A hybrid data-driven model for geotechnical reliability analysis

Tunnel boring machines are widely used to construct underground rail networks in urban areas. However, ground settlement due to complex geological conditions is an ever-present reality requiring continuous monitoring and management of risks. This paper addresses the following research question: How can we predict tunnelinduced ground settlement with engineering parameters, improve its predictive ability, and quantify its risks under uncertain parameters in complex geological conditions? To this end, we develop a hybrid data-driven model that considers prior domain knowledge to effectively and accurately quantify risk under uncertain parameters during a tunnel's excavation process. Our model comprises: (1) a deep neural network (DNN) to construct a ground settlement prediction model; (2) the incorporation of physical knowledge into the DNN-based prediction model; and (3) a Markov-chain-based importance sampling to analyze settlement reliability. We use the San-yang Road tunnel project in Wuhan, China, to evaluate the effectiveness and feasibility of our proposed approach. The results demonstrate that our hybrid data-driven model can accurately predict tunnel-induced ground settlement and quantify failure probability for geotechnical reliability under uncertain parameters during a tunnel's excavation process.

Automated summary

This study proposes a hybrid data-driven model that effectively and accurately quantifies the risks of tunnel-induced ground settlement under uncertain parameters in complex geological conditions, by considering prior domain knowledge. The model incorporates a deep neural network for ground settlement prediction, physical knowledge, and a Markov-chain-based importance sampling for settlement reliability analysis. Evaluation using the San-yang Road tunnel project in Wuhan, China, demonstrates the model's ability to accurately predict tunnel-induced ground settlement and quantify failure probability for geotechnical reliability under uncertain parameters.