Development of a 3D-point cloud-based spatiotemporal series model for tunnel rock mass discontinuities prediction

Predicting the three-dimensional (3D) distributions of discontinuities within rock masses is crucial for evaluating tunnel stability. However, this task is challenging due to the inherent opacity of rock, which prevents the direct observation of discontinuities. Most current methods for predicting discontinuities are based on extracting the two-dimensional intersection lines of spatial discontinuities. In this paper, we propose a novel, purely visual approach to analyze and predict the 3D distributions of discontinuities in rock masses. In this method, a 3D model of the tunnel face is constructed based on motion prediction and multi-view stereo vision, and the development of discontinuities is then predicted. Each set of discontinuities is projected onto the virtual tunnel face using a convex hull algorithm, creating a virtual trace. A newly developed algorithm for predicting spatiotemporal sequences, which incorporates a self-attention mechanism and a zigzag recurrent transition mechanism, is then applied to predict the evolution of discontinuities. For testing and verification, we used smartphones to collect surface data on multiple sets of excavated rock from the Bimoyuan Tunnel in Sichuan, China. Extensive experiments involving these surface data demonstrated the effectiveness of our proposed method. The findings provide technical support for predicting tunnel collapse and ensuring tunnel safety.