Predictive rail asset maintenance: from concept to reality

Rail infrastructure managers are under sustained pressure to improve reliability, safety and efficiency across increasingly complex, busy networks. Simultaneously, many maintenance decisions rely on reactive or scheduled inspections, predefined alarm thresholds and manual interpretation of large volumes of asset data.

These approaches have underpinned the development of modern infrastructure management, but the wider context is now changing. Networks are carrying increasing amounts of passengers and freight, maintenance access remains constrained, regulated and costly, financial pressures are intensifying, and teams are expected to extract greater value from existing assets and resources. In this context, existing ways of working are increasingly limited.

But rail has the foundations in place for a different approach. Remote Condition Monitoring (RCM) systems now collect enormous quantities of asset data. Additionally, advances in machine learning make it feasible to analyse that information at a scale and consistency that would have never previously been possible.

The opportunity is to combine these two advantages into reliable condition prediction. To make this a reality, Tracsis is actively developing an advanced predictive monitoring capability as an extension of its established RCM ecosystem to:

• Identify patterns associated with emerging faults before usual alarms are triggered

• Monitor thousands of assets simultaneously and direct engineers towards those presenting the clearest signs of developing risk

• Support more targeted inspections and interventions, reducing unnecessary maintenance activity and trackside access

• Improve asset reliability and potentially reduce service-affecting failures

• Provide a clear picture to infrastructure managers, giving a more consistent basis for balancing safety, performance and maintenance cost

• Extend RCM’s usefulness from an operational visibility platform into an intelligent maintenance decision-support capability.

The technology is in active development, but reaching the accuracy and reliability required demands extensive testing and high-quality data. To achieve this, Tracsis has developed a predictive platform for deployment in the field with select client partners. This is enabling the next key phase of training, evaluation and model refinement using data and feedback from some of the world's busiest rail networks in live operational environments.

This article explores why rail asset maintenance is at an important point of change, the opportunity created by predictive technology, and how Tracsis is making significant progress to bring the next generation of rail asset monitoring into operational use.

The operational challenge

Railway maintenance has always required a difficult balancing act. Infrastructure managers need to identify and address deterioration early enough to protect safety, reliability and network performance, but intervention itself also carries a cost.

Inspections require engineering resources, trackside activity can demand detailed planning and safe access arrangements, and component replacement generates labour and material costs. On a busy railway, maintenance activity may also compete with the operation of passenger and freight services for access to the infrastructure. In combination, these factors make both late and unnecessary intervention costly and time-consuming.

To prioritise work effectively, infrastructure managers need to answer one fundamental question: which assets require attention, why, and when? On modern rail networks, the answers are usually determined by one or more of three progressively sophisticated approaches:

1. Reactive maintenance, where an asset operates until a fault or failure requires intervention.

2. Preventative maintenance, where work is carried out at predetermined intervals in an effort to pre-empt failure.

3. Condition-based maintenance, the most advanced model widely available, where RCM platforms track asset data and help teams intervene according to the observed state of an asset.

But even with these strategies, asset insight is largely limited to real-time condition or performance, severely limiting efficiency across large-scale networks and thousands of assets.

The clear opportunity is therefore to become more targeted, moving away from legacy models and increasingly towards risk- and condition-led decisions, in part enabled by emerging predictive technology and represented by another stage of the progression scale:

4. Predictive maintenance, the next technological step, where historical and live condition information is analysed to anticipate deterioration and indicate when intervention is likely to become necessary.

Why now? The conditions are set for industry transformation

Predictive maintenance has been discussed in rail for years, with very little to show for it in terms of deployable solutions. Now, the combination of better data, new technology and growing operational need has meant viable predictive maintenance is moving from concept to reality.

Operators globally are entering a generational push for modernisation, with examples such GBRx in the UK specifically prioritising the exploration of artificial intelligence in various use cases, including asset management.

Growing pressures: a need to do more with less

Several changes are making innovation increasingly important. The first is network utilisation. This growing use of infrastructure makes reliability more crucial than ever, while putting greater pressure on the assets supporting the services. And in a purely practical sense, a busy operational railway is harder for engineering teams to access.

