Erweiterte Modellierungstechniken für Gaussche Prozessmodelle

Advanced statistical modeling techniques like Gaussian Process models are widely used to model complex system behavior. Applications are for instance the modeling and prediction of engine characteristics like fuel consumption or emissions during calibration or combustion development in order to derive optimal calibration settings in a subsequent optimization task. The main challenge, though, is to make this powerful technique easily applicable for all engineers. At the same time, automotive development poses particular challenges requiring for specific adaptions of the generic modeling techniques. Taking both considerations into account FEV’s methodology and calibration engineers have developed a new DoE tool which will be shortly summarized with focus on the above mentioned aspects. As one major development step, a higher degree of configurability of the Gaussian Process models is sought for in order to achieve better model fits while maintaining the high user friendliness and easy applicability of today. Consequently, advanced Gaussian Process models will be evaluated with regard to the ability to model certain output quantities. In order to guarantee ease of usage an automatized model selection criterion is proposed. The applicability of these novel approaches and the advantages over standard Gaussian Process models will be demonstrated in two use cases. Artificial modeling data will be assessed as well as data that originates from optimizations of a gasoline direct injection engine towards the upcoming Euro 6c emission legislation with stringent particulate emission limits.