Radical basis function neural network-based NOx soft sensor technique

With the advent of advanced diesel after-treatment technologies, select catalyst reduction (SCR) becomes dominant technology for the new emission legislation in china. Because of sophisticated NOx sensors are becoming a critical cost challenge to OEMs, open loop SCR is mainly used at present, and which lead to be difficultly adaptive to control emission reduction and calibration workload was heavy. Lean combustion technique for gasoline was applied, which lead to NOx gas obvious increase and needed NOx sensor to detect NOx concentration. This paper describes an approach for replacing the engine out NOx sensor with an artificial neural network (ANN) based NOx perception. A multi-layer perception network was trained to estimate NOx concentration from engine speed, load, exhaust temperature, and oxidation factor information. This supervised learning was conducted with measured engine data. The network was validated against measured data that was excluded from the training data set. The paper details application of this technique to a heavy duty diesel engine. Results show good agreement between predictions and measured data under the steady state conditions studied. The methodology will also be evaluated for use under transient operation, and over the operating lifetime of the associated application.