Session: 12-02 Risk Assessments and Management - Session 2

Paper Number: 134815

134815 - Typical Fault Diagnosis Model of Nuclear Power Plant Combined With Knowledge Driven and Data Driven 

Abstract: 

In order to maintain the normal operation of nuclear power plants and protect the safety of operators, it is very important to carry out fault diagnosis for nuclear power plant. Knowledge driven and data driven are the two mainstream methods for fault diagnosis in nuclear power plants, each with unique advantages and challenges. Knowledge-driven approach relies on prior professional knowledge, rules, and models for state monitoring and diagnosis, offering strong logic and interpretability. However, when dealing with complex systems, the performance of this method is often limited because there are artificially set thresholds in knowledge driven methods, which can easily cause false positives in the face of complex system structures and interrelationships. On the contrary, data-driven methods can more flexibly capture the mapping relationship between the characteristics and fault modes of complex systems through black box reasoning. This method is based on a large amount of data for learning, using machine learning and statistical techniques to reveal the potential patterns of the system. However, an obvious drawback of data-driven approach is its black box nature, where the mapping relationships learned by the model are difficult to explain. In high safety requirements industries, especially nuclear power, decision-makers often need clear explanations to understand the basis of fault diagnosis, so the interpretability of data-driven methods becomes a limiting factor.  Given the complementarity of the two methods, a combination of them is a worthwhile investigation. this article introduces the concept of graph neural networks and proposes an innovative fusion of knowledge and data-driven fault diagnosis method for nuclear power systems. This method embeds the fault propagation path into the data-driven inference process, and constructs a graph that includes feature nodes and fault type nodes, so that the structure of the graph corresponds to the fault label. The core of this fusion method lies in introducing prior information of knowledge within a data-driven framework, thereby improving the performance and interpretability of the model in complex systems.  In the study, we analyzed in detail the impact of data feature sampling and aggregation methods on the fault path inference process and diagnostic results.       To verify the effectiveness of the method, we implemented and validated the proposed scheme using nuclear power plant simulator. The research results indicate that this fusion method has achieved significant results in practical applications and successfully carried out typical fault diagnosis of nuclear power systems. In conclusion, by combining the interpretability of graph reasoning with the non-linear fitting ability driven by data, this study provides useful reference value for intelligent diagnostic models that integrate knowledge and data-driven approaches. The successful application of this method not only helps to promote technological innovation in the field of nuclear power, but also provides new insights for complex system fault diagnosis in other fields. 

Presenting Author: Xin Wang Harbin Engineering University

Presenting Author Biography: Currently pursuing a doctoral degree in Nuclear Science and Technology at Harbin Engineering University, majoring in fault diagnosis of nuclear power plants

Authors:

Xin Wang Harbin Engineering University
Minjun Peng Harbin Engineering University
Hang Wang Harbin Engineering University
Zikang Li Harbin Engineering University