It’s just over 40 years since London Underground equipped a fleet with a data capture system with the ability to download into a central computer. It’s probably fair to say that the ambition exceeded the capability of equipment then available. Now, however, the only practical limit to the data that can be accessed is the capability of the communications systems between train and fixed infrastructure.
Today’s trains have a variety of sensors. Some, such as tachometers on every axle, are required to make the train work. Others are fitted to detect unusual events. All of this information can be monitored (sometimes several times a second), stored, transmitted, and used to understand whether the train is working correctly or whether it needs attention. Indeed, sensors provided for one purpose can yield information about other issues. For example (spoiler alert) a pantograph monitoring system can infer subsidence in the infrastructure. All these data sets are often described as being poured into a lake. Stretching the analogy too far, specialised skills and equipment is needed to explore the depths of a real lake and so it is with data lakes.
An IMechE seminar explored all of this in November 2025, with speakers discussing standards, experiences, the future, and how one might rely on the output of AI or machine learning.
Setting the scene
Vaibhav Puri, director of Sector Strategy and Transformation at RSSB set the scene discussing automation and the challenge and opportunity of scaling the use of Al models and incorporating Al into systems or equipment. There was an underlying theme of being able to assure that infrastructure is always fit for operation and absolutely minimising ‘boots on the ground’ except for actual interventions to repair or renew things.
One aspect he emphasised was the small size of the rail market and how it is important that rail avoids bespoke solutions wherever possible. Vaibhav turned to the safety and change control approach to introducing AI. It is not a one-size-fits-all approach and there is a need to make sure that human factors principles can be applied to the design and operation of AI-powered applications. RSSB is trying to help duty holders navigate the assurance process for introducing AI with a toolkit, although he made no apology for its complexity. Finally, he introduced the Landscape of AI Applications in Rail, inviting organisations to share their experience of AI to contribute to research project T1395.
Moving onto some practical examples, James Hill and Emil Tschepp from Transport for London outlined many of the automated systems collecting data about the trains and infrastructure.
Automated track measurement systems (including automated visual inspection) are fitted to one or two trains on each of six lines, and dedicated engineering trains or specially instrumented out of service passenger trains are used on other lines. The latter collects collector shoe interaction and contact force data, carries out thermal surveys, and has high definition forward facing cameras. It also provides opportunities to try out new measurement systems in a low-risk environment including a recent test of location measurement underground using quantum computing. Drone surveys are also being used.
James and Emil said that collecting the data is just the start of the process and how it’s used is critical, adding that use cases other than the prime purpose of the data are often identified. TfL’s data science team use existing data sets to drive improvements to safety, operations, maintenance, and cost and they often find other use cases using existing data. A recent example has linked more consistent operation of ATO trains to the increase in the amount of corrugation on curves.
Timothy Mangozza from PA Consulting explored the question: “Is it worth automating maintenance planning and operations?”
His premise was that infrastructure issues cause just under 50% of delays on UK railways, maintenance costs just over £2 billion, and that improved sensing techniques and AI could deliver many times that sum in value. Timothy outlined the landscape for techniques in other industries particularly citing the incredible progress made by life sciences in using digital twins to accelerate development.
He proposed a five-step process to make better decisions and make them stick: (i) move away from Technical Readiness Levels into System Thinking; (ii) adopt an investment portfolio mindset; (iii) create the change programme; (iv) understand the success criteria; and (v) sponsorship pre- and post-delivery.
Trains with brains
Jarek Rosinski, founder of Transmission Dynamics presented his vision of the future of rail infrastructure monitoring with the strapline ‘Trains with Brains – Predicting the Unpredictable’.
He illustrated the Transmission Dynamics product range, and described the capability of PANDAS-V®, a pantograph/OLE monitoring system. It uses a roof-mounted camera system with embedded sensors and onboard edge processing, synchronised with a pantograph-mounted wireless accelerometer to monitor accelerations and analyse collected video footage and images generated to detect faults in the pantograph/OLE such as detached droppers or excessive arcing.
