NYC subway listens for broken rails with six Google Pixel phones

Between September and January, six Google Pixel smartphones rode four New York City subway cars on the A line — bolted inside plastic enclosures under the cars and in their interiors, accumulating accelerometer, magnetometer, gyroscope and (for exterior units) microphone data across the 32 miles between upper Manhattan and southern Queens. The experiment by the MTA and Google, called TrackInspect, asked whether cheap, mostly off-the-shelf technology could supplement track inspection, and Google Public Sector didn't charge the agency for it.

The scale of the alternative matters: human inspectors walk all 665 miles of subway track, and sensor-laden train geometry cars run only three times a year. Over four months, TrackInspect identified 92 percent of the defect locations human inspectors later pinpointed, per the MTA. One person did heavy labeling: NYCT assistant chief track officer Robert Sarno, a 14-year veteran nicknamed the track whisperer, tagged loose joints, loose bolts, rail and tie defects from audio snippets, topping 80 percent correlation with real defects.

In total the project combined 335 million sensor readings and 1,200 hours of audio with New York City Transit's defect database to train a group of about 200 individual prediction models. Because the models say what kind of trouble they hear, inspectors can arrive with the right tools — the stated goal being that they 'spend time fixing instead of identifying.' US regulators still require human inspection, and WSP consultant Brian Poston expects that to stay; New York is unique in using harmonics — audio plus vibration — rather than trackside cameras or fixed sensors.

Commodity Pixels bolted under subway cars hear track trouble: sensor feeds plus one veteran's ear labels train ~200 models that point inspectors at defects.

It turns the fleet itself into the sensor network: every scheduled trip becomes an inspection pass, at the marginal cost of a bracket and a phone.

The scarce resource — an expert ear — is used where it is irreplaceable (labeling), and machine learning propagates it across millions of readings.

Model output is actionable, not just an alarm: predicted defect types tell crews which tools to bring, collapsing the identify-then-fix loop.

A free initial experiment de-risked the idea before the MTA committed to a production pilot.

The sensor you need may already be in someone's pocket: pair cheap commodity hardware with scarce expert judgment — the expert labels, the models scale the expert's ear across the whole system.

The collaboration is expanding into a full pilot in which Google builds a production version and puts it in the hands of track inspectors themselves. MTA president Demetrius Crichlow frames the end state as a modernized system that automatically identifies and organizes track fixes; Google Public Sector's CTO says it hopes the project becomes a catalyst for cheaper, preventive railway-safety tooling across the industry.