Saturday, October 03, 2015

VW and the dark side of AI

When I heard VW got caught cheating on emissions tests, I was not surprised.

What was surprising is that I could smell french fries every time a coworker drove by in her 'clean' diesel VW bug running on biodiesel (a mixture of diesel and filtered used fryer oil from restaurants.)

How could the emissions from her car be clean if the aromatic chemicals that give off the characteristic french fry smell were not combusted beyond recognition?  How could such incomplete combustion pass the strict California emissions tests?

How did VW game emissions testing?

This Computer World article gave a log of possibilities but no definitive answer.
Arvind Thiruvengadam, a research assistant professor at West Virginia University, was involved in a project last year that evaluated tailpipe emissions of diesel cars made by European manufacturers for the American market.
Thiruvengadam said he hasn't researched the software that allowed Volkswagen to cheat on the tests. But he did say "there were lots of ways an electronic control unit could be programmed to identify testing and change its fuel mapping toward low emission in those rare scenarios."

For example, modern cars can sense when a hood is open for dynamometer testing, "so a smart hood switch could double as a defeat device."

Or, another sensor could detect when a vehicle's traction control unit was disabled, which is required during emissions testing, and place the emission system into a different mode.

"The possibilities are almost endless," he told Autoblog. "I'm pretty sure that if you're one of the largest car manufacturers, you could do a lot more."
Some of these methods could produce false positives--assume they are being tested when they are actually being driven on the road.

For instance, if the motor is running but the car is not being actively steered, can you assume the car is on rollers for a test?  What if the car is instead being driven on a straight desert road?

In that case, switching the car to 'clean' mode would drastically reduce the fuel efficiency and the car could run out of fuel in a remote location.  If many VW owners report the same problem when driving straight, often remote, roads, the gig would be up.

My aha! moment came when one article said the method involved steering, engine use, AND pressure sensor data.

According to the US EPA violation letter to VW, sensors that could be used in a 'defeat device' need to be disclosed along with the reason why the device is needed for a non-defeat purpose.

I could imagine pressure sensors being useful for a fuel injection system or a rough altimeter, but not while the car was in motion.  Other methods would work better due to Bernoulli's Principle.  So why would VW put pressure sensors in their cars?

A moving car is shaped a little bit like a wedge or an airplane wing.  The air above the car moves slightly faster than the air that flows below the car; Bernoulli's principle states that pressure exerted by a fluid decreases as the fluid velocity increases.

If the motor is running for many minutes without active steering AND the pressure sensor on top of the car does not drop, then the car is likely on rollers in an emissions lab.

Artificial intelligence, AI, in your car determines when to apply the ABS brakes, traction control and the ratio of air to fuel.  It can also be used to cheat on emissions tests.

How can we find a simple algorithm like this--that can be written in just a few lines--among the 100 million lines of code in some new cars?

The Upshot in the NYT answered (or tried to answer) Little Hunting Creek's question of excess mortality due to VW emissions.


  1. Since we now know that air pollution kills people, we should be able to quantify how many people VWs program killed. I find this kind of active evil distressing-so many people would have known what was happening there.

    1. Not to mention the person years scientists have wasted trying to figure out the sources of the mystery NOx found via remote sensing.

      We knew how much fuel was being consumed and what the emissions should have been given the auto inventory. Having those #s be off by a factor of 20-40 was such a huge time suck.

      Scientists have so much work to do with so little budget. To have wasted it for so many years due to cheats is infuriating.