Deterministic And Probabilistic Methods For Displacement Back Analysis Of Deep Excavations In Soft Soil

In design of deep excavations in soft soil,reasonable selection of soil properties is fundamental for the prediction of diaphragm wall deflection,which is crucial for the design optimization and construction safety assurance.However,it is difficult to estimate the wall deflection due to the uncertainty of the soil properties and the measurement error.In this study,deterministic back analysis method?improved GA-BP neural network considering the sensitivity factor of parameters?and probabilistic back analysis method?Bayesian method?are proposed to back-analyze soil properties.The feasibilities of the two back analysis methods are verified on the basis of two typical deep excavations in soft soil,i.e.,Jiyuqiao excavation case and Formosa excavation case.The main research contents are as follows:?1?Numerical models for the two selected cases are established using FLAC^3D to analyze the variation of wall deflection and strut stress with the excavation process.The results show that the wall deflection is large in the middle while small in both ends along the depth.The strut stress is far smaller than the yield strength of steel,which can ensure the stability of the excavation.Numerical simulation results are consistent with field observations,which ensure that numerical models can be used for subsequent parameter back analysis.?2?The improved GA-BP neural network considering the sensitivity factor of parameters is proposed to conduct the deterministic parameter back analysis.The applicability of the improved GA-BP neural network on back analysis is discussed through establishing the mapping relations between the multi-layered soil properties and wall deflection,and conducting comparative analyses with BP neural network.The results show that sensitivity factors are introduced to modify the fitness function and evaluate the behavior of the neural network,which can improve the inversion precision of the high sensitivity factor.The precision of the improved GA-BP neural network is better than that of BP neural network.?3?Bayesian method is adopted to conduct the probabilistic parameter back analysis.The maximum observed wall deflection and the deflection at multiple points are used to update the parameters,respectively.The reliability analysis is conducted using the updated soil properties.Failure probability is calculated by Monte-Carlo method and numerical integration method based on the sample statistic data,respectively.The results show that the soil properties can be updated more significantly using the deflection at multiple points.The numerical integration method is able to generate better precision than the Monte-Carlo method when calculating small probability events.