Mwd-based real-time identification of rock weathering: A comparison of supervised and unsupervised machine learning methods

A large-scale database linking six-dimensional tunneling parameters to rock weathering conditions was constructed from measurement-while-drilling (MWD) data during tunnel construction utilizing the drill-and-blast method. A data-driven predictive model was developed, integrating random search (RS) with a random forest (RF) classifier, for efficient, real-time identification of unexcavated rock weathering levels. Nine comparative models combining CART and XGBoost with hyperparameter optimization were also evaluated using the Macro F₁ score. The classification performance, applicability, and generalization capabilities of the proposed model were analyzed. Additionally, several unsupervised models were employed for clustering analysis, leveraging inherent data characteristics. Results demonstrated that the RS-RF algorithm outperforms other models in identifying rock weathering levels, with rotation pressure as the primary classification feature. For large, highly similar datasets, supervised algorithms exhibited superiority over their unsupervised counterparts. Consequently, the proposed model enables optimal classification of rock weathering conditions, thereby providing essential information for subsequent support and measurement.