Rapid Seismic Risk Assessment of Structures with Gaussian Process Regression

Abstract

Risk assessment of structures is an important step to predict damages and losses during natural hazards. The main idea is to fit fragility functions from existing earthquake records. Considerable research has been conducted in this area and various fitting approaches have been suggested. However, existing methods for deriving fragility functions require large training data sets and are designed for a single type of structures each time. This study aims to provide the loss estimation data using supervised learning techniques with a limited amount of observations. Observations are made after the strike of the earthquake based on the responses obtained from randomly selected damaged structures. These observations are used to train the model and prediction of the responses are made for the rest of the structures in the area. A case study is conducted based on the Fukushima 2011 earthquake. It is shown that Gaussian Process Regression (GPR), can provide an integrated model prediction for a group of structures with high accuracy and low uncertainties. This way, not only the predictions are made by a single model for different types of structures, but also the number of observations is very small compared to the existing methods.