Secondly, financial pressures are growing. Globally, the rail industry is being asked to deliver more, with greater efficiency, all within a challenging cost environment being driven by inflation and other financial constraints. Maintaining existing assets efficiently and sustainably helps operators achieve savings, so the pressure inevitably cascades down to on-the-ground infrastructure teams.

Finally, the technology needed to approach maintenance decisions differently is now developing rapidly. Artificial intelligence and machine learning technology are now vastly more affordable, accessible and capable than ever before, meaning the pressure to utilise it – and the opportunity cost of not utilising it – are building.

Better data: Remote Condition Monitoring has laid the groundwork

RCM has already changed the relationship between engineers and infrastructure. Rather than relying exclusively on physical inspection to understand asset behaviour, engineering teams can remotely view data from geographically dispersed equipment. Depending on the asset, this might include measurements such as electrical current, hydraulic pressure, position, movement, operating time and other technical variables. This provides a much richer picture of condition than was previously possible.

Current RCM approaches typically use configured rules and thresholds to help teams identify potential problems. When a measurement moves outside an expected range, an alarm can direct an engineer towards the affected asset. This is hugely valuable, but it has limitations as a threshold only provides part of the picture.

An alarm can show that a defined limit has been reached, but determining what that behaviour means may still require manual analysis. A skilled engineer will often need to compare a trace with previous operations to understand whether the change represents a real fault, a temporary variation, or an issue that doesn’t warrant immediate intervention.

Additionally, thresholds may flag more assets than limited experts can quickly assess. Across a small asset network, detailed manual interpretation may be practical. Across thousands of monitored assets generating a continuous stream of data, it may be prohibitively time consuming for engineers to perform consistently.

Essentially, RCM has created both the foundation for progression, as well as the next challenge. Rail now possesses far more condition data than its engineers can reasonably analyse manually. 

Technology: machine learning changes what can be done with the data

Modern machine learning technology provides a way to analyse repeated asset behaviour at a vast scale. Many rail assets generate recognisable traces during normal operation, but as components deteriorate, become contaminated, lose pressure, experience increased resistance or develop other faults, characteristics within those traces can begin to change.

Some changes are immediately obvious, while others are gradual or extremely subtle. A single unusual operation may have little significance, while repeated deviations, changes in the shape of a trace, combinations of measurements or a progressively changing pattern can provide much stronger evidence that an asset is deteriorating.

Machine-learning models can be trained on historical examples of normal behaviour and confirmed failure modes to identify these relationships. When similar patterns appear in new data, the technology can flag the asset for further investigation.

The significance of machine learning is its ability to extract additional intelligence from the information rail networks already collect, enabling experts to prioritise manual assessment and make more targeted maintenance decisions.

The opportunity: clarifying the data

The major potential benefit presented by predictive monitoring is earlier and more focused decision support for a given asset. Take a hydraulic points system as an example, which could fail for a variety of reasons, including hose faults, leaking pumps, lock-side seizure or contamination affecting the mechanism’s movement. Different faults may alter pressure, electrical current or movement characteristics in different ways.

A traditional alarm can identify when a measurement eventually exceeds a predefined threshold. A predictive model, however, has the potential to recognise the developing pattern before that threshold is reached.

By analysing large volumes of historical traces, a specialised model can learn which combinations of characteristics have previously accompanied particular faults. The objective is to recognise those patterns with enough accuracy and warning to help an engineer determine whether inspection, repair or another intervention is appropriate.

The central advantage of this is the scale of analyses it enables, rather than automation of expert decision-making. An experienced engineer may be capable of identifying subtle deterioration by studying the operating history of an individual asset in detail, but applying the same scrutiny continuously across thousands of assets is unrealistic.

Machine-learning algorithms can evaluate those assets simultaneously, compare developing behaviour against a far larger body of historical evidence and prioritise the cases most likely to warrant human attention. Engineers’ knowledge and discernment remain absolutely critical to that process, but predictive monitoring has the potential to present the most relevant information to the experts who can then concentrate their attention exactly where it is needed.

Instead of asking an engineer to find the significant trace among thousands of routine operations, the technology can help surface the assets showing the strongest evidence of developing deterioration and provide the supporting information required to investigate them.

