Research on sensor anomaly detection and fault location in nuclear power plants based on GAN-IAAKR model

To address the issues of accurate anomaly detection under minor drift and accuracy degradation faults in nuclear power plant sensors, as well as the lack of data directionality, model robustness, and real-time performance in fault localization algorithms for nuclear power sensors. A GAN-IAAKR-based method for sensor anomaly detection and fault localization is proposed in this paper to accurately and promptly detect anomalous states in nuclear power plant sensors during faults and to perform fault localization for the faulty sensors First, a time-series two-dimensionalization method is employed to encode normal sensor monitoring data into images. Next, evolutionary game theory is utilized, where multiple generators and discriminators are organized into two distinct groups to simultaneously engage in the game. During the game, a replicator dynamic equation is established, and the weights of each generator and discriminator are dynamically adjusted through a smoothing mechanism. The encoded images are then fed into an enhanced GAN model for training. Building on this, the GAN model reconstructs the images, and anomaly detection is carried out by calculating the MSE statistic (reconstruction error) between the original and the reconstructed images. Finally, the improved AAKR algorithm is applied to reconstruct the sensor monitoring data. By setting a reconstruction error threshold, normal and anomalous data can be effectively differentiated, thereby enabling fault localization. Experiments were conducted to compare anomaly detection performance between One-Class SVM, Autoencoder, and GAN models under conditions of minor drift and accuracy degradation faults. The experimental results demonstrate that the proposed anomaly detection model exhibits a high detection accuracy, exceeding 95%, while the fault localization models also show superior accuracy, real-time performance, and robustness, with an accuracy rate exceeding 90%.