Due to the ever increasingly stringent emission norms for passenger vehicles, the efficiency and performance of the Spark Ignition (SI) engines has been under the focus of the engine manufacturers. The quest of efficiency and performance has led to the development of increasingly complex powertrains and control strategies for these powertrains. The development process for layout and optimization, requires novel methods which feature a smooth transition between real and virtual prototypes. Furthermore, to reduce the development time and cost, having an engine simulator with low computational effort and good accuracy, which predicts the engine behaviour in the entire operating range, plays an important role.
The thesis proposes an Artificial Intelligence based engine simulator for a Gasoline Direct Injection (GDI) Turbo Charged (TC) engine. The simulator is a combination of certain quantities modelled by analytical models (previously developed by colleagues) and other by Artificial Neural Networks (ANN).
The activity began by analyzing the data acquired from the spark sweep tests (steady-state conditions) performed at the engine test cell at the University of Bologna. Consequently, a shallow neural network model was set up on Matlab for modelling combustion phase i.e., MFB50 and knock intensity i.e., MAPO98. Training of the network was done with a sweep test, where the different numbers of neurons were simulated using different training algorithms and activation functions available in Matlab. The output of the simulation (i.e., MFB50 and MAPO98) were then compared to the experimental values. Engine parameters such as load, speed, spark advance, lambda, and indexes such as MFB50, Pmax, IMEP, MAPO etc. have been normalized and the procedure adopted to normalize the variable has been mentioned in the relevant sections. The results were compared using Root Mean Square Error (RMSE) and R-square indexes, and the combination of the number of neurons, training algorithm, and activation function, which gave lowest RMSE and R-square, were selected.
The ANN models were then coupled to the already developed analytical models, which together form the complete engine simulator. The performance of this simulator was then tested by comparing the simulated results and the experimental data, both under steady state and transient conditions.