Significant movement of in-situ retaining walls is usually assumed to begin with bulk excavation. However, an increasing number of case studies show that lowering the pore water pressures inside a diaphragm wall-type basement enclosure prior to bulk excavation can cause wall movements in the order of some centimeters. This paper describes the results of a laboratory-scale experiment carried out to explore mechanisms of in situ retaining wall movement associated with dewatering inside the enclosure prior to bulk excavation. Dewatering reduces the pore water pressures inside the enclosure more than outside, resulting in the wall moving as an unpropped cantilever supported only by the soil. Lateral effective stresses in the shallow soil behind the wall are reduced, while lateral effective stresses in front of the wall increase. Although the associated lateral movement was small in the laboratory experiment, the movement could be proportionately larger in the field with a less stiff soil and a potentially greater dewatered depth. The implementation of a staged dewatering system, coupled with the potential for phased excavation and propping strategies, can effectively mitigate dewatering-induced wall and soil movements. This approach allows for enhanced stiffness of the wall support system, which can be dynamically adjusted based on real-time displacement monitoring data when necessary.

Reliable predictions of time-dependent diaphragm wall deflections in deep excavations in soft soils are crucial for managing potential damage to the surrounding environment. Bayesian updating offers a rational method for refining these predictions by using monitoring data. The inconsistency in monitoring data necessitates an examination of the impact of using different datasets on Bayesian updating. This paper presents a Bayesian updating of time-dependent deflections of diaphragm walls in deep excavations in soft soils using different datasets. The soft soil creep model is utilized to simulate the time-dependent behavior of soil. Bi-directional longshort memory neural networks are employed as surrogate models. Different updating strategies with varying numbers of data in the datasets are adopted for Bayesian updating and illustrated with the Taipei National Enterprise Center project. The results show that incorporating more monitoring data in the datasets for Bayesian updating does not guarantee better predictions unless the consistency of the monitoring data used is ensured. Additionally, the Bayesian updating process more accurately predicts short-term deflections than long-term ones, likely due to the higher consistency in short-term construction processes. It is advisable to review the construction processes to ensure the consistency of the monitoring data before selecting the appropriate dataset.