A robust diagnosis method specifically for similar faults in nuclear power plant multi-systems based on data segmentation and stacked convolutional autoencoders

Abstract

Nuclear power plants consist of multiple subsystems characterized by nonlinear correlations, numerous monitoring parameters, and faults with similar characteristics. These similar faults often exhibit closely distributed data points, resulting in unclear classification boundaries and reduced diagnostic accuracy for fault detection algorithms. A single deep learning model cannot establish a diagnostic model that is broad in scope, highly accurate, and robust. This paper proposes a robust, multi-system similar fault diagnosis model for nuclear power plants, combining K-Means and Stacked Convolutional Autoencoders (KSCAE). First, K-Means data segmentation model partitions complex multi-system fault data into several similar data clusters. Then, specialized diagnostic models are developed for each cluster using powerful stacked convolutional autoencoders to focus on learning and classifying similar faults. Testing across three nuclear power plant fault diagnosis scenarios based on data from the Fuqing nuclear power plant simulator demonstrates that KSCAE outperforms single deep learning models in diagnostic accuracy, particularly when fault severity differs between training and testing sets. The results show that the KSCAE algorithm achieves a maximum accuracy of 99.31% and a minimum accuracy of 79.11% under severe noise and data distribution differences. Compared to the baseline algorithms, KSCAE achieves an average accuracy improvement of up to approximately 20% across multiple tests, particularly in Scenario 2. This study demonstrates the effectiveness and robustness of the proposed model for diagnosing similar faults, providing a reliable approach for multi-system fault diagnosis in nuclear power plants.