ScholarWorks@성균관대학교

초록

The semantic segmentation of point clouds in building MEP (Mechanical, Electrical, and Plumbing) systems provides a system-level foundation for digital modeling and energy management. However, the densely interwoven and suspended nature of MEP components leads to fixed, irreversible occlusion patterns in point cloud data, introducing inductive biases that impair model generalization and segmentation accuracy. To address this, we propose a hybrid perception synergistic learning framework that integrates heterogeneous data sources. First, occlusion-aware synthetic point clouds are generated based on a multi-view convex hull algorithm, introducing compensation and loss simulation for the occluded regions in real point clouds, providing the model with enhanced implicit geometric priors. Then, a multi-granularity deep learning method, MEP-PNeXt, is proposed to establish a feature extraction mechanism that spans from global uniform sampling to local topology preservation. This approach reduces the reliance on similar geometric patterns caused by occlusion, while enhancing occlusion-aware topological modeling and the understanding of fine-grained features. Finally, the synergistic framework conducts hybrid training on MEP-PNeXt using both real and occlusion-aware data, followed by testing and transfer to real-world scenarios. Experimental results show that the framework yields a cumulative improvement of 12.99 % in mean intersection over union. The enhanced robustness and segmentation performance are particularly notable in small components (conduit and hangers) and heavily occluded tubular elements (round duct and pipe). This technology can be integrated into as-built MEP digital twins and automated inspection systems, facilitating the advancement of intelligent building maintenance and energy efficiency optimization. © 2025 Elsevier B.V.

키워드