Currently, much of this information is used in a reactive way but he discussed how the system might influence planning, become proactive, and then predictive. This included a future where every pantograph is monitored and connected via a gateway to a network of all pantograph monitoring systems. He explained Transmission Dynamics current work on the IntelliPan Network®, a fully connected ecosystem in which PANDAS-V systems communicate with the pantograph’s automatic drop device. In the event that an issue is detected, approaching trains within geofenced areas could be informed and would drop their pantographs before the location of the fault and raise them after.
Jarek discussed how analysis of the data from PANDAS-V fitted to revenue trains could be used to identify faults developing prior to potential derailments, building on the point that the TfL speakers had made additional use cases for existing data. While not the intent of the system, and based on post-event analysis, the data collected, for example, showed that the number of sway acceleration events in the range 0.05 -0.1g in the last 12 months suddenly increased in number over a two month period and, from the second month onwards, the values frequently exceeded 0.1g showing a clear adverse trend.
Company collaboration
Aamina Shah, systems engineer (infrastructure monitoring) at Angel Trains, and Lydia Parsons, senior account manager at One Big Circle, described how their companies have collaborated to provide infrastructure monitoring equipment and services across Angel Trains’ passenger fleet portfolio. Examples of collaborations to date include: Merseyrail’s Class 507 (thermal hotspot and conductor rail interaction monitoring); Southeastern Trains’ Class 707 (thermal hotspot and conductor rail interaction monitoring); enhancing a repurposed Class 153 (line-scanning technology to assess railhead treatment effectiveness); Avanti West Coast’s Class 390 (pantograph and OLE structure monitoring); and East Midlands Railways’ Class 360 (pantograph and OLE structure monitoring).
Their presentation detailed the collaborative steps and processes that are required to deliver enhanced infrastructure monitoring data. One case study explained that, using Southeastern’s Class 707 trains, a permanent installation of thermal and visible sensing technology has been installed for in-service infrastructure monitoring. A successful trial had been carried out using a portable equipment set on the tail lamp bracket, advancing to a permanent installation being designed and installed. It could not simply be fitted behind the windscreen because of the requirements for the thermal camera. Since commissioning, this system has identified loose third rail joints which, if not fixed, could lead to potential loss of electrical power as well as being electrically wasteful.
Future developments proposed include a full unattended track geometry measurement system on a Class 390; visual, thermal and shoegear camera installations on DC powered EMUs e.g., the Class 450; and further systems on Classes 357 and 360.
Digital transformation
Chris Beales, head of digital engineering at Porterbrook, and Jonathan Birch, technical director at Instrumentel, described how monitoring the engines on Turbostar units has resulted in improved performance. On the basis that the engines are the most complex part of these DMUs, and would yield greatest benefit, Instrumental’s data collection system was installed. They illustrated examples showing how the system identified faults before they might be picked up during maintenance. They can also save time. Analysis of in-service data eliminates the need for time consuming tests during routine maintenance, e.g., testing for air leaks. This eliminates the need to run the engines in depot and enables more work to be done in the downtime. They estimated that up to 2,000 maintenance hours per year could be saved for the average fleet.
Harry Shaw, mechanical reliability engineer at Cargill showed in a non-railway presentation how motion amplification can be used for detecting asset degradation. For your writer, this presentation was truly revolutionary and is very difficult to describe on the printed page. The principle is that vibrations can be detected but, when using conventional sensors, the results depend very much on the skill used in placing the right number of sensors in the right place.
What if, somehow, you could see the vibrations? This is the premise of motion amplification. It involves measuring movement not visible to the human eye using a high-speed machine-grade camera which uses processing methods and algorithms to turn every pixel into a sensor that measures vibration or motion. It is capable of sub-micron measurement. The results can be amplified for display and clearly show issues with movement/vibration where none was expected.
It is used by Cargill to:
- Visually demonstrate previously invisible phenomena that could previously only be captured graphically.
- Demonstrate the severity of an issue for key stakeholders.
- Identify issues, especially focusing on bad actors and repeat failures. It has become the first port of call on any asset for root cause analysis.
- More precisely fault find the issue in conjunction with traditional vibration analysis.
- To verify installation at commissioning of all new asset and CAPEX installations, e.g., it is now used as an exit criterion for project to maintenance handover.
