Intelligent identification of rock mass structural based on point cloud deep learning technology

• Published in: Construction & building materials 2024-12, Vol.456, Article 139340
• Main Authors: Li, Xu, Song, Zhanping, Zhi, Bin, Pu, Jiangyong, Meng, Chen
• Format: Article
• Language: English
• Subjects: Deep learning; Discontinuity of rock mass; Point cloud; RandLA-Net
• ISSN: 0950-0618
• DOI: 10.1016/j.conbuildmat.2024.139340

To accurately understand the mechanical behavior and stability of rock masse, identifying rock mass structures is crucial, as it directly influences the selection and performance of engineering materials, thereby affecting the design and safety of geotechnical engineering projects. Traditional image-based methods for extracting rock mass structures are often affected by environmental factors during data acquisition, and because these methods rely on two-dimensional (2D) imagery, they cannot fully capture the complex three-dimensional (3D) characteristics of rock structures. Additionally, existing 3D point cloud segmentation techniques are predominantly threshold-based, which limits their accuracy and efficiency, making them inadequate for addressing engineering challenges in complex geological environments. To overcome these limitations, this study introduces a deep learning-based approach for analyzing 3D point cloud data acquired using a Unmanned Aerial Vehicles (UAV)-mounted Light Detection and Ranging (LiDAR) system. The proposed method leverages enhancements to the RandLA-Net deep learning framework by incorporating advanced preprocessing and residual modules, thus enabling precise and efficient identification of structural planes within the LiDAR-generated point cloud data. When applied to a mining site in China, the segmentation approach achieved a prediction accuracy of 88 % and a mean Intersection over Union (mIoU) of 74 %. Ablation studies validated the improvements, and a comparative analysis with other point cloud-based structural identification algorithms demonstrated the superior performance of this framework in accurately delineating structural planes using UAV-mounted LiDAR data. •The point cloud data of rock mass structure are classified intelligently.•Deep learning techniques are used to automatically classify rock mass point clouds.•Randla-Net is improved.