New Approach to Ground Settlement Analysis During Box Jacking in Spatially Variable Soils

The three-dimensional random finite element method (3D RFEM) provides a robust tool for box jacking analysis considering spatial variability of soil parameters. However, it has a major drawback of being computationally very time-consuming. To address this common criticism, a novel surrogate model-based method that involves the data-driven random field and deep learning is proposed for the relationship between the ground settlement and random fields and the prediction of maximum ground settlement. The methodology starts from generating a set of random field samples using data-driven theory, which are then repeated finite element analyses to obtain the associated maximum settlement. Based on the mapping between features of input image data and 3D random fields, an innovative multi-channel fusion technique is developed to generate a multi-channel CNN (MC-CNN) model, which is trained to learn the relationship between the maximum settlement and 3D random fields and predict the maximum settlement. The effectiveness of the methodology is illustrated and validated using an underground cross passage project in Suzhou. The results show that MC-CNN model is effective in interpreting 3D random fields. In addition, through using the multi-channel fusion technique, the predictive capability of the MC-CNN model has been effectively boosted, especially compared with the single-channel CNN (SC-CNN) model, but the computational efficiency of the stochastic analysis has also been improved. Meanwhile, compared with the traditional regression model, MC-CNN is better at handling high-dimensional data.