Imbalanced sample distribution, with an abundance of normal samples and a scarcity of fault samples, along with uneven distribution among different types of faults, poses a challenge in analyzing operational data from nuclear power plants. Various methods have been proposed in current research to address this issue, including generative adversarial networks (GAN), under-sampling and over-sampling, and ensemble learning techniques. However, there lacks targeted research on the severity of imbalanced samples' impact on diagnostic models for nuclear power plants and a comprehensive performance comparison of various typical methods for handling imbalanced samples. This study focuses on typical approaches such as GAN, SMOTE over-sampling, and SMOTE-Boost ensemble learning, conducting simulations to assess the effects of imbalanced data, evaluate the performance differences, advantages, and disadvantages of these methods. Additionally, it proposes a novel imbalanced sample diagnostic method, CE-GAN-RF, incorporating Copula Entropy (CE) feature extraction module, GAN generation model, and random forest (RF) classification model, to offer new insights into imbalanced sample diagnosis techniques for nuclear power plants.
 

Generative Adversarial Networks; Copula Entropy; Imbalanced Sample Handling; Nuclear Power Plants