MS-RAG: A Multimodal Retrieval-Augmented Framework for Digital Twins with Knowledge Graph Reasoning

Traditional Retrieval-Augmented Generation (RAG) frameworks struggle to effectively handle multi-source heterogeneous data in digital twin scenarios, leading to insufficient accuracy in component relationship analysis and fault reasoning. This paper proposes a Multi-Stage Retrieval-Augmented Framework (MS-RAG), which establishes a knowledge base through a multimodal dataset fusing real-time equipment parameters and industry specifications. It models “fault-symptom-maintenance” causal relationships using a knowledge graph and generates enhanced contexts with causal chains via vector retrieval and knowledge graph logical reasoning. Experiments show that the method significantly improves fault diagnosis accuracy and reduces hallucination rates, providing an efficient solution for complex fault diagnosis in digital twins.