
Rob Sarno, an assistant chief track officer with the New York City Metropolitan Transportation Authority (MTA), has spent 14 years managing maintenance and emergency response. Recently, he played a pivotal role in training artificial intelligence systems to recognize damaged rail sounds.
In a pilot program launched in collaboration with Google Public Sector, artificial intelligence technology was utilized to address the persistent issue of subway delays. The initiative, known as TrackInspect, involved affixing Google’s Pixel smartphones to select subway cars to collect data on sounds and vibrations. This information was uploaded to Google’s Cloud for analysis, helping detect track defects before they became major problems.
“By identifying early rail defects, we save both time and money for crew members and riders,” stated Demetrius Crichlow, New York City Transit president.
This move aligns with global efforts to enhance transit systems using artificial intelligence. For instance, New Jersey Transit leveraged AI for crowd management, while Chicago’s CTA utilized AI for security enhancements. Beijing introduced facial recognition systems to expedite ticketing during rush hours.
Addressing MTA’s Persistent Delays
TrackInspect began as a proof-of-concept project initiated by Google Public Sector’s Rapid Innovation Team at no cost to the MTA. However, the future expansion of the project hinges on funding, as the MTA requires billions for ongoing projects.
Google has previously collaborated with transportation agencies, such as developing a chatbot for Chicago’s CTA and integrating Amtrak schedules with Google Maps. Yet, the MTA’s scale is unmatched, operating 472 subway stations and 237 local bus routes. In 2024 alone, there were over 1 billion subway trips.
Despite its extensive reach, the MTA grapples with service disruptions. Data from September to December revealed tens of thousands of delays each month. By employing artificial intelligence, TrackInspect aims to reduce these disruptions and enhance rider experience.
How TrackInspect Works
Between September 2024 and January 2025, six Google Pixel smartphones with standard plastic cases were mounted on four R46 subway cars. These devices collected 335 million sensor readings, 1 million GPS locations, and 1,200 hours of audio.
The smartphones inside the cars had their microphones disabled to prevent capturing conversations, focusing solely on vibrations. External microphones attached to the subway’s underside detected subtle track sounds. Inspectors then reviewed flagged areas for potential issues, training the artificial intelligence model in the process.
The A line was selected due to its combination of underground and above-ground routes, as well as areas with new construction. The line also experiences frequent disruptions, with thousands of delays recorded monthly.
Sarno’s role involved listening to audio clips and identifying defects such as loose joints or battered rails. His accuracy rate reached approximately 80%, aiding in refining the AI model. The TrackInspect system identified 92% of the defects found by MTA inspectors.
The Future of TrackInspect
The success of TrackInspect has sparked interest from other transit systems. While the pilot program has concluded, the MTA is actively seeking partnerships with technology firms to further enhance track improvement software using artificial intelligence.
Although preliminary data suggests a reduction in certain delay types, further analysis is needed to confirm the program’s impact. Nevertheless, the integration of artificial intelligence in transit maintenance holds promise for a more efficient and reliable subway system in New York City.