Some Practical Methods for Damage Assessment of Underground Structures using Machine Learning Techniques and Probabilistic Models.

Abstract This article reviews some practical methods of failure assessment for underground structures such as tunnels and deep excavations during construction stages in urban regions. Many factors, including random variables and parameters with statistically quantified values and laws of distribution, are tentatively considered to evaluate their effects on the failure of a specific objective (i.e., settlement of surface, the collapse risk of the diaphragm wall, etc.). A numerical model for a real sector (subsurface 2.9 km in length) of tunnel ‘Metro Line No1 Sai Gon‐Suoi Tien’ is created to estimate the reliability index of the tunnel sector and to predict possible risks for the structure system. By manipulating the input data (predictors) in the numerical model, data about the response (i.e., settlement of the existing buildings) could be collected that are sufficient for estimating the probability of failure, Pf, which is nearly 8 % for BaSon area, and particularly equals 23.8 % for Ben Thanh area; this would be compared to the probability Pf predicted by using some non‐parametric machine learning techniques such as multivariate adaptive regression spline (MARS). Besides, some probabilistic methods, such as the Taguchi method, are also reviewed for the failure of a deep excavation case study, from which the percentage of contribution of each predictor to the failure is quantified.