Unsupervised learning and nonlinear identification for in-cylinder pressure prediction of diesel combustion rate shaping process

Abstract

Combustion Rate Shaping (CRS) offers the potential to control in-cylinder fuel concentration gradient and distribution with new fuel injection strategies, consisting of a higher number of injections and smaller individual injection quantities. Developing a physics-based model for such strategies involves significant development efforts and is also associated with high computational cost. This paper proposes a black-box framework for CRS process control based on artificial neural network and principal components analysis. To identify the nonlinear system behavior of diesel combustion, a multi-input/multi-output empirical model has been developed. The cylinder pressure trace is transformed into principal components coefficients space, and the neural network is used to predict the coefficients from operation parameters of diesel combustion. The in-cylinder pressure trace is then reconstructed by the predicted coefficients and pre-extracted principal components. The model has been evaluated with specific combustion features of CRS diesel engine experiments. The results show that the model successfully captured the combustion characteristics of CRS based diesel combustion processes with sufficient generalization ability.