Performing Global Uncertainty and Sensitivity Analysis from Given Data in Tunnel Construction

This paper develops a novel hybrid approach that integrates metamodeling, machine learning algorithms, and a variance decomposition technique to support global uncertainty and sensitivity (US) analysis under uncertainty. It consists of three main steps: (1) metamodel construction; (2) metamodel validation; and (3) global US analysis. A multi-input and multioutput metamodel, with least-squares support vector machine (LSSVM) and particle swarm optimization (PSO) algorithms incorporated, is built in order to simulate system behaviors of tunnel-induced building damage. Three indicators-mean absolute percentage error (MAPE), variance of absolute percentage error (VAPE), and mean square percentage error (MSPE)-are proposed to test the prediction performance of the metamodel. The extended Fourier amplitude sensitivity test (EFAST) is used to perform global US analysis on the basis of the well-trained metamodel. The novelty of the developed approach lies in its capability of learning from given data to identify relationships between model inputs and outputs to provide an access for conducting global US analysis. The collected data from the construction of the Wuhan Metro system (WMS) in China are used in a case study to demonstrate the effectiveness and applicability of the developed approach. Results indicate that the developed approach is capable of (1) predicting and assessing the magnitude of tunnel-induced building damage in terms of the cumulative distribution function (CDF) of model outputs, and (2) identifying the most significant and insignificant factors for possible dimension reduction to improve the understanding of the model behavior. This research contributes to (1) the body of knowledge by proposing a more appropriate research methodology that can cope with aleatory and epistemic uncertainty and support global US analysis based on given data, and (2) the state of practice by providing a data-driven metamodel technique to simulate system behaviors of tunnel-induced building damage with high reliability and reduce dependency on domain experts. (C) 2017 American Society of Civil Engineers.