Utilizing Deep Learning for the Automated Extraction of Rock Mass Features from Point Clouds

Given that traditional joint set identification faces challenges due to labor intensity, safety concerns with accessing steep slopes, and limited sampling locations for measurements, there is an urgent need for more efficient techniques to align with the pace of mine development. This study proposes a deep learning technique for the automated identification of joint sets within 3D point cloud models of rock masses. The process commences with the classification of joints on the 3D surface of a rock mass, accomplished through the training of a deep neural network architecture, and validation using manually labeled datasets. Following this, individual joint surfaces are distinguished using the Density-Based Scan with Noise clustering algorithm. Subsequently, the orientations of the identified joint surfaces are determined by fitting least-square planes using the Random Sample Consensus method. Finally, the joint planes are categorized into distinct joint sets, and the dip direction and dip angle for each set are computed. The effectiveness of this methodology is assessed through a case study, demonstrating that the proposed approach is rapid, precise, and resilient.