Adaptive Sampling Methodology for Structural Identification Using Radial-Basis Functions

Abstract

The aim of model-based structural identification is to identify suitable models and values for model parameters that determine structure behavior through comparing measurements with predictions. Well-known methodologies, such as traditional implementations of Bayesian model updating, have been shown to be inaccurate in cases characterized by systematic uncertainties and unknown spatial correlations. Error-domain model falsification (EDMF) is another approach to structural identification. This approach is easy to understand for practicing engineers and can provide robust parameter identification without assumptions on spatial correlations. The performance of all approaches involving sampling is affected by the number of model evaluations that is generated based on prior knowledge of parameter-value distributions. This paper focuses on a new sampling technique, called radial-basis function sampling (RBFS), and its application to EDMF, to generate a set of candidate models that represents the behavior of the structure with a certain confidence level. Radial-basis function sampling provides a good exploration of the parameter space even with a limited number of samples, which results in reduced computation times. A full-scale bridge in Singapore has been tested and a new index of sampling quality is proposed to compare this approach with other sampling techniques such as Latin hypercube sampling (LHS) and Markov-chain Monte Carlo (MCMC). Finally, a cross-validation method is used to verify the robustness of the approach and the sensitivity of sampling on prediction reliability.