Operational efficiency & uptime

Predictive maintenance for rail infrastructure

Source New York City Subway (MTA) — public case. This is an industry example, not our project

How can AI predict rail infrastructure failures?

Rail and transit assets fail expensively and dangerously when maintenance is purely reactive or on a fixed calendar. A major transit operator used sensor data and pattern detection to flag defects early and prioritise which assets inspectors see first — moving from fixed schedules toward condition-based maintenance. The transferable capability is anomaly detection on commodity sensor data to catch wear before failure, with skilled inspectors making the call.

92%
of inspector-found track defects also flagged by the AI (2025 pilot)
335M
sensor readings collected in the initial pilot

The problem

Ageing rail infrastructure fails in ways annual visual inspection can’t catch — internal rail fatigue, bearing wear, switch degradation — and unplanned failures are dangerous and expensive.

The AI approach

Sensors (in the MTA pilot, Google Pixel smartphones mounted on subway trains) continuously capture audio and vibration data; machine-learning analysis detects the acoustic and vibration signatures of developing track defects before they become critical, generating prioritised maintenance.

Evidence it works

The MTA pilot mounted phones on A-train cars over several months to detect defects, shifting maintenance from reactive to proactive; MIT Sloan Management Review documented it as a successful proof of low-cost sensor-based predictive maintenance.

What “good” looks like

Earlier detection of developing faults, fewer service-affecting failures, and maintenance teams freed from manual inspection for planned interventions.

Feasibility & cost shape

The sensing can be remarkably cheap (commodity smartphones); the real cost is data engineering, model validation, and embedding alerts into maintenance workflows.

Our independent view

A useful reminder that the hardware is rarely the hard part — the workflow and data foundation are. We’d scope a single high-value asset class first and prove the false-positive rate before scaling.

Source & attribution

Based on publicly reported information about the New York City Subway (MTA) work.

This is an industry example included for illustration. It is not a Leia Intelligence project, and no client of ours is implied. Figures are as publicly reported by the original parties.

Sources: MIT Sloan Management Review ("A Maintenance Revolution: Reducing Downtime With AI Tools") · MTA / Google Public Sector ("Preventive Track Maintenance Pilot", press release, Feb 2025) · Mass Transit

This is the class of problem we help operators tackle — book a call.
Book a call