Abstract
The project addresses the challenge of integrating heterogeneous, multi-modal, and unstructured data into holistic digital twins for infrastructure. Its objective is to establish a robust data backbone (CARE-DS) and an interactive interlinking platform (CARE-DIVE) that enables seamless data management, analysis, and synthesis across diverse engineering domains.
Methodologically, the project develops a hybrid Retrieval-Augmented Generation (RAG)-inspired query layer that combines symbolic and sub-symbolic approaches. A GeoSPARQL-enabled RDF knowledge graph provides precise, ontology-driven information retrieval, while a vector database with domain-specific encoder models supports fuzzy matching and pattern recognition across incomplete or ill-structured data such as sensor readings, site photos, point clouds, text documents and design drawings. This dual strategy ensures both semantic transparency and flexibility, allowing queries across structured and unstructured modalities.
The work plan encompasses ontology-driven infrastructure, encoder training, cross-modal alignment, and validation through two practical use cases: minimally invasive bridge strengthening under live load and robotic assembly of modular concrete shells. Key contributions include (i) a unified ontology-based data backbone, (ii) lightweight encoder models for multiple data modalities, and (iii) a hybrid reasoning interface that integrates symbolic and sub-symbolic methods. Together, CARE-DS and CARE-DIVE provide scalable, interoperable foundations for managing complex infrastructure data and advancing digital twin applications.
