Abstract: Accurate classification of surrounding rock is essential for tunnel safety and optimal support design; Its mechanical properties can be collected consecutively using a three-boom jumbo that generates measurement-while-drilling (MWD) data. However, conventional data-driven models often suffer from a majority class bias, because of the inherent sample imbalances and complicated nonlinearities, and it fails to identify the classes of high-risk minority rock. This paper presents a LightGBM model that is integrated with Synthetic Minority Over-sampling Technique (SMOTE) and Grey Wolf Optimizer (GWO). SMOTE reconstructs a balanced data distribution via local feature interpolation, while GWO optimizes the LightGBM's hyperparameters adaptively, improving both the model’s accuracy and generalization in handling imbalanced datasets. Results show this new model achieves an overall accuracy of 94.6%, outperforming Random Forest (RF) and Gradient Boosting Decision Tree (GBDT) by 11.8% and 12.2% respectively. This approach overcomes data imbalance limitations effectively, and offers a reliable decision-making engine for accurate early warning, dynamic support of high-risk minority rock classes, and intelligent safety management in tunnel construction.

Key words: geological type identification, three-arm rock drilling rig, tunneling parameters, LightGBM ensemble learning, SMOTE algorithm