Real-time surface settlement prediction in shield tunneling using a time-series framework and interpretable machine learning

● A novel framework is proposed for real-time prediction of the entire surface settlement process. ● The time-series data structure captures the cumulative influence of multi-ring excavation. ● The optimized Bi-LSTM model excels at predicting the surface settlement process and MSS. ● SHAP analysis validates the model and links large settlements to construction anomalies. Current data-driven methods for surface settlement prediction in shield tunneling typically focus on static maximum values, often neglecting the cumulative influence of the continuous excavation process and the dynamic ground response. To address this, this study presents a time-series framework utilizing a Bayesian-optimized Bi-directional Long Short-Term Memory (Bi-LSTM) network to predict the entire settlement process. By modeling the input as a multi-ring sequence, the framework is designed to capture the temporal dependencies inherent in shield tunneling. The results indicate that the Bi-LSTM model delivers robust accuracy (MAE = 1.65 mm, R = 0.94), showing improvements over the baseline models in this study. Additionally, SHapley Additive exPlanations analysis suggests that the model captures physically meaningful mechanisms, identifying shield position and abnormal stoppages as key influencing factors. This approach shows potential for assisting in dynamic risk assessment and offers reference for operational parameter optimization in shield tunneling.