Accurately reconstructing Mechanical, Electrical and Plumbing (MEP) systems from laser-scanned point clouds is often hindered by structural occlusions, sensor noise, and extreme scale imbalance between large pipes and small fittings. This study presents a hybrid framework, driven by both knowledge and data, for automated pipeline BIM updating. To tackle scale variance, we implement a coarse-to-fine segmentation strategy using Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to isolate pipeline instances before segmentation with PointNeXt. Furthermore, a logic-based refinement module integrates geometric and topological priors from the design BIM to correct coordinate deviations in incomplete datasets. Finally, graph isomorphism analysis enables automated topological mapping between unstructured point cloud instances and structured BIM components. Experimental results from a dense shopping center case study demonstrate that the framework achieves a semantic segmentation mIoU of 74.45% and reduces the average spatial coordinate error to within 7 mm. Notably, the automated workflow compressed the modeling time from 3–5 days to approximately 3 h, offering a robust solution for digital twin-oriented facility management.
Wang et al. (Wed,) studied this question.