Hitting the Road
If handling unforeseen situations is important for a copy machine, it will be absolutely crucial for machines like the processor-heavy automobiles of the near future (see “The Networked Car,” TR September 2002). According to MIT’s Williams, officials from Toyota USA recently visited the Artificial Intelligence Laboratory looking for ways to reduce to zero the failure rate of the more than 30 computer processors inside each Toyota vehicle. “At that level,” says Williams, “control systems need to have the reasoning and smarts of a robot.” And although engineers are still decades away from putting an immobot into the driver’s seat-a concept that car buyers may never accept anyway-monitoring a car’s internal functions is a very different matter.
Drivers can’t spare the attention and don’t have the skills to sense problems with fuel pressure, emissions levels, antilock brakes, and any of the other complex systems that influence a car’s performance. Modern passenger vehicles contain dozens of processors, so-called electronic control units, that monitor and control these functions with conventional software. But researchers in Europe are already developing onboard diagnostic immobots that run on the same control units and respond to internal automotive failures as soon as (or even before) they occur.
Traditional heuristic programs fall short in such situations because they “fail to capture the first-principles knowledge human experts use” to diagnose problems and plan repairs, says Peter Struss, a computer scientist at the University of Technology at Munich. By “first-principles knowledge,” Struss means understanding of the structure, behavior, and interactions of valves, pipes, sensors, and other auto parts. At Occ’m, the Munich software company he cofounded, Struss has spent the last several years building and testing software models of auto parts in collaboration with companies such as BMW, Fiat, DaimlerChrysler, Peugeot, Citron, and Renault. He’s finding that his models identify problems that sometimes elude even expert mechanics.
Say the air conditioning in your vehicle won’t work. An onboard immobot might quickly deduce that the problem is a malfunctioning fuel-level sensor in the gas tank. “What’s the interdependency?” Struss asks. In some cars, he explains, “the AC control system has to ask the engine control unit whether, as a consumer, it’s allowed to come on. The engine management system will check whether there’s enough fuel. And if not, it will deny the request.”
A model-based diagnostic system knows such details in advance. Therefore, if you’ve got a full tank and the air conditioning still won’t work, the model-based diagnostic makes an educated guess at the cause. “Workshop people can read out the diagnostics from control units and see six or seven trouble codes, but it may not be obvious what the ultimate cause is,” says Struss. “They cannot see all the interactions.” But an immobot can.
In two prototype self-diagnosing vehicles Struss helped build for a project of the European Commission, the immobot software ran on laptops inside the passenger cabins. However, Struss predicts that as early as 2004, similar model-based programs will be built into new cars’ onboard diagnostic modules. Meanwhile, he is developing programming tools to help automotive designers build cars with diagnostic systems in mind, rather than (as is the case today) adding them as an afterthought. This means, for example, making sure sensors are always placed where they’ll return the data an immobot needs to identify-and someday soon, correct-component failures.
The focus of Struss’s work and related studies in Europe-including a project at France’s Centre National de la Recherche Scientifique to build a model-based system that will help garage mechanics isolate faults in electronic circuits-indicates that when immobot software finally makes its mass-market debut, it will almost certainly be in the automotive realm, says computer scientist Louise Trav-Massuys, director of the French project. “The major problem in automobiles,” she says, “is the increasing complexity of electrical systems, which means that companies are looking for intelligent tools to help them analyze their products.” The key economic and safety advantage of model-based diagnosis, she argues, is that “faults and symptoms do not need to be anticipated.” In other words, software designers need only spend their time learning how a car’s systems should act, not how they might go wrong.