A two-dimensional spatial-temporal decision-support framework for pipe jacking in weathered rock

Jacking force prediction for pipe jacking in highly weathered rock is challenging because closed-form mechanistic models require reliable strength parameters that may be unavailable when intact cores cannot be recovered. Using two full-scale microtunnelling drives in weathered phyllite (120 m) and sandstone (140 m), this study develops an interpretable deep learning framework to relate operational parameters to drive-specific jacking force variability. Before developing the attention-based interpretation framework, GRU, LSTM and Conv1D baseline models were first compared using separate random grid search hyperparameter optimisation, with validation data derived only from the training set and the original testing data retained as unseen records. The selected GRU model was then integrated with Bahdanau attention for one-step-ahead prediction. To prevent temporal information leakage, the section-wise train-test split was completed before training-only up-sampling and sliding-window reconstruction, and no window was allowed to cross the train-test boundary. The GRU model achieved the best overall performance, with R 2 of 0.82 and RMSE of 7.93 ton for phyllite, and R 2 of 0.97 and RMSE of 12.22 ton for sandstone. Spatial attention was used to identify operational parameters with higher predictive contribution along the drive, while temporal attention identified when each parameter became predictively influential. These outputs were combined through a post-hoc dot-product mapping to produce a spatial–temporal influence score. Cross-drive transfer learning showed poor direct transfer in both directions, although predicted-drive fine-tuning improved performance, indicating limited direct generalisability but improved adaptability after fine-tuning. As the two drives also differ in burial depth, pipe diameter and machine configuration, the attention results are interpreted as drive-specific model-attributed patterns rather than lithology-only effects. The phyllite model relied more strongly on face support and thrust response, whereas the sandstone model showed greater predictive contribution from drive length, elapsed time and jacking speed. The resulting spatial–temporal influence map provides a practical framework for interpreting operational variability and stage-dependent predictive sensitivity during pipe jacking under changing geological and project-specific conditions.