Deep Learning–Based Few-Shot Method for Detection and Evaluation of Tunnel Overbreak and Underbreak

Accurate quantification of overbreak and underbreak is essential for tunnel quality, cost, and stability. Point-cloud methods struggle with sparse labels, heavy site noise, and strict profile continuity. We introduce a three-dimensional (3D)–constrained, multidomain augmentation framework that expands few-shot data while keeping engineering fidelity. We propose a tunnel-aware pipeline with three contributions: (1) a 3D-constrained, multidimensional augmentation that enlarges few-shot data while preserving annular continuity and realistic site disturbances; (2) a geometry- and density-aware network tailored to tunnel point clouds; and (3) a scale-coupling scheme that aligns augmentation perturbations with the network’s receptive fields. On a challenging in situ dataset, the method reaches 88.0% accuracy, outperforming iterative closest point (ICP) + moving least squares (MLS) and vanilla PointNet++ by 36.0% and 9.5%, respectively. Ablations attribute an 18.4% gain to geometric augmentation and a further 3.1% to the dynamic attention module. The framework is robust under few labels and high noise, enabling practical, real-time monitoring with transfer potential to other underground inspections.