Dynamic and explainable deep learning-based risk prediction on adjacent building induced by deep excavation

Abstract

Since deep excavation will inevitably generate great disturbances in the surrounding environment, excessive vertical displacement of adjacent buildings is one of the crucial concerns in ensuring excavation safety. Therefore, this study proposes a hybrid deep learning model called Att-Bilstm under the integration of the bidirectional long and short-term memory neural network (Bi-LSTM) and the self-attention mechanism, aiming to dynamically and accurately predict the upcoming excavation-induced risk to adjacent buildings. To better improve the explanation of the Att-Bilstm model, the Shapley Additive explanations (SHAP) analysis is conducted to measure the contribution of variables in risk prediction, while the Latin hypercube sampling (LHS) approach is employed to assess risk evolution. The developed deep learning approach is verified in a case study about the excavation of Area A of the Jing'an Temple Station on Shanghai Metro Line 14. Results indicate that: (1) The data denoising technique fast discrete wavelet transform (FDWT) and the automatic hyperparameter tunning method Bayesian optimization are beneficial to establish a highly accurate Att-Bilstm predictor with the proper time sliding window, reaching the MAPE smaller than 0.062. (2) Variables identified as having a greater influence through SHAP analysis require increased attention for effective control of excavation-induced risks. The simulation-based approach provides early risk warnings, contributing to risk mitigation in the uncertain and dynamic underground environment. (3) Compared to three other popular deep learning models, Att-Bilstm exhibits greater robustness in handling different levels of incomplete monitoring data, achieving a MAPE lower than 0.070 even with 60% missing data. In short, this study contributes to the dynamic perception and understanding of potential excavation-induced risks to adjacent buildings, significantly enhancing the reliability of risk mitigation decisions.