Within the next year or two, you may be able to hop into a ride-share car, look at your smartphone and tell the driver that the left front tire needs air. This technology can send a car’s diagnostics to any car that a person is in. They wouldn’t need to know the car’s history or connect it manually anyway. The information would be gathered from the car’s sounds and vibrations that are measured by the phone’s microphone and accelerometers.
MIT researchers have developed a smartphone app that combines the various diagnostic systems that the team developed. The app would save the average driver $125 a year and improve overall gas mileage by a few percentage points, according to Joshua Siegal, Ph.D. ’16, co-author of the study. For trucks, the savings could run to $600 a year, not counting the benefits of avoiding breakdowns that could result in lost income.
With today’s smartphones, Siegal says, “The sensitivity is so high, you can do a good job [of detecting the relevant signals] without needing any special connection."
For some diagnostics, mounting the phone to a dashboard holder would improve the level of accuracy. The accuracy of the results from the diagnostic systems they have developed is in the excess 90 percent. Tests for misfire detection have produced no false positives where a problem was incorrectly identified.
The basic idea is to provide diagnostic information that can warn the driver of upcoming issues or needed routine maintenance before these conditions lead to breakdowns or blowouts.
So how can a phone tell that a car filter is getting clogged? "We're listening to the car's breathing, and listening for when it starts to snore," Siegel says. "As it starts to get clogged, it makes a whistling noise as air is drawn in. Listening to it, you can't differentiate it from the other engine noise, but your phone can."
To develop and test the various diagnostic systems, including detecting engine misfires that signal a bad spark plug or the need for a tune-up, the team tested data from a variety of cars, including some that ran perfectly and others that had issues. In order to test the models, the researchers rented cars, created a condition they wanted to diagnose and then restored the car to normal.
"For our data, we've induced failures [after renting] a perfectly good vehicle" and then fixed it and "returned the car better than when we took it out. I've rented cars and given them new air filters, balanced their tires, and done oil change" before taking them back, recalls Siegel.
Some diagnostics require a complicated multistep process. In order to tell if a car’s tires are getting bald and need to be replaced or if they are overinflated and might risk a blowout, the researchers use a combination of data collection and analysis. First, the system uses the phone’s built-in GPS system to monitor the car’s actual speed. Then vibration data can be used to determine how fast the wheels are running. That can be used to derive the wheel’s diameter, which can be compared with the diameter that would be expected if the tire was new and properly inflated.
Many of the diagnostics are derived by using machine-learning processes to compare many recordings of sound and vibration from well-tuned cars with similar ones that have a specific problem. The machine learning systems can then extract subtle differences. For example, algorithms designed to detect wheel balance problems did a better job at detecting imbalances than expert drivers from a major car company.
A prototype of the smartphone incorporating all these diagnostics tools is being developed and should be ready for field testing in six months. A commercial version should be available within a year after that.
A paper on this research was published in the journal Engineering Applications of Artificial Intelligence.