ABSTRACT
Many excavations are made adjacent to existing foundations, especially in urban areas. In order to prevent the damage of the adjacent structures induced by the excavation, it is important to predict the retaining wall deformations caused by excavations. However, the uncertainty of subsurface conditions and the nonlinear interactions between multiple agents (e.g., soils, groundwaters, excavation support structures, and excavation activities) make the prediction of the response of retaining wall induced by excavation a rather difficult and complex task. This paper proposes a method to solve this problem by using machine learning techniques, where information is generated from the interaction processes between soils, structures, and excavation activities for predicting retaining wall deformations. First of all, an artificial neural network (ANN) model is proposed, and the soil properties, groundwaters, excavation support structures and excavation activities are considered. Then a mixed sample database of retaining wall deformations is generated based on high-fidelity numerical simulations and field measurements in Shanghai. Finally, the model is trained by the mixed database to recognize the patterns of the complex interactions between retaining wall deformations, excavation geometries, support structures, and soil properties. The results show that the accuracy of predicting the retaining wall deformations is less than 2 mm compared with high-fidelity numerical simulations. In the meanwhile, the mean accuracy of predicting the retaining wall deformations is 2.62 mm compared with the field monitoring data of an engineering case in Shanghai. The strategy of machine learning techniques with the model can predict the maximum retaining wall deformations induced by excavation with high accuracy.
