A spatiotemporal deep learning method for excavation-induced wall deflections

• DOI: 10.1016/j.jrmge.2023.09.034
• Publication Date: 2024-01-29T07:43:19Z
• Authors: Yuanqin Tao; Shaoxiang Zeng; Honglei Sun; Yuanqiang Cai; Jinzhang Zhang; Xiaodong Pan
• Keywords: Robustness

Data-driven approaches such as neural networks are increasingly used for deep excavations due to the growing amount of available monitoring data in practical projects. However, most network models only use from a single point and neglect spatial relationships between multiple points. Besides, lack flexibility providing predictions days after activity. This study proposes sequence-to-sequence (seq2seq) two-dimensional (2D) convolutional long short-term memory (S2SCL2D) predicting spatiotemporal wall deflections induced by excavations. The model utilizes all points on entire extracts features combining 2D layers (LSTM) layers. S2SCL2D achieves long-term prediction through recursive seq2seq structure. excavation depth, which has significant impact deflections, is also considered using feature fusion method. An project Hangzhou, China, illustrate proposed model. results demonstrate that superior accuracy robustness than LSTM S2SCL1D (one-dimensional) models. demonstrates strong generalizability when applied an adjacent excavation. Based results, practitioners can plan allocate resources advance address potential engineering issues.