Moving from more alarms to better decisions

The ability to surface asset insights earlier is essential for successful predictive monitoring, but more alarms isn’t the answer. Maintenance teams rarely need another source of unqualified alerts, and a predictive system that simply generates additional alarms based on hyper-sensitive measurements would only increase workloads, not reduce them.

In short, useful predictions need context. An effective predictive alert could factor in and provide that context by combining varied information including:

1. Asset type, location and history

2. The pattern or behaviour detected

3. Similarity to historical examples of known faults

4. How the behaviour has changed over time

5. Confidence in the prediction

6. Asset criticality

7. Operational consequences of failure

8. Supporting information for an engineer to review.

This creates the basis for a much more sophisticated maintenance workflow. An emerging problem on a highly critical asset could be escalated and scheduled for inspection at the next suitable opportunity. A lower-confidence indication on a less critical asset could remain under observation while further data is collected. A stable asset showing no meaningful deterioration could be monitored remotely with no need for immediate intervention. The objective is to make intervention more proportionate to actual condition and risk and a well calibrated predictive system could help teams find a better balance.

What predictive maintenance could change for infrastructure managers

If predictive monitoring can be demonstrated to operate consistently and accurately and can be integrated effectively into maintenance processes, its impact could extend across reliability, workforce productivity, cost, safety, and passenger experience.

Earlier warning and more controlled intervention

The clearest potential benefit is providing infrastructure teams with more time.

Identifying subtle deterioration earlier gives maintenance teams more opportunity to investigate a potential fault, understand its significance and decide how and when to intervene. This could make all the difference between planned work during an appropriate access window and an urgent, inconvenient and service-disrupting response being required following a failure.

While predictive insight will never eliminate unexpected failures completely, it could increase the number of cases where emerging deterioration is visible early enough to support a planned response, avoiding disruption.

More targeted maintenance

Fixed maintenance intervals are still appropriate for some asset types, but time elapsed since an inspection does not always correspond neatly with asset condition. For example, two seemingly similar assets can deteriorate at vastly different rates simply as a result of operating frequency, location, environment, installation history and numerous other variables.

Condition monitoring and predictive insights can help teams distinguish between assets requiring earlier attention and those that continue to operate within expected parameters.

Better use of engineering expertise

Predictive monitoring also has the potential to change how the time of highly specialised engineers is used. As the volume of monitoring data grows, asking experienced engineers to manually examine every asset trace becomes increasingly inefficient.

Models capable of identifying and prioritising unusual or deteriorating behaviour can reduce that burden. Engineers can then focus on interpreting higher-value cases, assessing operational context and using their expert judgment to decide on appropriate responses. 

Reduced unnecessary trackside activity

Greater certainty about condition can also reduce inspections being undertaken due to uncertainty about asset condition, helping teams identify what requires attention before sending teams onto the infrastructure. This could positively impact safety, productivity and working conditions for pressured maintenance teams.

Making predictive monitoring a reality

The potential of this technology is clear, but building a predictive model is only part of the challenge. Rail is a safety-critical environment, and a vital pillar of global economies, so any model used to influence maintenance decisions needs to demonstrate sufficient accuracy, consistency and explainability to give engineering teams confidence in its outputs. This makes the quality of the underlying data and the way models are trained fundamentally important.

A reliable machine-learning model requires large quantities of asset data, but beyond that it needs relevant examples of healthy operation, natural operational variation, developing degradation and confirmed failures. Those examples also need to be labelled accurately enough for the model to distinguish meaningful fault characteristics from routine variation.

Models also need to cope with differences between individual assets and the broader operational context around them. Equipment of the same general type can operate differently because of configuration, age, environment, usage or installation, so a model performing well on one set of assets cannot automatically be assumed to perform reliably elsewhere. The only way to address this is to learn from rail experts and actual operational environments.

This means predictive solutions must be built in close partnership with industry, from concept to delivery. Collaboration with infrastructure teams is vital to understand the exact operational challenges and context, access training data, and define the solution. Then, real-world testing is required to evaluate predictions against known outcomes, understand false positives and false negatives, refine how different failure modes are identified and determine how predictive outputs should fit into existing engineering processes.

