Welcome to the PSAM 18 Abstract Status page.
Abstract SM154Full Paper + Presentation
LLM-Based Retrieval-Augmented Construction of Dynamic Master Logic Models for System Diagnostics of Complex Systems
Abstract
Functional modeling is widely used in reliability and safety assessment to describe how engineered systems achieve their objectives and how failures propagate through their functional structure. Dynamic Master Logic (DML) is a hierarchical functional modeling framework that links high-level objectives to supporting components through explicit logical relationships. However, constructing DML models from technical documentation remains a labor-intensive process, limiting its applicability to large systems. Recent advances in large language models (LLMs) have created new opportunities to assist in extracting structured knowledge from unstructured engineering texts. In this work, we present a retrieval-augmented LLM-based framework for constructing DML models from system descriptions. Retrieved document evidence is used to ground and constrain model generation. The resulting DML logic, including its hierarchical and logical relationships, is encoded as a knowledge graph (KG-DML). The approach employs layer-specific information retrieval to incrementally build the model along the DML hierarchy. The resulting KG-DML supports an interaction phase in which an LLM agent invokes predefined tools to generate diagnostic insights, enabling upward propagation of failure impacts and downward identification of success paths. A case study involving the Low-Pressure Coolant Injection system of a nuclear power plant demonstrates consistent reconstruction of the DML model and stable structural performance across repeated runs. The results suggest that documentation-driven DML construction can scale to larger, document-intensive systems while maintaining the structural consistency required for reliability analysis.
← Check another abstractFunctional modeling is widely used in reliability and safety assessment to describe how engineered systems achieve their objectives and how failures propagate through their functional structure. Dynamic Master Logic (DML) is a hierarchical functional modeling framework that links high-level objectives to supporting components through explicit logical relationships. However, constructing DML models from technical documentation remains a labor-intensive process, limiting its applicability to large systems. Recent advances in large language models (LLMs) have created new opportunities to assist in extracting structured knowledge from unstructured engineering texts. In this work, we present a retrieval-augmented LLM-based framework for constructing DML models from system descriptions. Retrieved document evidence is used to ground and constrain model generation. The resulting DML logic, including its hierarchical and logical relationships, is encoded as a knowledge graph (KG-DML). The approach employs layer-specific information retrieval to incrementally build the model along the DML hierarchy. The resulting KG-DML supports an interaction phase in which an LLM agent invokes predefined tools to generate diagnostic insights, enabling upward propagation of failure impacts and downward identification of success paths. A case study involving the Low-Pressure Coolant Injection system of a nuclear power plant demonstrates consistent reconstruction of the DML model and stable structural performance across repeated runs. The results suggest that documentation-driven DML construction can scale to larger, document-intensive systems while maintaining the structural consistency required for reliability analysis.