Promotor: Prof.dr. T.H.W. Bäck
|Links||Thesis in Leiden Repository|
Over the last 20 years large efforts have been made in developing and optimising modelling techniques for DoE usage in engine calibration. A prerequisite for optimally applying DoE test designs is the detailed knowledge of the engine’s operating boundaries enclosing the ‘design space’. Known boundaries can be implemented in the DoE test plan such that no test points are planned in a region where the engine cannot or should not be operated. Four mathematical approaches have been analysed and compared on the problem statement ‘Which design space description method is most suitable for combustion engine calibration test applications?’: - Convex hull method (CH) - Prediction error variance (PEV) - Support vector machine (SVM) - Support vector machine with a leave-one-out optimisation (SVM-LOO) It can be concluded that the two SVM-based methods are the most suitable methods to describe a boundary. The quality of the assessment will be the primary selection criterion and therefore the SVM-LOO shows the most potential. Since the initial application of a design space description method will be offline, a longer boundary training time and new test point allocation time, e.g., required for the SVM and SVM-LOO methods, are acceptable for the current calibration problems.