Digital twin with uncertainty-informed deep learning for prompt quantitative risk assessment of deep excavation

Abstract

Deep excavation projects are inherently fraught with significant risks due to complex geological conditions, intricate construction processes, and uncertain surrounding environments. This study proposes a novel digital twin-driven framework that integrates the physical excavation process with a virtual digital model to enable prompt and closed-loop risk management. The framework combines building information modeling (BIM), Internet of Things (IoT), and uncertainty-informed deep learning models. First, multisource and heterogeneous data from the physical excavation process are captured via the BIM-IoT system to form a comprehensive database. The database is then input to the virtual digital model to facilitate synchronous risk prediction, influence factors identification, and risk control. A key innovation lies in the incorporation of an uncertainty-informed long-term time series model, which combines bidirectional long short-term memory-based point prediction and deep Gaussian process regression-based interval prediction. This hybrid model effectively addresses data and model uncertainties, bolstering forecasting precision. Finally, the proposed digital twin with uncertainty-informed deep learning model is applied to a practical deep excavation project, showcasing its capability to improve risk assessment and management throughout the entire construction process. These findings provide a solid foundation for enhancing risk management level with increased intelligence, promptness, and reliability in underground construction through the application of digital twin technology.