineside structures such as platforms, tunnels, bridges, signposts, and so on. In Britain, clearance distances are typically tight, which means that vehicles must be assessed to each of their routes, while trackside infrastructure must be continuously monitored and maintained. To this end, the Office of Rail Regulation (ORR) requires Network Rail to maintain a list of assets proximate to the railway, including a detailed ‘gauging profile’ recording their measured outline relative to the track. These gauging profiles are used to predict clearances between the infrastructure and rolling stock, avoiding unsafe operation.   

As the railway has grown, so too has the number of assets in its vicinity which need to be catalogued. Complicating matters further, track positions drift over time due to the forces imparted by rail traffic, so surveys must be periodically repeated to ensure their continued validity.

Technology has helped – laser scanners are in continuous use to generate gauging measurement data – but all the data has to be processed before entering the national gauging database. Currently this requires skilled technicians to select and categorise the measurement data, including determining the structure type: be it a bridge, platform, tunnel, and so on.

It’s a major bottleneck for Network Rail, with long lead times before the data is available. As a result, data may be out of date when it is needed by engineers, necessitating urgent catch-up surveys which impact project costs and timescales. Identifying and processing the assets is time-consuming, laborious work, and, with skilled staff in short supply, Network Rail is under constant pressure to keep up with infrastructure gauging and ORR’s regulatory requirements.

Automating better

Alternatively, new technology offers a better way of processing the railway gauging data, so that cataloguing the assets isn’t such a difficult task. Innovations in software, data, and internet infrastructure can automate the process of cataloguing assets. In November 2020, Atkins applied to an Innovate UK Small Business Research Initiative (SBRI) competition, which called for entries showcasing innovation in automated survey processing for railway structure gauging. We were successful in our funding application for both phase one and phase two of the project, collaborating with Network Rail to deliver a successful project outcome.

Our goal was twofold: to demonstrate innovations capable of enhancing Network Rail’s interpretation of point cloud data, and to accurately locate and identify trackside features from point cloud data, enabling accurate gauging clearance processing. This would help Network Rail to survey its network and assets remotely, providing enormous benefits: less time on track for staff would save money, as well as reducing the safety risks for staff. An automated process would help to reduce human error and it would enable Network Rail to meet the ORR’s regulations and de-risk the railway as a whole, providing an up-to-date inventory of all the assets in proximity to the tracks.

The new platform takes the survey data from Network Rail and prepares it (a form of data management which ensures it has the right IDs, etc.) before feeding into an AI system that processes it with a high degree of accuracy. The system deploys a set of tools that enables us to validate the data against the current national gauging database, where all the most recent gauging profile data is stored. This ensures that identified assets are matched to Network Rail’s previous records, enabling seamless replacement of gauging data.

“We have really enjoyed working as partners and co-investors with Network Rail bringing innovation to the gauging data processing stream of Infrastructure Monitoring. We have delivered an automated LiDAR data interpretation tool by training artificial intelligence with our leading gauging data expertise. This capability, combined with our proven ability to retrofit affordable sensors to in-service trains, is an exciting leap forward in improving the efficiency, timeliness, and accuracy of the gauging database process.”

Mark Fielding-Smith, National & Digital services Director, Atkins

Putting data in the driving seat

The machine learning model was trained using manually classified point cloud data. After assessing its initial accuracy, we classified more training data for the least-accurate assets and retrained the model. This allowed an iterative improvement in accuracy and this cycle was repeated until classification accuracy was maximised for each asset.

To deliver the best solution, we needed to build a bespoke product. We combined the computational geometry techniques into our own methodology for processing the geometry accurately. And, as even the best AI system will never be 100% accurate, we developed a human-in-the-loop checking of the AI outputs to pick up misclassified assets in the process, creating a ‘fail safe’ system where any errors would quickly be exposed and corrected, managing the risk downstream.

This solution enables vast volumes of data to be processed quickly, accurately, and without using large teams of people. What used to take a trained employee up to half a day to manually identify and produce gauging profiles, now takes the inference system a matter of seconds. By further training the AI using future project data, and by keeping the machine learning model at the cutting-edge using Atkins’ expert data science team, the accuracy of asset detection will continue to improve. The output geometries we create have already been validated against the baseline of contemporary survey methods in the national gauging database in a study overseen by the technical authority at Network Rail.

Yet the benefits from processing greater amounts of data more quickly extend still further. Atkins has applied rolling stock asset management retrofit expertise to improve the data capture process. We have successfully supported Angel Trains during the design, approvals, and installation of a LiDAR (Light Detection and Ranging) scanner, super HD Forward Facing CCTV, and associated computers and antennas on a GWR Class 165 unit. This equipment captures gauging structure data during regular passenger service and automatically relays the data for analysis. These projects show that, when engineering and IT integration capabilities are brought together, we can combine traditional and innovative ways of working to deliver high value benefits that maximise the efficiency of the operational railway.

It’s a work in progress, but survey data automation is a case study in how digital technology can facilitate intelligent automation that increases safety, efficiency, and productivity, eliminating repetitive and laborious manual processes that slow things down and sap momentum. All so that Network Rail can keep the tracks safe, without so much strain.