Beyond the technical design, the user experience is also extremely important. Even a technically accurate model will deliver limited value if engineers cannot easily use the solution or quickly understand what has been detected and what evidence supports the prediction. Any truly useful predictive capability therefore needs to sit seamlessly within practical maintenance workflows, presenting information in a way that supports rather than complicates decision-making.

How Tracsis is positioned to deliver the next era of asset monitoring

This necessary combination of high-quality data, rail expertise, frontier technology and close industry collaboration is where Tracsis has a significant advantage.

Predictive monitoring is a natural extension of Tracsis’ RCM platform – and accompanying Tracsis hardware – already deployed across some of the world’s busiest passenger and freight rail networks, where it has processed billions of asset data points over many years of operation.

This industry leading asset-monitoring ecosystem provides the established wide-scale infrastructure needed to collect, process, visualise and manage information from rail assets. Predictive technology adds a new analytical tier to that, learning from historical and live behaviour and using that knowledge to identify emerging risks within new data.

Our development work benefits from a huge, varied and high-quality volume of rail asset information, continually supplemented by live monitoring data. This provides an exceptionally strong base for training models against real asset behaviour rather than simulated conditions that lack full context.

Beyond the data, Tracsis brings extensive rail expertise and innovation built over more than two decades, a proven history of leveraging machine learning to improve asset insight, an industry-leading RCM platform, and direct relationships with the infrastructure teams who will ultimately use the technology. Together, this puts Tracsis in a uniquely strong position to deliver predictive insights that will give teams the confidence they need to make decisions.

Development alongside working infrastructure teams

Tracsis’ Predictive Maintenance capability is being developed by a team of data-science and engineering experts in close collaboration with industry partners. Live and historical asset information is being used to train models, evaluate their performance and understand how reliably different patterns can be associated with confirmed failure modes.

Tracsis’ early-stage predictive platform has already been released to partners’ infrastructure teams to use in their day-to-day workflows, allowing live feedback to inform optimisation of the interface, usability, accuracy, and insight presentation.

These partners are providing the operational context that isolated raw data, including what happened to an asset, how a fault presented itself, how the infrastructure team responded, what operational and financial consequences followed, and whether an earlier prediction would have changed the maintenance decision. Additionally, they offer valuable feedback on the platform’s usability and intuitiveness.

These insights allow development to focus on enabling decision-making confidence, operationally useful outcomes and user experience rather than solely technical accuracy in isolation.

Project Overview

• Phase 1 (Complete) – Discovery and proof of concept: Assessing asset data, developing initial models and validating the technical approach through early design and testing. 

• Phase 2 (In Progress) – Limited deployment, testing and refinement: Testing the developing platform with selected industry partners using real-world asset data, feedback and operational use cases. 

• Phase 3 – Productisation and integration: Developing a fully functional predictive capability integrated into Tracsis' RCM, Centrix, and ready for wider operational deployment.

• Beyond – Expansion and optimisation: Extending predictive monitoring across additional asset types, improving model performance and continually refining the capability.

Conclusion: the next logical step for rail asset management

The rail industry has already made substantial progress in understanding infrastructure through remote monitoring, with extensive benefits for infrastructure teams and operational outcomes. The next step is to use the data yielded by that progress to deliver predictive insights that help teams make decisions more effectively.

Predictive monitoring has great potential, but for it to be trusted, safe and useful for operators, it needs to meet a high bar. Success needs to be measured by its impact on real maintenance decisions and engineering confidence, not just technical accuracy achieved in isolation from the wider context, so industry collaboration is an essential part of making it viable.

Scheduled and preventive maintenance will remain essential, as will the established expertise of infrastructure teams, and unexpected failures will still occur. However, carefully developed predictive technology will provide an extra level of decision-making confidence and the opportunity to reduce disruptive impact. With that confidence, engineering resources can be applied more effectively, the risks of trackside activity mitigated, and disruption for operators and passengers reduced.

For Tracsis, this is the logical evolution of more than two decades of rail asset monitoring experience, allowing us to build on our established monitoring platform, leading rail expertise, and access to high-quality asset data. By combining these advantages with specialised machine-learning and developing the technology directly with the rail industry and infrastructure teams, we are working hard to make this new generation of asset management a reality.