Observation Updating of Model Parameters for Consolidation Settlement Using Adaptive Surrogate Model

Data assimilation or observation updating is a method for improving prediction by using observation data to reduce the uncertainty. Particle filter (PF) is one of data assimilation methods based on Bayesian inference and Monte Carlo approach. In PF, posterior probability distributions (PDFs) are expressed by many sample realizations. PF has a problem known as degeneracy, where weights tend to concentrate into only a few particles, which causes poor computational performance. If many particles are used, the degeneracy can be avoided, however it requires high computation cost. Adaptive surrogate model approach for reliability analysis have attracted attentions for their low computation cost recently. We introduce a method to estimate posterior probability distribution directly by using surrogate model constructed by adaptive Gaussian Process Regression (GPR) with multiple random fields. The learning function is also very important in active learning, and has two roles in estimation of posterior PDF: a) finding the location of large values of posterior PDF, b) avoiding the location where posterior PDF is once calculated. The proposed method is demonstrated through a consolidation settlement problem. The settlement is predicted by using soil/water coupling FEM analysis. Model parameters of FEM are updated by synthesized observation data of settlements.

Keywords: Data assimilation, Gaussian process regression, Observation update, Particle filter.