TBM rock mass classification using XGBoost and Interpretable Machine learning

Accurate classification of surrounding rock grades during Tunnel Boring Machine (TBM) operations is essential for optimizing excavation parameters and ensuring construction safety. However, traditional classification methods face challenges, particularly in adapting to complex geological conditions and achieving high efficiency in model optimization. To address these issues, this study proposes an identification methodology based on the COA-XGBoost framework. Initially, raw data is preprocessed to eliminate anomalous instances, such as those arising from equipment malfunctions or maintenance activities, using a state discrimination function. To better capture the interaction dynamics between the TBM and the surrounding rock, a novel feature parameter, specific energy, is introduced. The Relief algorithm is employed to evaluate feature importance and assign weights to TBM operational parameters, amplifying the role of critical features while minimizing interference from redundant ones. Subsequently, the Coati Optimization Algorithm (COA) is integrated with the XGBoost classifier, enabling global parameter optimization and enhancing the model’s performance. The proposed method was validated using data from the water conveyance tunnel section of the Yinchuojiliao Project. Results demonstrate that the Relief algorithm effectively enhances the model’s ability to identify critical parameters through feature weighting. The COA-optimized XGBoost model achieves a significant improvement in recognition accuracy compared to traditional approaches, with accuracy gains of approximately 8%, particularly excelling under complex geological conditions such as Class IV and Class V surrounding rocks. The final model achieves an accuracy of 91.5%, with recall and F1 scores reaching 91.5% and 91.4%, respectively. Additionally, five-fold cross-validation highlights the model’s robustness and generalization capabilities. Furthermore, a SHAP-based interpretability analysis clarifies the relative contributions of each feature to the classification outcome, offering valuable insights for practical TBM parameter optimization.