Classification of Rock Hardness at Tunnel Faces Based on a Drilling Parameter Cloud Map and Convolutional Neural Network

Abstract

Accurately measuring rock hardness at a tunnel face is crucial for evaluating rock mass quality and ensuring construction safety. This study developed a fast and accurate classification method for rock hardness. Firstly, six drilling parameters (DPs)—of hammer pressure (Ph), feed pressure (Pf), rotary pressure (Pr), feed speed (Vp), water pressure (Pw), and water flow (Qw)—and the rock hardness were collected from 1,894 tunnel face boreholes. Then a new method for generating cloud maps of DPs was developed using data correlation analysis and feature transformation. Finally, six algorithms, i.e., Inception-V3 convolutional neural network (CNN), random forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting (LightGBM), were utilized to construct hardness classification models of rocks at tunnel face using DP cloud maps or DPs as input parameters and rock hardness as output parameters. Furthermore, the performance of models before and after DP cloud map generation was analyzed using accuracy, recall, precision, F1 value, receiver operating characteristic (ROC) curves, and area under curve (AUC) values as evaluation indexes. The outcomes showed that the Inception-V3 CNN model with DP cloud maps as input parameters performed optimally, achieving an accuracy of 94.18%.