An intelligent recognition and classification method for TBM tunnel surrounding rock based on cross-attention transformer and multi-source data fusion

Accurate classification of surrounding rock is vital for ensuring the safety and efficiency of TBM operations. To address the limitations of existing methods, which often overlook ascending-stage dynamics, encounter difficulties in multi-source data fusion, and lack interpretability, this study proposes a Cross-Attention Transformer with XGBoost (CA-Trans-XGBoost). The study uses Section IV of the Yinsong Water Diversion Project as a case study, collecting and organizing 802 days of operational data and cutter replacement records. A Transformer encoder was applied to extract dynamic features from the ascending stage, while an MLP modeled structured features from the stable stage. A cross-attention mechanism was introduced to enhance feature interaction, and fused features were further processed with XGBoost for classification and feature importance analysis. Results show that CA-Trans-XGBoost achieves the best performance among six models, with an Accuracy, Precision, Recall, and Macro-F1 of 95.0 %, 93.6 %, 92.6 %, and 92.9 %, respectively. The model shows clear advantages in identifying minority classes II and V. Further analysis confirms that a 30-s ascending stage is the optimal temporal window. The proposed method balances predictive accuracy and interpretability, providing support for intelligent TBM excavation and parameter optimization. • A cross-attention Transformer improves the accuracy and interpretability of TBM surrounding rock classification. • Multi-source fusion integrates TBM operating parameters with disc cutter replacement records. • Causal learning and ablation experiments confirm robustness and generalization of the framework. • Cross-attention captures interactions among features, thereby enhancing feature representation.