Metamodel based robustness design is commonly used to mitigate the consequences of variability without removing its underlying sources. However, uncertainties introduced by metamodels should be properly addressed before conducting design optimization. In addition, the measurable data uncertainties from physical tests and computer simulations are also unneglectable. Relevance Vector Regression (RVR) is a probability model based on the Bayesian learning framework. It shows potential in estimating the prediction uncertainty. This paper proposes an alternative RVR based robustness design procedure considering the design variables uncertainty, data uncertainty and metamodeling uncertainty. Based on the fundamental theory of RVR, simplified expressions of the response mean and variance is derived for robustness design accounting for three kinds of uncertainties. The formulation of RVR based robustness design is then built. Double loop Monte Carlo sampling is used to solve this optimization problem. An engineering example is used to demonstrate the proposed method and comparative studies are conducted between the proposed method and traditional robustness design.

This content is only available via PDF.
You do not currently have access to this content.