An investigation of semiphysical artificial neural networks for multi-fuel combustion phasing control of spark ignition engines

Abstract

The number of engine control actuators and potential fuel sources are constantly increasing to meet fuel economy targets and global energy demand. The increased engine control complexity resulting from new actuators and fuels motivates the use of model-based control methodologies over map-based empirical approaches. Purely physics-based control techniques have the potential to decrease calibration burdens but must be complex to represent nonlinear engine behavior with low computational requirements. Artificial neural networks are recognized as powerful tools for modeling systems which exhibit nonlinear relationships, but they lack physical significance. Combining these two techniques to produce semiphysical artificial neural network models which provide acceptable accuracy while minimizing the artificial neural network size, the calibration effort and the computational intensity is the focus of this research. To minimize the size of the neural network, sensitivity analyses are carried out on the critical inputs and the minimum number of required neurons. The most critical physical parameters are selected as follows: the laminar flame speed; the turbulence intensity; the total in-cylinder mass. The control algorithm derivation is described, and the process validated in real time using an engine dynamometer. The real-time experimental results demonstrate that the semiphysical artificial neural network approach can produce accurate ignition timing control for both gasoline and E85. Robustness of the semiphysical neural network approach is also discussed on the basis of the real-time experimental results.