A multi-module deep learning framework with graph-based network and crack attention for tunnel lining crack segmentation from LiDAR point cloud

Cracks in railway tunnel linings pose significant safety risks and demand timely, condition-based maintenance. While LiDAR point clouds offer high-resolution structural data, crack detection remains challenging due to severe class imbalance, and interference from surface contamination, construction joints, and other non-crack features. This study presents PointCrackNet, a novel deep learning framework for accurate crack segmentation in LiDAR-acquired railway tunnel point clouds. PointCrackNet integrates Dynamic Graph Convolution, Graph Attention, and Spectral Convolution with a specialized crack-enhancement attention module, capturing fine-scale crack details while preserving global structural context. A cross-attention mechanism enables efficient feature propagation across network layers, and a multi-component loss function composed of weighted cross-entropy, edge-aware, and continuity terms is introduced to address class imbalance, improve boundary localization, and enhance structural consistency along crack paths. This framework adapts convolutional and attention mechanisms originally developed for 2D images to unstructured 3D point clouds, overcoming challenges posed by geometric irregularity and spatial discontinuity. Experimental validation used our self-built dataset of over 627 million annotated points from a 1-kilometer railway tunnel, acquired using kinematic high-precision LiDAR scanning and multiview fusion. PointCrackNet achieved a crack region IoU of 65.5 %, surpassing other segmentation models. Ablation studies highlight the unique contribution of each module, with the crack attention mechanism providing the largest performance gain. PointCrackNet demonstrates superior capability in detecting crack continuity and edge precision, offering a robust and transferable AI module for digital twin platforms in tunnel health monitoring. This workflow enables automated, large-scale diagnostics and supports smart, data-driven maintenance of underground infrastructure.