Towards Automated Lithology Classification in NATM Tunnel: A Data-Driven Solution for Multi-dimensional Imbalanced Data

To fully grasp the lithology of unexcavated tunnel geology, a correlation database using measurement-while-drilling (MWD) information from the NATM tunnel excavation process was established, resulting in a multi-dimensional imbalanced dataset consisting of 7216 entries. By integrating borehole imaging and expert interpretation, drilling parameters were aligned with lithology data. A hybrid ensemble model, combining adaptive synthetic sampling (ADASYN), grid search (GS) hyperparameter optimization, and eXtreme gradient boosting (XGBoost), is proposed for intelligent lithology classification. Various machine learning models, incorporating hyperparameter optimization and oversampling algorithms, were employed, cumulatively generating 12 classifiers for Macro F₁ performance comparison. Comprehensive analysis showed that the GS-ADASYN-XGBoost algorithm outperformed the other hybrid models in classifying different lithologies. Water pressure was identified as the key feature influencing lithology classification, followed by water flow. Setting the oversampling proportion to 0.2, the ADASYN method effectively optimized the data imbalance ratio, significantly enhancing classifier performance. This improvement was most notable for the least represented lithology category, chlorite, with an increase of 1.27 times compared to no oversampling. The proposed model provides valuable insights for geological interpretation of the tunnel face.