Neuro-fuzzy model tree approach to virtual sensing of transient diesel soot and NOx emissions

Abstract

Diesel engine combustion and emission formation are highly nonlinear and thus create a challenge related to engine diagnostics and engine control with emission feedback. This article describes the development of neuro-fuzzy models for prediction of transient NOX and soot emission from a diesel engine. The modeling techniques are motivated by the idea of divide and conquer the input–output space. The complex problem is divided into multiple simpler subproblems, which are then identified using simpler class of models. This article explores two different choices of local models, specifically polynomial and neural networks. The modeling technique is augmented with input relevance algorithm to select the most relevant input regressors. Two algorithms, namely, orthogonal least square and automatic relevance determination, are introduced. The models are data driven, and an advanced experimental setup incorporating a medium duty diesel engine and fast emission analyzers for soot and NOX is used to generate training data. The choice of local models and input relevance algorithm is validated with instantaneous emission recorded during transient schedules different from those used in development. High prediction accuracy, both qualitatively and quantitatively, is demonstrated with low computational cost.