Abstract
Understanding as-built conditions of mechanical, electrical and plumbing (MEP) systems is important for operation and maintenance of buildings and infrastructures. However, acquiring large-scale annotated point clouds for training robust segmentation models is labor-intensive. This paper presents a simulation-to-real approach for manual-annotation-free and cross-domain semantic segmentation of MEP point clouds. A sensor perception model is developed to generate BIM-based synthetic point clouds that reflect real scans by modeling realistic sampling density, occlusion and sensor characteristics. Iterative self-learning is introduced to support cross-domain representation learning, in which feature representations are refined across synthetic and real point clouds. A probability-driven label updating mechanism is proposed to exploit spatial consistency and structural context in MEP systems. The approach jointly addresses data and feature learning to enhance the semantic segmentation performance, improve transferability of deep learning models trained on synthetic data, and mitigates domain gap induced by complex MEP geometries and sensing artifacts.
| Original language | English |
|---|---|
| Article number | 107143 |
| Journal | Automation in Construction |
| Volume | 190 |
| DOIs | |
| Publication status | Published - Oct 2026 |
Keywords
- BIM
- Domain adaptation
- MEP system
- Self-learning
- Semantic segmentation
- Simulation-to-real
- Synthetic point cloud
ASJC Scopus subject areas
- Control and Systems Engineering
- Civil and Structural Engineering
- Building and Construction
Fingerprint
Dive into the research topics of 'Iterative self-learning with BIM-based sensor perception for manual-annotation-free cross-domain segmentation of MEP point clouds'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver