Large-scale baseline model exploration from structural monitoring based on a novel information entropy-probability learning function

Abstract

In this paper, an active learning framework for structural baseline model exploration is proposed based on the Kriging method. The framework is built to solve the problem that when the traditional Kriging approach needs to calibrate more structural parameters, the performance of Kriging predictors hinges very much upon the number of samples obtained from the finite element analysis. A novel information entropy-probability learning function is derived based on Bayesian inference and information entropy. To give a full picture, the uncertainties of different responses should also be quantified during the updating process. Then the effects of uncertainties are considered in both the learning function and objective function constructed from the posterior probabilities. The proposed algorithm is first verified experimentally by a two-span continuous beam considering different types of responses. The framework is then further improved for updating the baseline model of a cable-stayed bridge using field data. The active learning approach, as compared to the ordinary Kriging method, can achieve good performance without undue computational cost and the need for imposing weights on different responses. The proposed baseline model exploration method can be extensively applied to bridge engineering because it facilitates the calibration of numerical models using field measurements and response simulation of extreme loading with significant improvement in computational efficiency and performance of Kriging predictors.