Comparative Analysis of Performance of Neural Estimators for Diagnostics in Engine Emission System

This article describes the results of a comparative performance analysis on the use of neural estimators to accurately estimate the Differential Pressure (DP) signal from diesel engine systems equipped with a Diesel Particulate Filter (DPF) aftertreatment system. For most systems, there are known and modeled relationships between system inputs and outputs; however, in the case of nonlinear, time-varying systems a detailed modeling of the system might not be readily available. Therefore, Artificial Neural Networks (ANNs) have been used for developing critical relationship between system inputs (engine and aftertreatment parameters) and system output (DP signal). Both batch (offline) and online learning ANN estimators have been proposed. A control-oriented engine out DPF-DP model is desirable for on-board applications as a virtual DPF-DP sensor which could be used in parallel as an alternate analytical redundancy-based sensor. Furthermore, in order to limit the online computational effort, a limited set of inputs has been selected after a detailed correlation analysis. The experimental validation demonstrate that the online learning estimators provide better overall results in terms of accuracy and overall robustness to time varying and non-linear conditions.