Muck volume measurement of earth pressure balance shield using 3D point cloud based on deep learning

Earth pressure balance shield over-excavation leads to ground loss, causing ground collapse and posing a serious threat to infrastructure and transportation. Existing shield muck volume control techniques, which are based on weighing or observing the muck box, are unreliable, time-consuming and have large errors. In this paper, a muck volume measurement algorithm and a 3D point cloud segmentation model are proposed to measure the muck volume and control the shield excavation with high accuracy and automation. The model constructs a point-serialized attentional interaction approach that addresses the limitations of the disordered and sparse properties of the point cloud on the local attentional receptive field. In order to segment the dynamic boundary between muck and box, the model designs normal conditional positional encoding to enhance the spatial characteristic representation of the point clouds. The model was trained and tested based on a real shield muck point cloud. The mean accuracy, mean intersection over union, and overall accuracy of the model are 0.987, 0.971, and 0.983, respectively. The maximum calculation error of the muck volume is 3.65% and the average error is 2.71%, which are less than the errors of existing construction technology. The average processing time of the segmentation model for a single sample is 32.5 s, which is about 60 times faster than manual segmentation. These findings have significant engineering value for shield construction control. The code and muck point cloud samples can be obtained from https://github.com/posuifeng/Muck-point-cloud-segmentation.git, password: 3jpi.