Constructing Dynamic Master Logic Models as Knowledge Graphs for Complex System Diagnostics Using Retrieval-Augmented Large Language Models

Modern engineering systems, especially in critical sectors like nuclear energy and aerospace, are becoming increasingly complex, involving numerous interconnected components and dynamic behaviors. Traditional diagnostic approaches, such as fault tree analysis (FTA) and event tree analysis (ETA), rely heavily on predefined failure sequences and event-based logic. While effective in certain contexts, these methods struggle with scalability and adaptability when faced with large-scale, evolving systems. They often cannot account for unanticipated failures or cascading effects, limiting their utility in ensuring safety and reliability.

To address these challenges, recent advances in functional modeling frameworks, notably Dynamic Master Logic (DML), have provided a hierarchical approach to represent system objectives, functions, and dependencies explicitly. DML captures the causal and structural relationships within a system, enabling more comprehensive reasoning about failure propagation and safety. However, manual construction of DML models remains labor-intensive, requiring expert interpretation of extensive technical documentation. This bottleneck hampers the widespread adoption of DML in complex, document-rich environments.

In this context, the integration of artificial intelligence, particularly large language models (LLMs) like GPT-4, offers promising solutions. These models excel at semantic understanding, text summarization, and content generation. Yet, their unstructured output and potential for hallucination pose challenges for reliable system modeling. To overcome these issues, this study introduces a retrieval-augmented generation (RAG) framework that combines the strengths of LLMs with targeted information retrieval. By retrieving relevant document segments based on semantic similarity and guiding LLMs to generate structured, hierarchical representations, the framework automates the construction of Knowledge Graph-based DML (KG-DML) models.

The proposed methodology involves preprocessing system documentation, embedding texts into vector stores, and sequentially retrieving and generating model components at each DML layer. The process ensures that the generated models preserve hierarchical dependencies and logical relationships, crucial for accurate diagnostic reasoning. Validation metrics, including precision, recall, and logical consistency, are employed to assess the structural quality of the models. Experiments conducted on the Low-Pressure Coolant Injection system of a decommissioned boiling water reactor demonstrate that the automated KG-DML models achieve high accuracy and stability across multiple runs, confirming their applicability to large, complex systems.

This work significantly advances the field by providing a scalable, automated pipeline for functional system modeling, reducing reliance on manual effort, and enhancing diagnostic capabilities. The integration of retrieval-augmented generation with hierarchical modeling opens new avenues for real-time safety assessment, fault diagnosis, and system optimization. Future directions include incorporating real-time sensor data, expanding multi-modal information integration, and developing automated validation tools, aiming to realize fully autonomous, intelligent system management in safety-critical industries. Overall, this research marks a pivotal step toward intelligent, scalable, and reliable system diagnostics driven by AI and formal modeling principles.