PhyGeo-KG: Physics-Regularized Distant Supervision for Multimodal Geometric Knowledge Graph Construction in Catenary...

Area of Science:

• Engineering
• Computer Science
• Infrastructure Management

Background:

• High-speed railway catenary maintenance demands integrated knowledge bases linking records and geometric models for digital twin decision support.
• Current methods face challenges like label sparsity, weak instance grounding, and poor physical interpretability in engineering knowledge graph construction.

Purpose of the Study:

• To introduce PhyGeo-KG, a novel physics-regularized distant supervision framework for high-fidelity multimodal geometric knowledge graphs in catenary maintenance.
• To address limitations in existing knowledge graph construction for engineering applications, focusing on interpretability and physical consistency.

Main Methods:

• Developed a Semantic-Geometric-Physical-Procedural ontology to unify heterogeneous engineering data.
• Implemented a deterministic grounding strategy aligning text with Industry Foundation Classes (IFC)/Building Information Modeling (BIM) entities via geometric interfaces.
• Employed a physics-aware refinement and ontology-driven evolution process to ensure physical plausibility and graph expansion.

Main Results:

• PhyGeo-KG demonstrated improved relation precision and physical consistency on a real-world dataset.
• The framework effectively suppressed semantic hallucinations, enhancing the reliability of the constructed knowledge graph.
• A case study confirmed its utility for instance-level fault localization within semantic digital twins.

Conclusions:

• PhyGeo-KG offers an interpretable and transferable foundation for physics-regularized multimodal geometric knowledge graph construction.
• The framework supports digital twin-enabled decision support in catenary maintenance, with potential for future sensor-integrated applications.
• This approach advances the reliability and physical consistency of knowledge graphs in complex engineering domains.