Towards trustworthy excavation-induced risk warning for adjacent building: A Bayesian reasoning based probabilistic deep learning method

Foundation pit excavation for underground space development inevitably disrupts the surrounding soil, raising safety concerns for adjacent buildings. To address the need for an intelligent and trustworthy warning of the excavation-induced risk for adjacent buildings, this study develops a hybrid deep learning framework for probabilistic modeling (PM) with a long short-term memory (LSTM) neural network (termed as PM-LSTM). The proposed framework incorporates Bayesian reasoning and a bidirectional mechanism to enhance its predictive capabilities. The forward learning process enables the dynamic estimation of the probability that adjacent buildings will experience varying levels of risk over time, as new data is incorporated. Meanwhile, it can precisely calculate the first exceeding probability of the adjacent building entering an extremely high-risk level daily, facilitating early warning triggers. Besides, the reverse learning process leverages Bayesian reasoning to identify the most influential response parameters of the foundation pit, serving as key checkpoints for excavation monitoring. It further calculates the posterior probabilities and their intervals for each response parameter under the assumption of a specific risk state for adjacent structures. These insights enable the formulation of proactive risk mitigation measures. The proposed PM-LSTM framework is validated through a case study of the excavation project at Zone A of Jing’an Temple Station on Shanghai Metro Line 14. Comparative analyses further demonstrate the robustness of the framework, underscoring its potential as a reliable decision-making tool for risk analysis and management in complex and uncertain underground engineering projects.