Internet of things
Matt Weingarth, North Europe Director at KONUX, explained the benefits of using devices such as KONUX Switch Internet of Things to provide early warnings of issues. The device is battery powered with a lifespan of over five years, and is bolted to the track. It contains sensors such as accelerometers and, combined with ‘cloud’ based software, provides maintainers with a time history of impact loads and displacements. Rather than measure loads of a measurement train, this device measures the loads of the trains normally running on the line. This can be used to identify particular trains that might impart a much higher load than usual and detect gradual deterioration over time, e.g., crossing nose wear.
The machine learning software is trained to detect early signs of wear before visible defects appear, distinguish between different failure modes, guide the right maintenance action, and replace reactive repairs with an AI-driven predictive maintenance regime. The system is in extensive use with Network Rail.
Daniel Pyke, marketing lead at Sensonic explored vibration sensing using optical fibres. Fibre-optic cables are in use throughout industry as they are robust and reliable. They are also passively safe, resistant to EMC issues, and tamper resistant. They have been in use on railways for decades. Less well known is their use for sensing including acoustic/vibration, strain, and temperature. Daniel described the former, known as distributed acoustic sensing (DAS).
DAS uses trackside fibre-optic cable, turning the fibre into many virtual microphones listening to vibrations. It is effectively a continuous sensor which can cover up to 50 miles (80km). Daniel described several applications for DAS. The system learns the vibration signature of the ‘intrusion’ and delivers a precise location to within a few metres. Where appropriate, drones can be flown to identify the exact issue. Where permitted, drones can fly autonomously in response to a threat. Daniel showed several examples.
Positive train control
Ajtony Farkas, a principal engineer at Cordel described how the company’s trainborne LiDAR system is being used to carry out trackside asset mapping of Positive Train Control (PTC) lineside assets on railroads in the USA, as part of a programme called PTC Asset Connect. PTC was mandated by the US Federal Railroad Administration and was installed between 2008 and 2020. PTC is designed to trigger a brake application if it detects that the engineer (driver in UK English!) doesn’t start braking early enough before a signal at danger.
US railways are required to carry out an annual audit of all relevant assets including: signals, switches, speed signs, mileposts, derailers and level crossings. The LidAR system collects this and a great deal more information as it collects its point cloud. The point cloud data can be analysed to identify individual assets and, on subsequent runs, changes can be identified, e.g., mile post collapse, or obscured by vegetation encroachment. This data will be added to a master PTC database which combines mobile LiDAR inspections with other material. The tools were developed over the summer 2025 which involved training the AI machine learning system to recognise new asset types specific to PTC. The information available to operators of the system combines data from various data sources to allow better decision making.
Rough riding
A rough ride on a train might arise from a train issue, a track issue, or a combination of both, so a holistic approach to identifying the cause is needed, as Mani Entezami from the University of Birmingham and MoniRail Ltd explained. Rough riding can result from the interaction of track and train issues and, as dedicated measurement trains do not always behave in the same way as the regular trains on the line, they do not provide enough of the right information. It is therefore more valid to fit equipment to a line’s regular trains.
MoniRail’s system, which uses three types of vehicle-mounted sensors as shown in Table 1, provides more frequent results that are comparable with those obtained from Network Rail track recording vehicles. Mani described several case studies covering fitment to a freight vehicle, Classes 158 and 170 DMUs, and high-speed trains. He showed the results and reasons for rough ride events that the system identified, including short wavelength faults or squats, hunting linked to high equivalent conicity, and the consequences of long wavelength track faults on high-speed lines.
Cohesion is key
The presentations and discussions clearly demonstrated that there is a wide array of advanced technologies available to support railway maintenance, such as IoT devices, distributed acoustic sensing, trainborne LiDAR, and sophisticated vehicle-mounted sensor systems.
However, for these innovations to truly benefit the railway, it is essential that all these tools and the data they generate are seamlessly integrated into a cohesive platform. Only by combining these technologies into an accessible and actionable system can maintainers efficiently diagnose issues, prioritise interventions, and ultimately improve the safety and reliability of rail operations.
Image credit: iStockphoto.com