Typical Fault Diagnosis Model of Nuclear Power Plant Combined With Knowledge Driven and Data Driven

Abstract In order to prevent potential radioactive leaks and maintain personnel safety, fault diagnosis in nuclear power plants is crucial. Knowledge-driven and data-driven approaches are two predominant directions in nuclear power plant fault diagnosis. Knowledge-driven approach relies on prior professional knowledge, rules and model for state monitoring and diagnosis, offering strong logic and interpretability. However, it often yields poor results in complex systems. On the other hand, data-driven approach utilizes black-box inference to establish the mapping relationship between complex system characteristics and fault modes. However, this approach lacks interpretability in its operation mechanism, and it is deemed unacceptable for high-security industries such as nuclear power. To address these challenges, this paper introduces the concept of graph neural network and proposes an innovative idea of integrating knowledge-driven and data-driven approaches for fault diagnosis in nuclear power systems. The proposed method embeds the fault propagation path into the data-driven inference process by constructing a graph based on feature nodes and fault type nodes. Furthermore, we analyze how different methods of sampling features and aggregating data influence the diagnostic results during the process of inferring fault paths using graphs with corresponding fault labels. We implement and verify our scheme using a comprehensive simulation simulator for nuclear power devices. The results demonstrate that our method effectively carries out fault diagnosis by combining both interpretable graph-based inference with powerful nonlinear fitting capabilities offered by data-driven approaches.