Application of Artificial Neural Networks to Aftertreatment Thermal Modeling 2012-01-1302

Accurate estimation of catalyst bed temperatures is very crucial for effective control and diagnostics of aftertreatment systems. The architecture of most aftertreatment systems contains temperature sensors for measuring the exhaust gas temperatures at the inlet and outlet of the aftertreatment systems. However, the temperature that correctly reflects the temperature of the chemical reactions taking place on the catalyst surface is the catalyst bed temperature. From the Arrhenius relationship which governs the chemical reaction kinetics occurring in different aftertreatment systems, the rate of chemical reaction is very sensitive to the reaction temperature. Considerable changes in tailpipe emissions can result from small changes in the reaction temperature and robust emissions control systems should be able to compensate for these changes in reaction temperature to achieve the desired tailpipe emissions.

This paper presents an artificial neural network based model for estimating the catalyst bed temperature in aftertreatment systems. The artificial neural network was used to model the functional relationship of the catalyst bed temperature at different axial locations as a function of the exhaust gas mass flow rate, catalyst inlet exhaust gas temperature, catalyst outlet exhaust gas temperature, ambient temperature, and rate of heat generation from chemical reactions. A physics based one-dimensional thermal model for flow through aftertreatment systems was also developed and used to motivate the structure of the neural network model. Temperature measurements from diesel oxidation catalysts of different sizes during engine steady state and throttle snap experiments were then used to generate training samples for calibrating both models. The physics based one-dimensional thermal model and artificial neural network based model were then evaluated with data from transient experiments. When compared with the reduced order thermal model, the results show that the artificial neural network was able to achieve better accuracies.