ABSTRACT
Accurate rock mass classification is crucial in tunnel boring machine (TBM) tunnelling, offering effective support to the surrounding ground and safe excavation. Traditional borehole drilling methods, which assume rock mass classifications remain constant between boreholes, can be misleading. To address this, we define reliable rock mass classification within a 2-metre radius around boreholes. In the Yinsong water diversion project, 12962 boring cycles were conducted alongside 275 boreholes, yielding 447 reliable rock mass classifications. This created an imbalance between operational and geotechnical data, with 447 labelled samples and 12515 unlabelled samples. A random forest classifier trained on limited labelled data is prone to overfitting and lacked robustness. We employ semi-supervised learning to incorporate unlabelled samples, resulting in significantly improved accuracy. The findings offer valuable insights into accurate rock mass classification in TBM tunnelling.
