Exploring Computer-based Rock Mass Characterization

The precision and representativeness of manual rock mass characterization result are often limited by site constraints or practices due to accessibility, safety, time, and unfavourable environmental conditions. In this study, we have formulated a computerized workflow for rock mass characterization by adopting the NGI Q-system, with the aid of laser scanning that could be easily deployed on underground construction sites. This study also explores the application of artificial intelligence, specifically deep learning, in the proposed workflow for rock mass characterization based on the relationship between Rock Quality Designation (RQD) and number of discontinuity sets (Jn) and the block shapes and sizes. In order to effectively train the deep learning model, we propose a novel technique called Synthetic Rock Mass Point Cloud (SRMPC) to generate point clouds model of rock mass with varying RQD and discontinuity sets. SRMPC utilizes an implicit approach with signed distance functions to model rock mass geometry and employs the sphere tracing technique for efficient point cloud generation. A synthetic dataset comprising 35 classes of RQD and Jn combinations was created to train and test a deep learning model, PointNet++ (Qi et al., 2017), for classification. The deep learning model exhibited a certain ability to identify the features of the synthetic rock mass with varying RQD, but faced challenges in distinguishing rock masses with similar RQD and discontinuity sets. To address this problem, a revised 20-class classification model was developed and trained, resulting in improved accuracy.