Uncertainty propagation of frequency response functions using a multi-output Gaussian Process model. (June 2019)

• Record Type: Journal Article
• Authors: Lu, Jun; Zhan, Zhenfei; Apley, Daniel W.; Chen, Wei
• Is Part Of: Computers & structures. Volume 217(2019)
• Page Start: 1
• Page End: 17
• Publication Date: 2019-06
• Subjects: Uncertainty propagation -- Frequency response function -- Nonparametric approach -- Gaussian Process model -- Functional output; Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2019.03.009
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: A MOGP model is used to speedup uncertainty propagation of structural dynamics. Modal approach is employed to avoid the curse of dimensionality of FRFs output. A vector mapping relation is created between random variables and modal parameters. Enjoys computational advantages of predicting individual FRFs and their statistics. Abstract: Uncertainty propagation of frequency response functions (FRFs) under parameter variations is crucial for structural design and reliability analysis. However, obtaining sufficiently large samples of FRFs from a high-fidelity finite element (FE) model can easily become unaffordable. To reduce the computational cost, much interest is focused on metamodeling techniques, but how to efficiently create a metamodel over the whole frequency domain still remains challenging. In this work, a novel nonparametric approach based on multi-output Gaussian Process (MOGP) modeling to speedup uncertainty propagation is proposed. First, modal decomposition together with a low-dimensional representation method is used to address the curse of dimensionality of FRF output. Subsequently, a MOGP model is employed to provide a fast vector-output approximation of the random modal parameters, which are needed thereafter for reconstruction of the FRFs at any frequency level. To demonstrate the numerical accuracy and computational efficiency of the proposed method, both a discrete and continuous numerical example are investigated. It is shown that the proposed approach not only achieves accurate estimation of FRF variability, but also greatly improves computational efficiency.