• Physics-based Sim-to-Real paradigm resolves data scarcity in civil engineering DL. • GridSegNet eliminates semantic ambiguity at complex node-rod interfaces via priors. • Topology-guided workflow maps discrete scans to high-fidelity geometric digital twins. • Graph-based topological reconstruction builds deformation fields under severe occlusion. • Model-free framework achieves sub-millimeter accuracy in 90s for million-scale data. Automated deformation monitoring of large-span space grid structures using point clouds is frequently constrained by data scarcity and a strong reliance on prior design models. To address these, a model-free geometric perception framework is proposed. First, a virtual scanning mechanism generates synthetic data with non-structural artifacts, solving data scarcity with zero annotation cost. This enables GridSegNet to achieve 95.52% mIoU, effectively eliminating semantic ambiguity at complex node-rod interfaces. Subsequently, an automated workflow is established: utilizing pre-trained models for semantic perception; employing graph theory-based topological reconstruction to build high-fidelity geometric digital twins under severe occlusion. By autonomously quantifying the discrepancy between the reconstructed axis and the inferred topological baseline, the framework generates self-referenced deformation vector fields, effectively isolating and capturing local geometric anomalies (e.g., rod buckling). The framework achieves a 0.72 mm RMSE in synthetic evaluations and is further validated through a field deployment on an actual 32 m × 21 m space grid structure. In this zero-shot real-world scenario, the system successfully identifies physical buckling with an average deviation of 0.82 mm compared to expert manual measurements while processing million-scale point clouds in 90 seconds. By bridging the semantic and topological mapping from raw scans to digital twins, this work provides robust theoretical support for constructing cognitive construction automation systems with self-referenced diagnostic capabilities.
Zhang et al. (Wed,) studied this question.