Cross-Domain Few-Shot Anomaly Detection for equipment in nuclear power plants

Abstract

In Nuclear Power Plants (NPPs), operating data from equipment may shift due to changes in environmental conditions, device degradation, or component replacements. These shifts can impact the performance of data-driven monitoring models trained solely on source domain data, leading to increased false alarms and reducing both the effectiveness and reliability of the models. Furthermore, the amount of shifted data in real-time monitoring is limited and cannot meet the demands for deep learning model’s training process. To address the problems of Cross-Domain Few-Shot Anomaly Detection (CDFS-AD), we propose a Deep Temporal–Spatial Transfer Learning Network (DTSTLN). The proposed model leverages an improved transformer model to achieve temporal–spatial feature extraction and reconstruction of input operating data. And Maximum Mean Discrepancy (MMD) based loss function is utilized to achieve domain adaptation, enabling knowledge transfer and effective training with limited data. Comparative experiments on real operating data from the reactor coolant pump in NPPs demonstrate the effectiveness of DTSTLN in monitoring shifted data, as evidenced by higher F1-scores and lower False Alarm Rates (FARs) compared to other baseline methods, highlighting its potential for anomaly detection of NPP equipment in real scenarios.