Abstract
Deep excavation often induces soil deformation (e.g., ground settlement), which might lead to serious damage to adjacent buildings, resulting in significant economic losses and even casualties. Therefore, accurate prediction of ground deformation induced by deep excavation is crucial. Over the past several decades, Observation Method (OM) has been increasingly used in geotechnical engineering, particularly for deep excavation. Engineers use OM during construction by developing models for predicting data related to deep excavation and then comparing the predicted values with actual monitoring results to adjust construction activities. However, real geotechnical observations from a site are often spatially sparse (e.g., site investigation data) and spatiotemporally variable (e.g., settlement monitoring data). Therefore, the results from model prediction often differ from the values observed in the field. This requires a continuous modification of the design model and parameters, which is challenging and subjective.To address these issues, a new machine learning-aided analysis of deep excavation is proposed in this study to predict ground settlement and wall displacement using limited observation data from the field. In the proposed method, a novel machine learning method called Iterative convolution eXtreme Gradient Boosting (IC-XGBoost) is adopted to explicitly model complex stratigraphic variations from limited borehole data. The quantified stratigraphic uncertainty then propagates through a finite element method (FEM) model under a Monte Carlo Simulation (MCS) framework. The results demonstrate that, when stratigraphic uncertainty is neglected, the ground settlement might be significantly underestimated. Using the results from such a stochastic FEM analysis that accounts for stratigraphic uncertainty, a data-driven and physics-informed Bayesian Sparse Dictionary Learning (SDL) is further adopted to improve the prediction of ground settlement and displacement by integrating monitoring data sequentially obtained from the field during construction.
The proposed approach is demonstrated using a deep excavation case history from Guangzhou, China. The results show that machine learning of monitoring data obtained from previous construction stages can greatly improve prediction of ground settlement and displacement in the subsequent three to four construction stages. The proposed method enables geotechnical engineers to dynamically update the prediction of ground deformation induced by deep excavation, as construction stages proceed and monitoring data are sequentially collected. This further facilitates timely and necessary modifications of construction activities during deep excavation, without a need for manual modifications of input soil parameters to FEM model, as usually do in traditional OM methods.
| Date of Award | 25 Sept 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Yu WANG (Supervisor) |