中文
English
In the 1992 film, “Sneakers,” a team of security specialists scans surveillance footage of Dr. Gunter Janek, a mathematician who’s invented a revolutionary codebreaker and hidden it in a black box—somewhere.
Whistler (David Strathairn), the group’s blind audio expert, listens patiently to the group, including Dan Aykroyd’s character, Mother, as they feverishly scour the video of Janek talking with a colleague for clues to the box’s location.
“Fellas,” Whistler interrupts them. “Janek’s little black box is on his desk between the pencil jar and the lamp.”
“Whistler, I hate to tell you this, but you’re blind,” Mother cracks.
“Don’t look—listen,” Whistler says, clinking a pair of tuning forks. (Spoiler: Thanks to Whistler, the group uses conversation clues in the video to find the box).
Thirty-three years later, Delaware-based startup Sonibel Instruments is doing just that—not to break into encrypted websites, but rather to sonically identify inconsistencies in semiautomatic MIG welds before they get to the inspection process.
The founding team of CEO Sophia Millar, Chief Technical Officer George Hallo, and Chief Product Officer Hooman Piroux originally got the idea to use sound to identify weld defects from a job shop owner they knew who reportedly would listen to the welds that new hires were doing and could tell just by the sound if they were messing up.
“He joked about how he could hear crappy welds from his office, and he’d have to go down the hall and tell the new guys that he just hired that they were doing something wrong,” Millar said with a laugh.
Standing apart from vision-based systems and reliance on cameras, the Sonibel system uses an aluminum-cased sensor that’s mounted on the welding torch itself. It focuses on how the droplets of molten metal are vibrating when they hit the weld puddle, Hallo said. The frequency variations in those vibrations indicate whether the weld is sound or is suffering from perforations or lack of proper fusion. Using Sonibel’s proprietary algorithm, which Hallo and Piroux have been working on for almost a year, the sensor’s software analyzes the feedback from the puddle, interprets it, and displays the result on a small monitor.
“We take small chunks of audio at a time and run it through our model to get the output,” Hallo said. “That output gives us a classification saying whether it’s a defect or it’s a good weld.”
Standing apart from vision-based systems and reliance on cameras, the Sonibel system uses an aluminum-cased sensor that’s mounted on the welding torch itself.
Hallo noted that the system currently indicates “good weld” or “defect” after the weld bead has been completed. However, if there is a defect, the system tells the welder at what point in the weld bead (say, between 40% and 55% of the total length) that the problem occurred. The next iteration of the system will tell users specifically whether the defect is porosity, lack of penetration, or some other type of error, Hallo said.
“Sometimes it’s very clear that this is a terrible, porosity-filled weld,” Millar said. “We can just label that right away. But other times, it’s subsurface, and that’s where our tool would help most.”
Hallo said that the strong work the company has done to build up its database of welds and inspection results—the base material for the system’s algorithm—is largely to thank for that.
“A lot of work has been going into getting that variability, covering all the settings, making sure … it doesn’t only catch the really bad defects, but when we have audio that just has a little bit of subsurface porosity, we’re still able to extract that and verify that this sound correlates to this. And with our processing, we’re able to find the certain characteristics that point to porosity.”
So far, Sonibel is so young, it only has a few of its units in the field actually being tested. Between lab and field tests and feedback from early test projects, the company is building a database from which its AI software draws conclusions about the accuracy of its users’ welds. Early on, the team found some development partners on a wait list that wanted to buy the product even before it was ready for market. So, in exchange for beta-testing the system, those partners offered scads of feedback and data—just what the team needed.
“Right now, a big part of what we’re doing is going back to that wait list and seeing who wants to convert, who wants to actually buy the unit,” Millar said. “We’re still pretty limited in the manufacturing, but right now we’re doing demos, and then we’ll do the pilots. Because it’s such a new technology, a lot of people don’t know that it works—they just want to test it. So, we’ll give it to them for two or three weeks, and then if they like it, they’ll buy it.”
