Fault Detection and Location Method for Nuclear Power Plant Main Systems Based on Low-Dimensional Manifolds

Abstract The nuclear power plants have characteristics such as variable operating conditions, many large and complex system, high-dimensional fault data, and complex nonlinearity. Classical fault diagnosis methods such as BP neural networks and support vector machines lack excellent feature extraction capabilities and have limited data learning abilities, making them unsuitable for multi-fault diagnosis tasks in complex, nonlinear nuclear power plants with multiple parameters. To address these issues, this study proposes a multi-fault diagnosis model for the main system of nuclear power plants based on low-dimensional manifolds. Following the concept of distributed diagnosis, the study divides the large and complex main system of the nuclear power plant into several subsystems, transforming the complex multi-system diagnosis task into simpler fault detection and localization tasks for each subsystem. Subsequently, within each segmented subsystem, considering the strong nonlinear coupling relationships among fault features and the difficulty in feature extraction, the study employs a stacked autoencoder method to reduce the dimensionality of fault data and uses the random forest method for classification to achieve fault detection and localization in each subsystem. Through simulation data testing, the fault diagnosis method developed in this study proves to be feasible, providing a reference for addressing the large-scale fault diagnosis challenges in complex nuclear power plants.