Deep learning-based identification of rock discontinuities on 3D model of tunnel face

Abstract

Discontinuity mapping on tunnel faces is crucial for assessing stability and determining the need for additional reinforcement during tunnel construction. The traditional manual mapping approach is time-consuming and error-prone, necessitating a more accurate and efficient approach. This study explores a novel approach using photogrammetry to reconstruct digital 3D models of tunnel faces, enabling comprehensive discontinuity characterization without any time restriction. Despite challenges in image data collection and processing procedures, photogrammetry proves to be a viable alternative to LiDAR scanning for reconstructing precise 3D models of tunnel faces. Additionally, a deep learning technique is proposed to automatically identify rock mass discontinuities departing from massive random fractures on the 3D tunnel face. Since working directly with 3D models in deep learning is still challenging, the 3D tunnel face model is projected into four 2D images (i.e., RGB, depth map, normal vector, and curvature images) encompassing all necessary information of the 3D model. Afterward, a 2D semantic segmentation deep learning model is trained to identify areas of discontinuity based on the projected multi-2D images. Finally, the identified discontinuities are re-projected onto the 3D model to accurately reflect their original 3D context. Our results indicate that the proposed approach not only automatically and accurately quantifies rock discontinuities but also minimizes subjectivity inherent in manual judgment.