Interpretable machine learning for automatic rock mass classification in TBM tunneling: a case study of Beishan URL

Abstract Accurate rock mass classification is essential for safe tunnel boring machine (TBM) excavation, yet raw TBM monitoring data are often affected by shutdowns, machine adjustments, and transient disturbances, weakening their correlation with geological conditions. This study proposes an interpretable data-driven framework for rock mass classification using stable TBM operational data from the Beishan underground research laboratory (URL). A total of 14,157 stable segments were extracted from 4421 excavation cycles, and 13 representative features were selected using Pearson correlation and mutual information. Due to limited Class IV samples, the task was formulated as a three-class problem. Four ensemble learning models (random forest, gradient boosting decision tree, extreme gradient boosting, and light gradient boosting machine) were developed, with the light gradient boosting machine achieving the best performance (accuracy 94.468%, macro F1-score 92.756%). Misclassifications mainly occur between adjacent classes due to overlapping TBM responses and alignment effects, while SHAP analysis identifies machine attitude, thrust, gripper response, displacement, and support behavior as key predictors. The framework provides an interpretable approach for extracting rock mass information from TBM data.