关键词: 盾构隧道, 渗漏检测, 稀疏卷积, 激光雷达

Abstract: Traditional convolutional neural networks face challenges in accurately detecting leakage in shield tunnels,particularly due to feature distortion and inefficiencies in handling sparse point cloud data.To address this,we propose a Sparse U-Net based leakage detection method,leveraging three-dimensional LiDAR point clouds.This method incorporates voxelization,hash table and rule-based sparse convolution operations to efficiently capture linear leakage features.An encoder-decoder architecture is employed for precise leakage segmentation,and Focal Loss is introduced to address class imbalance.Experimental results demonstrate the proposed method significantly improves both accuracy and computational efficiency,achieving increases of 5.52% in IoU and 3.41% in accuracy compared to traditional methods,providing an efficient and reliable solution for leakage detection in shield tunnels.

Key words: shield tunnel, leakage detection, sparse convolution, LiDAR