Data-driven predictions of tunneling parameters for large-diameter slurry shield—A Bayesian perspective

Data-driven prediction of shield tunneling parameters has attracted great attention recently in the field of underground engineering. Existing machine learning methods, including both shallow and deep learning approaches, often lack the capability for dynamic updating with newly assimilated records and fail to provide explicit uncertainty quantification. This paper proposes a Bayesian framework integrated with a multivariate probabilistic model for shield tunneling parameter prediction, enabling joint modeling of interdependent parameters and dynamic updating with new data, thereby overcoming the lack of parameter correlation modeling and uncertainty quantification in traditional methods. Real-time tunneling parameter databases are firstly established for the three ground soil types along the tunneling longitudinal direction based on the borehole data. Multivariate probability distribution models for 8 key tunneling parameters are then built using the Johnson distributions, which offers the prior distributions for later Bayesian inferences. Then, the posterior distributions of tunnelling parameters of the ring to be excavated are further updated when the records of real-time tunneling parameters are collected. The performance of the proposed method is examined on the Zhuhai Tunnel. The obtained results demonstrate that the proposed method can effectively assimilate real-time tunneling parameter records to reduce the predictive uncertainty of key parameters, significantly improves the accuracy and interpretability of shield tunneling parameter predictions.