Development of a diesel engine simulation tool using artificial neural networks

He, Y. Development of a Diesel Engine Simulation Tool Using Artificial Neural Networks. University of Wisconsin-Madison, 2002.

A study was conducted to develop an accurate simulation tool with a small resource footprint for diesel engine design. A general and systematic methodology that integrates powertrain system modeling and CFD engine modeling has been formulated.

The model integration methodology employed Artificial Neural Networks (ANN) to represent engine in-cylinder physics by training the ANN to approximate CFD simulation results of the engine. The ANN approach was applied to model a turbocharged DI diesel engine over a wide range of operating conditions. The seven most important diesel engine control parameters were varied over their possible ranges: engine speed (RPM), engine load (Mf), start of injection (SOI), injection pressure (Pinj), mass in the first injection pulse of a split injection (M1), boost pressure (P bst), and EGR. The studied engine responses included cylinder pressure (Pcyl), cylinder temperature (Tcyl), cylinder wall heat transfer (HT), NOx emission (NOx), and soot emission (Soot).

An efficient data collection methodology using the Design of Experiments (DOE) techniques was developed to select the most characteristic engine operating conditions and hence the most representative data to train the ANN. In total 71 engine steady-state operating conditions were simulated with CFD, and five Multi-layer Perceptrons (MLP) were trained individually to approximate the five engine output parameters (Pcyl, Tcyl, HT, NOx, Soot) as a function of the seven engine control parameters (RPM, Mf, SOI, Pinj, M1, Pbst, EGR). The testing results showed that the five trained MLPs achieved satisfactory capabilities of predicting engine responses and hence representing the characteristics of the engine over a wide range of operating conditions. The ANN modeling accuracy was improved by incorporating prior knowledge into the ANN design and using a committee of networks instead of the best single network to make predictions.

The trained MLPs were then integrated together to develop the ANN cylinder model that demonstrated the capability of simulating engine performance and emissions over the standard heavy-duty Federal Test Procedure (FTP) transient cycle.