ABSTRACT
Point cloud data generated from laser scanners is vital for shotcrete validation in tunnel construction due to its accuracy, aiding in precise application and reducing waste. It ensures quality control by comparing data with design specifications, enhancing safety by identifying potential hazards and facilitating documentation for monitoring and maintenance. Efficiency is improved through rapid data processing, streamlining the construction process, and reducing costs. However, achieving high data accuracy requires processing large volumes of data. Manual processing is time-consuming, delays real-time results, and introduces inconsistencies due to human variability. Deep learning models that directly ingest raw point clouds address these challenges by preserving the full spatial detail of the data while identifying scan contents such as shotcrete, raw rock, rebar mesh, and rock bolts. These semantic segmentation models, including adaptations of U-Net for 3D point clouds and transformer-based architectures, facilitate feature extraction by utilizing the inherent structure of the point clouds along with additional attributes like intensity or RGB. Scalable deep learning techniques make tunnel point cloud processing faster and more consistent, enabling real-time decision-making for safer, more efficient construction. In the Norwegian tunnel project Andfjord Salmon Tunneler for the land-based fish farm Andfjord Salmon, as well as the Kolltveittunnelen project, which are part of the Norwegian RV 555 Sotrasambandet project, LNS has been utilizing AI to remove human-led manual point cloud processing while maintaining accuracy and quality during drill and blast construction. AI-enabled (artificial intelligence) point cloud processing can save 0.5 hours per blast and 1 hour per 50 meters of the tunnel (concrete scans), approximately 12 hours per 100 meters of construction (20 blasts). By semi-automating this workflow, they achieve greater operational flexibility and anticipate synergy effects on other survey workflows such as rock bolt conformance, consistent calculation of concrete thickness, and scan-to-BIM modeling.
