Automated Identification of Rock Discontinuities from 3D Point Clouds Using a Convolutional Neural Network

Accurate and quick mapping of the rock discontinuity is essential for rockmass stability analysis. Traditional methods for rock discontinuity identification are often subjective, time-consuming, and dangerous. To improve the efficiency of rock discontinuity identification, a deep learning algorithm was developed. A case in Tianjin, China was utilized to demonstrate the method, and the GoogLeNet convolutional neural network (CNN) was employed to identify rock discontinuities. The xyz-coordinates and point normal were calculated as input data, with 100 randomly selected points used to create training sets (70%) and testing sets (30%). The CNN model was built, followed by being trained using the training sets and verified using the testing sets. Three natural group discontinuities were predicted from the point clouds using the trained model, and individual discontinuities were extracted from each group discontinuity using the density-based spatial clustering of applications with noise (DBSCAN) algorithm. Subsequently, the orientation of each discontinuity was determined based on the principal component analysis (PCA) algorithm. In this case, three sets of rock discontinuities and 486 individual discontinuities were identified. The average error degrees in dip direction and dip angle are 2.36 degrees and 1.30 degrees compared with manual measure orientations using least squares results, indicating deep learning methods have excellent accuracy. A public case conducted in Ontario, Canada was applied to comparatively analyze different algorithms to strengthen the accuracy argument of the method. Furthermore, this study discussed the respective strengths and limitations of deep learning and shallow neural network (SNN) algorithms in the identification of rock discontinuities.