Deep learning-based point cloud semantic segmentation for tunnel face excavation areas in drilling and blasting tunnels

Abstract

The tunnel face excavation area in drilling and blasting tunnels contains rich construction and geological information, with each part requiring unique measurement indicators and analysis. Precise segmentation of these components is urgently needed to enable automated measurement. This paper proposes a deep learning-based point cloud semantic segmentation method to automatically and accurately segment different parts of the tunnel face excavation area during the construction phase of drilling and blasting tunnels. First, point cloud data of the excavation area were collected and labeled the tunnel face, excavation profile, ground surface, and initial support section. A sample dataset was then created by calculating normal vectors, applying the synthetic minority oversampling technique (SMOTE), and using data augmentation techniques. A point cloud semantic segmentation network was subsequently constructed and trained, and its performance was subsequently evaluated using metrics such as training accuracy, testing accuracy, mean intersection over union (mIoU), and confusion matrices. The experimental results demonstrate that the proposed method achieves high-precision semantic segmentation in the complex, irregular tunnel face area of drilling and blasting tunnels, with a maximum training accuracy of 0.9854, a validation accuracy of 0.9536, a testing accuracy of 0.9551, and all mIoU values above 0.91. The discussion examines the impact of normal vector features and data augmentation on model performance and demonstrates that data augmentation enhances training effectiveness and generalizability. These findings suggest that the proposed method has substantial potential for automated measurement in drilling and blasting tunnel construction.