Efficient statistical parameter calibration for complex structural dynamics systems underconsideration of model uncertainty

点击翻译按钮获取中文摘要

Increasing the virtualization of the industrial product development process requires accurate mathematical models for the description of the dynamical behavior of structural systems. At the same time, ever-shorter development cycles as well as an increasing number of product recalls and the resulting economic damage pose a challenge for the industry to improve the expressiveness of mathematical models in the decision-making. The consideration of the data uncertainty and model uncertainty inherent to the models is therefore getting more and more important. By means of a statistical parameter calibration, the uncertainty of the model parameters can be reduced and quantified simultaneously in order to increase the predictive accuracy of the model. Methods for statistical parameter calibration presume computationally inexpensive models or use fast surrogate models of complex models that can be evaluated thousand fold. Novel multi-fidelity methods combine the numerous evaluations of an inexpensive low-fidelity model with a few evaluations of an expensive, more precise high-fidelity model. Thus, the computational effort for a statistical parameter calibration can be reduced while at the same time ensuring a high precision of the results. Consequently, also complex structural models can be calibrated efficiently. However, existing approaches neglect the model uncertainty, resulting in the calibration to be biased and the physical parameters to forfeit its meaning. Therefore, the goal of the project is the development of a multi-fidelity method for the efficient statistical parameter calibration of computationally expensive models under consideration of model uncertainty. The testing of the method on the demonstrator of the CRC 805, which has been designed with similar requirements as an airplane landing gear will enable the transferability on comparable structural systems. By means of the developed method, it will be possible to quantify the data uncertainty as well as model uncertainty for computationally intensive models.