DEARBORN, Mich. Ford Motor Co. is applying a neural-network technology developed at NASA's Jet Propulsion Laboratory (JPL) to meet the more stringent automotive-emissions standards of the next millennium. The recurrent neural network receives raw sensor data from a vehicle's engine and learns to detect and identify, in real-time, malfunctioning components that contribute to pollution emissions.
"Current on-board diagnostics using conventional technologies are, overall, not satisfactory," said Ken Marko, a corporate technical specialist at Ford. "Massive engineering efforts are required to craft the algorithms, and the situation will only worsen as emission regulations become more demanding in the 21st century."
The problem today is being attacked in a costly, piecemeal fashion by adding sensors and digital signal processors to tackle specific areas. Marko said the neural network does a better job, with no additional costs to the vehicle, "and learns on its own rather than requiring a massive engineering effort."
Future emissions regulations will demand that malfunctioning engine components be detected and identified quickly so that service personnel can replace or repair them before they contribute significantly to pollution. But the hundreds of thousands of events taking place in engines every minute make detection a formidable task using conventional analytical techniques. "Even if the next century's cars have Pentiums built into their dashboard, it would add a lot of cost to bulletproof the processors for mission-critical tasks like engine control," said Lee Feldkamp, a senior staff technical specialist at Ford. "We need dedicated controllers."
The researchers claim multilayered recurrent neural networks can solve the problem. Such networks are structured with a feedforward multilayer perception plus a one-layer time-lagged recurrent neural network. The architecture is ideal for identifying malfunctions that vary over time a situation typical with engine control.
Typical automobile engine functions such as misfire detection and air/fuel mixture optimization are readily monitored and corrected with neural nets. The technology also lends itself to fuel-canister purge, wherein leftover fuel is dumped into the cylinders to prevent its leaking into the environment.
Ford contracted with the Jet Propulsion Lab to adapt the JPL recurrent neural-network technology to engine control. The resultant neural design, licensed to Ford from the lab, not only can address fixed emissions standards but also can progressively learn to solve a car's emission problems better as it logs more time on the road.
"Our design will drop right into Ford's existing ASICs so that it adds no cost to the vehicle as would a faster DSP with sensors plus it will be able to quickly switch between a variety of engine-control tasks," said Jet Propulsion lab researcher Raoul Tawel.
Engine-combustion events occur on a millisecond time scale at maximum engine speeds, which is relatively slow for multilayered recurrent neural networks but would consume the lion's share of the computational power of current DSPs and high-end microprocessors. Because of the relatively slow timing requirements, Jet Propulsion opted to use an inexpensive bit-serial implementation.
The main occupation of the neural ASIC core is a series of parallel multiply-and-accumulate (MAC) operations to simulate the actions of living neurons and their synaptic weights. Neurons gather numerous parallel inputs together after multiplying each by a separate synaptic weight. The neurons then fire an output pulse if the sum exceeds the non-linear threshold level, which is preset by a sigmoid transfer function.
JPL chose a parallel intralayer topology with a single-instruction-multiple-data (SIMD) architecture, allowing all neurons to multiply their inputs simultaneously by separate synaptic weights. A slower, bit-serial fixed-point computation and multiplexed interlayer techniques were chosen to conserve ASIC real estate.
To conserve computational resources further, the non-linear transfer function is handled by a ROM-based lookup table. The multiplexing requires that the inputs and outputs to the 16 physical neurons be stored in RAM, to be doled out to the neurons in bit-serial fashion.
To accommodate the multitude of required diagnostic operations for emission control including misfire detection, air-fuel mixture optimization and fuel-canister purge many different neural architectures are stored in a global synaptic ROM and then referenced by pointers at the command of the global controller. The temporary neuron state RAM is 64 x 16 bits, compared with the 2-k x 16 bits of the global synaptic weight ROM.
The prototype chip was implemented on 0.5-micron CMOS, resulting in a 20-MHz device with an 8-mm2 die that could perform all 16 MAC operations in about 1.6 microseconds. That prototype was sped up to 50 MHz before its was handed over for manufacturing by Ford's wholly owned Visteon subsidiary.