The application of dynamic neural networks to the estimation of feedgas vehicle emissions

We describe the application of dynamic neural networks to the estimation of moment-by-moment (here referred to as “instantaneous”) feedgas emissions. We base such estimates on engine variables available in normal operation to the powertrain processor. Training data were acquired from a single vehicle on a chassis dynamometer facility using standard driving schedules (as used for emissions certification tests). The trained networks were tested using driving trajectories both similar to those used in training as well as trajectories that are distinctly different. The method described allows us to estimate instantaneous levels of carbon monoxide (CO), total hydrocarbon (HC), and oxides of nitrogen (NO_x) in the feedgas (i.e., in the exhaust stream prior to the catalytic converter) of a particular vehicle for a wide variety of trajectories, including cold start, using information readily available to the vehicle´s powertrain control module. We discuss briefly the prospects for further generalization of this capability