SINGAPORE – Public transport operator SMRT is stepping up its use of artificial intelligence to raise service standards, a move that may take years but could reduce train delays and keep operating costs in check once deployed at scale.
Through a new AI-powered platform called Jarvis, SMRT is tapping data from across its system to predict faults before they occur.
Using a simple chatbot-style interface, maintenance teams can quickly access insights and predictions on when equipment might fail, enabling them to act before breakdowns happen and help reduce train delays, improve reliability and make journeys safer for passengers.
Developed by SMRT’s technology arm, Strides Technologies, in collaboration with software and cloud computing company Oracle, Jarvis is now being piloted with about 50 users.
Announced during Oracle’s AI World Tour in Singapore on April 14, the multi-year initiative aims to apply AI to railway engineering while ensuring it meets the strict safety standards required for train operations.
The move will broaden and elevate SMRT’s use of AI to improve operations. The company already uses an AI system to identify potential train problems or overcrowding before they happen.
It also reflects how AI is spreading beyond sectors such as banking and healthcare into areas like public transport. However, its effectiveness still depends on having sufficient data and supporting infrastructure, and it is best used to support, rather than replace, human decision-making, at least for now.
A key goal for Jarvis is to enable engineers to move away from traditional time-based maintenance – where parts and equipment are checked at fixed intervals regardless of their condition – towards monitoring how the equipment is performing in real time.
This would enable maintenance to be carried out only when there are signs of wear or potential failure.
“If we can assess the condition and predict failures, we don’t need to carry out maintenance every three or six months,” said Mr Albert Soh, head of business operations and analytics at Strides. “We do it when it is required.”
For a start, the new system will focus on mechanical failures that follow predictable wear-and-tear patterns. One example is the point machine – a critical track component that switches trains between lines. While failures are rare, they can cause significant disruption when they occur, Mr Soh explained.
Using historical and sensor data, Jarvis is being trained to flag early signs of deterioration days in advance, allowing maintenance crews to fix issues during scheduled windows instead of reacting during breakdowns.
Another application involves platform screen doors across the North-South and East-West lines, where more than 2,000 doors are monitored. By analysing how long doors take to open and close, the system can detect subtle performance degradation and prioritise maintenance before faults arise.
This would help engineers deploy resources more precisely, while reducing inconvenience to passengers, Mr Soh said. “For commuters, the outcome is fewer disruptions and a smoother journey,” he added.
The shift is also expected to help SMRT slow the rise of its operating costs, particularly manpower, by allowing maintenance teams to prioritise work more efficiently.
“It is about enabling our people to do more, and to support a growing rail network without a corresponding increase in manpower.”
The Jarvis pilot is expected to run until the end of 2026, with broader deployment expected to take several years, as SMRT integrates decades-old systems, improves data quality and installs sensors where needed.
Mr Soh said training AI models can take between six months and a year, depending on the amount of data gathered and the complexity, as systems are built, tested and validated before deployment.
“AI can help us stay ahead of problems, rather than reacting to them,” he said, adding that with more than two million passenger journeys supported daily, even small improvements in reliability can have a significant impact on passengers’ daily lives.
Still, Jarvis will not be able to predict all problems, especially sudden electrical faults or unexpected incidents.
A lack of data remains one of the key constraints, with organisations still working to gather information from different systems to train more accurate models.
Mr Chris Chelliah, senior vice-president of technology at Oracle Japan and Asia-Pacific, said this reflects how larger and more critical uses of AI are being rolled out – not as a one-off move from pilot to full implementation, but as a system that evolves and learns over time, gradually taking on more complex functions.
In SMRT’s case, early versions of Jarvis were used to help engineers search through technical manuals and retrieve instructions more quickly, before progressing to analysing data, flagging potential faults and recommending actions in advance.
“Eventually, Jarvis will learn enough and evolve to augment decision-making,” Mr Chelliah said.
