[2F5-OS-19a] OS-19

4:45 PM - 5:00 PM JST(7:45 AM - 8:00 AM UTC)

[2F5-OS-19a-06] Multimodal Data Knowledge Graph Integration to Enhance Reliability of Equipment Maintenance Support

〇Tatsuya Baba¹, Takumi Uezono¹, Kentarou Yoshimura¹ (1. Hitachi Ltd.)

[[online]]

Keywords:

Knowledge Graph,GraphRAG,Ontology,Knowledge Integration

In this study, with the aim of advancing facility maintenance and inspection operations, we proposed a knowledge structuring and retrieval methodology that logically links complex technical drawings with textual information. The novelty of our approach lies in the combination of a four-stage knowledge graph (KG) construction process comprising (1) ontology expansion, (2) KG generation, (3) ontology mapping, and (4) KG integration, and an efficient retrieval method. Specifically, we proposed Multi G-Retriever, which utilizes the Prize-Collecting Steiner Tree (PCST) algorithm to perform subgraph extraction based on multiple subqueries on the constructed KG. The proposed method was evaluated using a multi-hop question-answering task designed for piping systems. The experimental results confirm that our approach demonstrates improved effectiveness compared to the conventional multimodal RAG baseline, enabling reliable inference across both drawings and text.