A Methodology to Enhance Design and On-Board Application of Neural Network Models for Virtual Sensing of No _x Emissions in Automotive Diesel Engines

The paper describes suited methodologies for developing Recurrent Neural Networks (RNN) aimed at estimating NO_x emissions at the exhaust of automotive Diesel engines. The proposed methodologies particularly aim at meeting the conflicting needs of feasible on-board implementation of advanced virtual sensors, such as neural network, and satisfactory prediction accuracy. Suited identification procedures and experimental tests were developed to improve RNN precision and generalization in predicting engine NO_x emissions during transient operation. NO_x measurements were accomplished by a fast response analyzer on a production automotive Diesel engine at the test bench. Proper post-processing of available experiments was performed to provide the identification procedure with the most exhaustive information content. The comparison between experimental results and predicted NO_x values on several engine transients, exhibits high level of accuracy.