Bayesian graph attention network for segment uplift prediction in shield tunnelling

This study proposes a Bayesian Graph Attention Network (BGAT) framework for segment uplift prediction with uncertainty quantification in shield tunnelling. The proposed approach formulates the segmental lining as a mechanics-informed graph, where rings are represented as nodes and physical adjacency is encoded as edges to capture inter-ring deformation compatibility. Unlike conventional sequence models, the graph attention mechanism adaptively learns the condition-dependent contributions of neighbouring rings under varying geological and construction conditions, while Bayesian inference quantifies epistemic uncertainty through Monte Carlo sampling. The framework was validated using field monitoring data from 800 segmental rings in the Guangzhou intercity railway tunnel project, with a 70%/30% sequential split for training and testing. Comparative experiments with seven benchmark models demonstrate that GAT achieves superior performance with R² = 0.932, outperforming BiLSTM (R² = 0.868) by 7.4%. The BGAT model provides well-calibrated 95% confidence intervals that adapt to local prediction difficulty. The proposed framework offers a valuable tool for uncertainty-aware prediction and risk-based decision-making in intelligent shield tunnelling construction.