Uncertainty Analysis of Structural Dynamics by Using Two-Level Gaussian Processes
Uncertainty analysis is an important part of structural dynamic analysis in various applications. When a large complex structure is under consideration, component mode synthesis (CMS) is frequently used for reduced-order numerical analysis. But even so, in some situations the computational costs are still high for repeated running of a computer code which is required in uncertainty analysis. Gaussian processes offer an emulation approach to realization of fast sampling over a given parameter configuration space. However, both the low-fidelity data obtained by CMS and the corresponding sample obtained by Gaussian process emulation need to be assessed by comparing with high-fidelity data which can be obtained but are usually very expensive. When obvious bias exist in the low-fidelity data, two-level Gaussian processes are introduced for processing both the low- and high-fidelity data simultaneously to make more accurate predictions of quantities of interest. CMS can serve not only to provide low-fidelity data but also to locate problematic areas on complex structures. Comparisons of the results obtained by Monte Carlo sampling, which is performed using both a full finite element model and a CMS model, indicate that two-level Gaussian processes can be an efficient tool to emulate high-fidelity sampling with guaranteed accuracy.