A deep learning-based prediction model for rock-breaking efficiency of milling excavators

Abstract This study proposes a high-precision prediction framework for rock-breaking efficiency in geotechnical engineering, aimed at advancing intelligent construction technologies. This study utilizes Convolutional Neural Network (CNN) to extract deep features-such as edge textures, fracture distributions, and mineral compositions-through its convolutional and pooling layers. These features are used to construct high-dimensional representation vectors of the rock mass. The extracted features are then combined with construction parameters and input into an eXtreme Gradient Boosting (XGBoost) model. An ensemble learning strategy is employed to improve the model’s generalization performance. The model is trained and evaluated using 11,898 tunneling cycles from the Yinsong Project, with the dataset split 70:30 into training and testing sets. Experimental results demonstrate that the CNN-XGBoost model achieves a Mean Squared Error (MSE) of 0.085 and a coefficient of determination (R 2 ) of 0.912 on the test set, significantly outperforming alternative models. The Mean Absolute Percentage Errors for cutterhead rotation speed and advance rate are 5.72% and 8.28%, respectively-both below the 10% threshold-confirming the model’s high predictive accuracy in real-world scenarios. Overall, this study offers a viable approach for integrating deep learning with numerical simulation in geotechnical engineering and provides practical insights for modeling and optimization in related domains.