1 week Ago By Stephen Nessen
In one of the oldest railroad stories, folk legend John Henry races against a machine hammering holes for the construction of a railroad tunnel, only to die from exhaustion after just barely beating the machine. In the modern version of this tale, Google’s artificial intelligence algorithm could go up against the experienced ear of an MTA track maintainer. “ I was a skeptic to say the least,” Rob Sarno, a third-generation transit worker and assistant chief track officer for system maintenance, told Gothamist.
But last September, Sarno agreed to work with Google to install six of its Pixel smartphones on an A train. They attached some of the devices on the car's axle. They put one phone on the train's bonnet, which older A train models have.
The phones were set up to record hours of sounds and vibrations. Then the MTA sent Google hours of training audio of what track defects sound like, so it could feed its AI algorithm to detect defects based on the sounds or vibrations. The MTA has well-trained inspectors that do this job by regularly walking all 865 miles of track including train yards, looking for problems.
Sarno argues that’s the best way to spot an issue. Inspectors can see, for example, that if there’s white dust near the tracks, the ballast that holds up the tracks is getting too much pressure from a train that’s bouncing too much, and the track would need to be repaired. Likewise, Sarno can hear when there’s too much jangling when a train goes over the tracks, which indicates there are loose joints in need of tightening.
The agency also relies on three geometry cars — special train cars that cost millions of dollars and can scan the tracks, analyze that data using special computers and find defects the human eye can’t see. But the MTA is also frequently criticized for buying expensive, custom-made equipment, rather than so-called “off-the-shelf” technology, and nothing is more off-the-shelf than a smartphone. Still, Sarno was skeptical that the phone would record any audio that was worthwhile.
“If you've ever been in a subway system in New York, everything is loud. So I assumed there was going to be one blanket noise of loudness,” he said. Turns out he was wrong.
The audio was high quality, and the artificial intelligence had no problem deciphering it. Google sent what it believed were all the track defects to inspectors. Sarno listened to the same recordings, then did the same.
Workers looked at the tracks to see if they were right. By the end of the experiment in December, Google was able to identify track problems with 92% accuracy. Sarno, who listened to the same recordings, was only 80% accurate.
The findings were first published in Wired . Still, he insists his job shouldn't be replaced by a smartphone. “This cannot replace a track inspector.
I cannot say that enough,” he said. “It's giving the track inspectors more tools to do their job, but it cannot replace them.” But Sarno thought it was successful enough that he wants the MTA to see what other similar technology is out there, and asked the agency to issue a formal request from other vendors to see if any other companies have a similar system.
“ I think it would be worthwhile expanding it out,” Sarno said..
