MIT researchers are working on an app that uses a smartphone’s microphone and accelerometers to diagnose impending maintenance problems.
As cars get smarter, more and more of them are going to give their owners preventative maintenance alerts. It’s one of the benefits to consumers regularly touted by advocates of the connected car, and even some older cars can get in on the action via aftermarket units that connect to a car’s onboard diagnostics port.
However, that last one might not be necessary if a technique being developed by some researchers at MIT pans out. Rather than plugging a diagnostic dongle into a car’s controller area network—with the attendant hacking risk—Joshua Siegel and his colleagues reckon a smartphone’s microphone and accelerometers could be sufficient.
Some of his research has just been published in Engineering Applications of Artificial Intelligence; specifically a paper that shows that audio data collected by a smartphone alone can diagnose an air filter that needs to be changed.
The idea behind it is quite simple. A dirty or occluded filter—blocked by leaves, for example—will let a different flow of air through it than one that’s working as designed. And that difference will result in different auditory and vibratory signals. (This is important because a dirtier or occluded filter won’t send the optimal amount of cold fresh air to the engine, which means worse fuel economy and increased wear.)
Siegel and his colleagues tested the idea using a Mazda 2 and Honda Civic, recording engine noises with a stand microphone as well as an iPhone 6. In addition to getting audio samples from the engine bay under normal conditions, they also tested the cars after covering the air filters with about 2mm thickness of carbon filter material to simulate “uniform particulate buildup.”
The team also sampled the cars’ sound with a 10cm square piece of paper on top of a (clean) filter to simulate the presence of a large leaf or other blockage. Once armed with these audio samples as training sets, they created an algorithm that could learn to discriminate between a filter in optimal condition versus one that wasn’t working properly.
OK, it’s still a bit of a leap to go from waving a phone over an engine bay to listen to an air filter to an app that could keep a constant ear on your car’s health to warn of impending trouble, but it’s not that far-fetched. Siegel’s paper cites previous work that has shown neural networks can use sound to identify “air intake manifold leaks, ECT and camshaft sensor failures, and cylinders with accuracies exceeding 95 percent accuracy.”
Other work has also shown algorithms that are better at detecting an out-of-balance wheel, and Siegel intends to commercialize the idea through a company he founded called Data Driven. He plans to begin testing an app that integrates a number of those functions in the next six months, and the researcher suggests a commercial release might be ready within 18 months.
As with Nexar’s phone-based vehicle-to-vehicle (V2V) solution and some of the driver behavior monitoring being done by Zendrive, it’s another example of trying to get the good parts of the connected car without exposing drivers to an unnecessary hacking risk.