Probabilistic prediction of vertical deflection for bridges using Gaussian process regression

Abstract

The prediction of the long-term deflection of concrete bridges is not easy because it is induced by several complex physical phenomena such as creep and shrinkage. Several physics-based equations have been suggested in various standards. However, the predictions based on these equations can be different from actual measurements owing to various uncertainty sources including material properties, traffic loads, and temperature. In this study, a probabilistic method is proposed to provide a reliable probabilistic prediction on the long-term vertical deflection of bridges. The proposed method adopts Finite Element (FE) analysis model based on a conventional physics-based equation as a basis function and introduces a Gaussian process to construct a probabilistic prediction model. Based on the actual measurements of bridge vertical deflection, the parameters of the Gaussian process model are determined through optimization to maximize the probability of observing the given measurement data. The constructed Gaussian process model can provide 95% and 99% prediction intervals as well as the predictive mean on bridge vertical deflection. The proposed method is applied to an actual bridge in the Republic of Korea, and the prediction results show good agreement with the actual measurements.