Combinatorial Techniques for Fault Diagnosis in Nuclear Power Plants Based on Bayesian Neural Network and Simplified Bayesian Network-Artificial Neural Network

Abstract

Knowledge-driven and data-driven methods are the two representative categories of intelligent technologies used in fault diagnosis in nuclear power plants. Knowledge-driven methods have advantages in interpretability and robustness, while data-driven methods have better performance in ease of modeling and inference efficiency. Given the complementarity of the two methods, a combination of them is a worthwhile investigation. In this work, we introduce two new techniques based on Bayesian theory (knowledge-driven) and artificial neural network (data-driven) for fault diagnosis in nuclear power plants. The first approach exploits an integrated technique, Bayesian Neural Network (BNN), which introduces Bayesian theory into the neural network to provide confidence in diagnosis. The second approach, denoted as Simplified Bayesian Network-Artificial Neural Network (SBN-ANN), adopts a hierarchical diagnosis idea, which firstly uses a simplified Bayesian network to diagnose fault types and then a neural network to diagnose the severity of faults. The two new techniques are implemented and verified with simulated faults data of a typical pressurized water reactor. Compared with single-algorithmic diagnostic approaches such as Bayesian network and neural network, the new combinatorial techniques show better performance in diagnostic precision. The results suggest the feasibility to develop the data and knowledge dual-drive technologies for fault diagnosis.