Modeling and Optimization for Stationary Base Engine Calibration

Abstract

This thesis presents new approaches and results for modeling and optimization for stationary base engine calibration.At first, the requirements on the modeling are discussed, in order to determine the most suitable modeling technique for this topic in an extensive comparison.The Gaussian process modeling can be identified as the most promising approach.With the Gaussian process modeling, the highest precision with a low amount of measurements can be achieved, the modeling can be fully automated with the maximum marginal likelihood method, a dependable performance on complex problems can be obtained and an accurate prediction of the uncertainty of the model can be estimated.However, recent approaches in engine calibration do not consider outliers in the measurements and an automatic adaption to bad distributed data.Therefore, based on the results of the model comparison, the most promising modeling technique is enhanced in order that a new robust modeling framework for stationary base engine calibration can be obtained.An automatic transformation of the measurement data ensures that the modeling assumptions on the data distributions are met.Even if outliers are contained in the data set, a robust Gaussian process formulation guarantees that the modeling is asymptotically unbiased, meaning that the model is tending to the real engine behavior as the number of measurements tends towards infinity.Since state of the art model-based online optimizations for engine calibration do not use a fully probabilistic approach and can only handle a single objective function, a new, improved online optimization approach is introduced.As a Gaussian process modeling is used, additional information, such as an accurate prediction of the variance and the marginal likelihood probability density function of the model parameters, can be exploited for the online modeling, in order to obtain an increased performance at a lower amount of measurements compared to other approaches.With the new multi-objective online optimization, more objectives can be regarded and the Pareto optimal areas can be determined.All these new contributions enhance the performance for modeling and optimization, and therefore they are able to reduce time and costs on the test bench, improve the reliability of modeling and optimization results, assist the calibration engineers and increase the user acceptance of model-based techniques in engine calibration.Various theoretical examples and practical applications demonstrate the performance of these new approaches.A very special appreciation goes to my friends and my family for their understanding and support.Extraordinary thanks go to my parents, Bruno and Edith Berger, and to my sister Stefanie Berger for their love, care and support throughout all my years of education.My deepest thanks go to Carina Freutsmiedl.I know that I spent very much time on working in the last months.