A Construction-Phase Reliability Framework for Hard Rock TBM Penetration Rate Prediction Under Delayed UCS Information

Reliable construction-phase prediction of hard rock tunnel boring machine (TBM) rate of penetration (ROP) remains difficult because ground–machine interaction changes along the alignment and uniaxial compressive strength (UCS) is often incomplete or delayed at ring scale. This study proposes a construction-phase reliability framework that integrates sequence deep learning, inverse-distance-weighted UCS completion, chronological rolling evaluation, PassRate monitoring, and performance-triggered updating. The framework was developed from a granite-dominated TBM drive in northern Thailand and evaluated under a delayed-UCS information policy. In the Phase-2 forward deployment-style evaluation, the selected gated recurrent unit (GRU) model achieved a root mean square error (RMSE) of 0.1639 m/h, a mean absolute error (MAE) of 0.1186 m/h, and 62.63% within a symmetric ±10% accuracy band over 990 evaluated rings. Direct static application of representative theoretical and empirical models produced substantially lower within-band performance of 11.92–20.71%. One early reliability trigger occurred at Ring 3409, after which UCS updating, retraining, and redeployment restored the monitoring process without further intervention triggers. The results show that construction-phase TBM prediction should be managed as an auditable reliability workflow with explicit information boundaries, rather than as a single static accuracy score.