머신러닝과 SHAP를 활용한 터널 발파 영역별 비장약량 예측 및 영향 인자 해석

초록

In tunnel blasting, the cut, stoping, lift, and contour zones serve different purposes, and the required charge differs by zone. However, a single specific charge based on the whole face has commonly been applied uniformly to every zone. In this study, a data-driven approach was applied to all four blast zones to examine the inter-zone differences and the dominant factors of each zone. A database of 334 datasets was constructed from 24 domestic NATM tunnels. Eleven machine learning algorithms were compared for each zone through 5-fold cross-validation, with strict train/validation/test separation, and XGBoost was selected for every zone. On the test set, the coefficient of determination was 0.95, 0.84, 0.86, and 0.85 for the cut, stoping, lift, and contour zones, respectively, with MAPE between 5.0% and 8.9%. The zone-dependent distributions showed distinct central tendencies and variabilities, and the dependence on rock-mass grade differed by zone; the contour zone remained nearly constant regardless of RMR. The SHAP analysis showed that the dominant factor differed by zone: the cut method for the cut zone, the cross-sectional area for the stoping zone, the unit weight for the lift zone, and the rock type for the contour zone.

키워드