Intelligent classification of tunnel surrounding rock based on multivariable TBM tunnelling features and Optuna-XGBoost

Reliable classification of tunnel surrounding rock using tunnel boring machine (TBM) operational data remains challenging because field records contain non-excavation data, outliers, transient responses, and imbalanced samples, while adjacent rock grades often exhibit overlapping parameter distributions. This study develops and validates an intelligent classification framework integrating multivariable feature extraction and Optuna-based XGBoost optimization. Field data from the Daliangshan No. 1 Tunnel were cleaned using operational-state criteria, combined Z-score and interquartile-range methods, and steady-state tunnelling-stage extraction. Six representative tunnelling parameters were transformed into 36 statistical features, and ADASYN was applied to alleviate class imbalance. The proposed model was compared with unoptimized XGBoost, random forest, K-nearest neighbours, PSO–XGBoost, and BO–XGBoost. On the hold-out test set, the multivariable-feature Optuna–XGBoost model achieved an accuracy, macro-averaged F1 score, and Kappa coefficient of 0.9600, 0.9600, and 0.9467, respectively. Five-fold cross-validation further yielded values of 0.934, 0.938, and 0.927 for the same metrics. Ablation analysis showed that Optuna optimization and multivariable feature expansion increased these metrics by 8.4, 9.0, and 10.9% points over the original-feature XGBoost baseline. Sensitivity analysis of maximum tree depth, learning rate, and number of base learners indicated that model performance remained stable near the Optuna-derived settings. The proposed framework improves the representation of rock–machine interactions and shows potential for online surrounding-rock classification during TBM tunnelling.