Every day, millions of passengers tap through fare collection gates and validators to begin their daily commute. In London alone, Transport for London reported more than 3.5 billion journeys on their services in 2024/25.
Behind every seamless tap lies a complex machine working tirelessly, but constant use – and occasional acts of vandalism- inevitably lead to mechanical or electronic failure. For each unplanned breakdown, the ripple effect can be significant: gates go offline, queues build, boarding delays increase, travelers get frustrated, and the transit agency’s bottom line can take a hit.
Traditionally maintenance on these machines has been reactive: engineers review incident reports, travel to sites, and attempt on-the-spot fixes. If the right spare part isn’t on hand, a second visit is required. The result is longer downtime, higher operational costs, and preventable inconvenience for passengers.
What if, instead, we could foresee problems before they occur and send an engineer with exactly the spare part needed before it requires replacement? That’s the promise of predictive maintenance, powered by modern AI.
Working in collaboration with independent researchers from the Artificial Intelligence and Data Analytics (AIDA) Lab at Imperial College London, our team set out to apply large language models to fare gate logs and error reports to explore whether we could accurately anticipate issues before they occurred and dispatch engineers with the exact part they need – transforming the economics and reliability of public transport.
What is Predictive Maintenance?
At its heart, predictive maintenance is about replacing parts not because of a fixed calendar schedule, or after a breakdown, but because the data tells you a failure is imminent. Think of it as the difference between changing your car’s oil every three months “just in case” and replacing it when your car’s warning light tells you there is an issue.
In the public transit domain, a fare gate is an ideal candidate for this approach as it represents a hybrid of mechanics, electronics and software. Instead of conducting blanket part replacements, or waiting for a gate to jam then sending an engineer to identify the problem, recommend a part replacement, order the part and return to install it (2 total visits), the system could leverage logs from hardware and human reports to forecast and order parts required, allowing teams to intervene with the right parts in hand (1 total visit).
This predictive maintenance approach goes a step beyond preventative maintenance (which is schedule-based) and a shift away from reactive maintenance. This proactive model does more than improve reliability. It reduces the need for routine blanket part changes and costly emergency callouts, and it keeps passengers moving smoothly through stations. In short, predictive maintenance is smarter, cheaper, and more customer-friendly than the old “break-fix” approach.
Why Transit Agencies Stand to Gain
From a transit agency’s point of view, the benefits of predictive maintenance are compelling:
-
Operational Efficiency:
AI driven predictions allow maintenance teams to plan interventions during off-peak or low traffic periods and arrive fully prepared with all necessary equipment. Our research showed the potential to reduce on-site engineering visits by up to 70 percent. -
Cost Savings:
Fewer call outs and optimized spare-parts inventory means lower labor, inventory and storage costs, plus less revenue lost to gate downtime. According to the Wall Street Journal, “Unplanned downtime costs industrial manufacturers an estimated $50 billion annually. Equipment failure is the cause of 42 percent of this downtime.” A report from McKinsey & Company suggests reducing equipment failures through predictive maintenance can reduce maintenance costs by 18-25%. A recent report by the U.S. Department of Energy echoed these cost savings, estimating that predictive maintenance could generate 8-12 percent more cost savings than preventative maintenance, and up to 40% more than reactive maintenance. -
Reduced downtime and improved availability:
The logic is straightforward: predicting and preventing faults reduces the number of service interruptions. Some organizations report up to 50% fewer unplanned outages after moving to predictive approaches. -
Better passenger experience:
Less downtime means fewer gate outages, fewer station delays and fewer complaints from passengers. Over time, reliability bolsters public confidence in transit systems. The strategic value is not only the savings, but the reputational gains – especially in jurisdictions under pressure to improve service and ridership metrics.
How we used AI to Power Predictive Maintenance
The leap from concept to real-world impact comes from artificial intelligence – specifically large language models (LLMs) that can read and interpret text with near-human understanding. Specifically, this model allows the part required for fare gate maintenance to be predicted ahead of a site visit. The LLM our team used was affectionately nicknamed “PartLlama”.
Here’s how it works:
1. Ingesting maintenance data
Fare -gate incident logs contain a mix of structured error codes and free-text narratives. PartLlama ingests these descriptions and converts them into “tokens” that AI can analyze.
2. Learning to classify parts
Instead of predicting the next word (as a typical LLM does), PartLlama classifies each log entry into one of a dozen possible spare part categories. For example, a maintenance data input might read “Device crashed, cold start failed, requires PC.” This model outputs a precise part match like: “part needed: 9504-08019-1 QTY1 (AFC PM).
3. Efficient finetuning with Low-Rank Adaption
Because general-purpose language models are trained on vast and diverse datasets, adapting them to specific use cases, like fare gate maintenance logs, requires targeted fine-tuning. Training an entire 8-billion parameter model from scratch is expensive, so we used Low-Rank Adaptation (LoRA), which updates only 0.0524% of the model’s parameters. This approach preserves the model’s general language capabilities while keeping computing costs lower than typical LLM training budgets.
The Results: High Accuracy at a Low Cost
The results spoke volumes, showing a high level of accuracy at a low cost the model delivered:
- 83.5 percent Top 1 accuracy – meaning the model’s first guess of a required part was correct most of the time
- Over 95 percent Top 5 accuracy – meaning the correct answer almost always appeared in its top five suggestions of a part required for maintenance.
Due to the lightweight nature of the system, it is highly scalable and can run in real time on standard IT infrastructure, integrating seamlessly with existing maintenance flows.
The model also demonstrated robustness across a variety of inputs. Whether the maintenance logs were written by staff or generated by automated sensors, PartLlama was able to deliver consistent accuracy, ensuring reliable predictions could be achieved in the diverse and sometimes messy data environments typical of public transport.
This study showed our AI backbone could move predictive maintenance from theory to practice. Instead of wading through thousands of incident reports, engineers could receive an actionable shortlist of likely fixes before they ever left the depot.
From Proof of Concept to Everyday Operations
This research is only the beginning. Our current model was trained on only a year of maintenance logs and focused on the 26 most frequently requested parts. Next iterations of the model would seek to add:
- Real-time integration: Building a deployment pipeline that ingests live maintenance logs and instantly returns ranked predictions.
- Broader coverage: Extending predictions to hundreds of rarely used parts, supported by continuous retraining as new data arrives.
- Explainability features: Highlighting the words or error codes that drive each prediction so engineers can validate AI recommendations with confidence.
The long-term potential of this research goes beyond predicting parts needed for service from maintenance logs - it includes anticipating external factors that may impact gate performance and contribute to failures.
A Smarter Future for Transit Maintenance
Public transport is the backbone of modern cities, and fare collection equipment are the silent workhorses that keep people moving. By pairing the deep operational knowledge of transit agencies, with the predictive power of AI, we can move from reactive maintenance to a proactive, data driven model.
The benefits are clear: fewer disruptions, lower costs and a better customer experience. For Cubic and our agency partners, this approach is more than a proof of concept – it’s a blueprint for the future of transit maintenance.
As this technology scales, passengers may never notice that gates don’t break down. But behind the scenes, AI will be working tirelessly alongside field technicians to keep their journeys smooth, reliable and stress-free.
Download the full white paper from the the Artificial Intelligence and Data Analytics (AIDA) Lab delivered via Imperial Consultants, for a deeper insight into the application of AI in predictive maintenance.