基于改进PointNet++隧道衬砌渗漏水识别方法

A Tunnel Lining Leakage Detection Method Based on the Improved PointNet++

摘要: 本研究针对传统点云投影方法在隧道衬砌渗漏水识别中的信息损失和投影角度依赖性强等问题，提出了一种基于改进PointNet++的直接三维识别方法。该方法通过在三维空间中直接处理点云数据，避免了降维带来的几何信息损失。通过引入强度感知采样策略、反射强度加权聚合机制和病害注意力机制，充分利用渗漏水的物理特性(低反射强度、竖向流淌方向)增强特征表达能力。进一步采用主成分分析(PCA)降维优化计算效率，并使用Focal Loss解决样本不均衡问题。实验结果表明，所提方法在自建隧道衬砌渗漏水点云数据集上的F1-Score达到85.8%，IoU达到73.9%，相比基线PointNet++模型分别提升7.3和9.2个百分点，满足实际工程应用需求。该方法为隧道衬砌渗漏水的自动化检测提供了有效的技术参考。

Abstract: This study addresses the issues of information loss and strong projection angle dependence in traditional point cloud projection methods for identifying tunnel lining leakage water. It proposes a direct three-dimensional recognition method based on the improved PointNet++. This method directly processes point cloud data in three-dimensional space, avoiding the geometric information loss caused by dimension reduction. By introducing intensity-aware sampling strategy, reflection intensity weighted aggregation mechanism, and disease attention mechanism, it fully utilizes the physical characteristics of leakage water (low reflection intensity, vertical flow direction) to enhance feature expression ability. Further, principal component analysis (PCA) is used to reduce dimensionality and optimize computational efficiency, and Focal Loss is employed to solve the problem of sample imbalance. Experimental results show that the proposed method achieves an F1-Score of 85.8% and an IoU of 73.9% on the self-built tunnel lining leakage water point cloud dataset. Compared with the baseline PointNet++ model, it improves by 7.3 and 9.2 percentage points respectively, meeting the requirements of practical engineering applications. This method provides an effective technical reference for the automatic detection of tunnel lining leakage water.