Heuristics-Enhanced Model Fusion Considering Incomplete Data Using Kriging Models

Anton V. Beek,Mian Li,Chao Ren

DOI: https://doi.org/10.1115/1.4038596

2017-01-01

Journal of Mechanical Design

Abstract:Simulation models are widely used to describe processes that would otherwise be arduous to analyze. However, many of these models merely provide an estimated response of the real systems, as their input parameters are exposed to uncertainty, or partially excluded from the model due to the complexity, or lack of understanding of the problem's physics. Accordingly, the prediction accuracy can be improved by integrating physical observations into low fidelity models, a process known as model calibration or model fusion. Typical model fusion techniques are essentially concerned with how to allocate information-rich data points to improve the model accuracy. However, methods on subtracting more information from already available data points have been starving attention. Subsequently, in this paper we acknowledge the dependence between the prior estimation of input parameters and the actual input parameters. Accordingly, the proposed framework subtracts the information contained in this relation to update the estimated input parameters and utilizes it in a model updating scheme to accurately approximate the real system outputs that are affected by all real input parameters (RIPs) of the problem. The proposed approach can effectively use limited experimental samples while maintaining prediction accuracy. It basically tweaks model parameters to update the computer simulation model so that it can match a specific set of experimental results. The significance and applicability of the proposed method is illustrated through comparison with a conventional model calibration scheme using two engineering examples.