Neural Network Soft Sensors for Gasoline Engine Exhaust Emission Estimation

Abstract

Worldwide research and development programs aim to reduce harmful emissions from transportation vehicles. Soft sensors have shown great potentials to reduce cost and improve onboard diagnosis for vehicle emission control. In this work, two sets of soft sensors are proposed to predict the emissions and exhaust heat flux of a gasoline engine. Extensive steady-state measurement points over the entire engine operating conditions are collected for model training and validation, and the locally linear model tree learning method is adopted. The CO, NOx, hydrocarbon, exhaust temperature, and exhaust heat flux are estimated by the soft sensors under steady-state conditions. Training of CO, exhaust temperature, and exhaust heat flux models has achieved high model accuracy over the entire engine map. Local models are developed for NOx and HC emissions to improve model performance at different engine operating speed/load conditions, especially in the low emission zone. Model validation has shown correlation coefficients ranging 0.983 ∼ 0.999