Real-time rockmass classification in TBM tunneling using a dynamically weighted LightGBM model optimized by Bayesian algorithm

The construction of ultra-long water diversion tunnels in hydraulic engineering often faces challenges such as high overburden depth, large cross-sections, and complicated geological conditions. Real-time perception of rock mass class is critical to ensure the safety and efficiency of Tunnel Boring Machine (TBM) excavation. An intelligent rock mass classification model based on Bayesian-optimized LightGBM was developed using 4984 TBM tunneling cycles from the main water conveyance tunnel of the Yinchuo–Jiliao Water Diversion Project (YCJL Project). Eight key parameters, including tool penetration index and gripper pressure, were selected as input features through Pearson correlation analysis and Random Forest feature importance ranking. A dynamic weight allocation strategy was developed to capture the temporal dependency of TBM tunneling data. Bayesian optimization automatically identifies the optimal phase ratio (20% rising stage + 80% stable stage data) by balancing the transient dynamics and steady-state characteristics. Combined with Monte Carlo cross-validation for hyperparameter tuning, the final model achieves a prediction accuracy of 92.88% on the test set, which can significantly outperform conventional algorithms such as Random Forest and XGBoost. The Shapley additive explanations analysis reveals distinct feature contributions to rock mass class prediction: gripper pressure and steel arch pump pressure dominate in Class II, while TPI shows heightened sensitivity in Class IV/V. The research outcomes provide an efficient solution for intelligent construction and disaster prevention in ultra-long hydraulic engineering tunnels.