Deep Learning-Based Rock Muck Detection on Point Clouds During TBM Excavation

During tunnel boring machine (TBM) excavation, rock mucks will be produced by rock-machine interaction and shipped out of tunnels. The morphology characteristics of rock mucks are essential evidence to indicate whether the TBM operating parameters can adapt to the rock mass conditions ahead of the tunnel face. Therefore, knowing this information can help determine better TBM operation parameters and guarantee safe and stable TBM construction. Previously, some works used image processing methods to obtain the particle size distribution of rock muck. Considering the difficulties of capturing rock images in a dim tunnel environment and using rock images to depict 3-D rock morphology characteristics, we propose to use a laser line-scanning sensor to capture 3-D point clouds of rock muck and leverage deep learning to detect rock’s bounding boxes. We prepare a 3-D point cloud dataset of 263 rock muck samples through our designed simulation system. To reach real-time detection for rock muck, we adopt a one-stage object detection neural network with a YOLO3D structure. Furthermore, we incorporate an additional curvature feature to enhance the discrimination of adjacent rocks and tailor specific anchors to facilitate the determination of bounding boxes. In addition, we employ the generalized intersection over union (GIoU) loss function to achieve improved performance. Then, we can calculate the volume size distribution for the rock muck through detected bounding boxes. Finally, the efficiency of our proposed method is verified through comprehensive experiments on the proposed rock muck point cloud dataset.