Few-Shot Learning for Abnormal Event Detection in Nuclear Power Plants

Abstract Recent research on abnormal event detection in nuclear power plants (NPPs) increasingly adopts the data-driven method, aiming to identify anomalies by harnessing vast operational data without needing prior knowledge of plant system dynamics. However, the limited availability of abnormal samples within NPPs affects the detection accuracy of data-driven methods. This paper introduces a few-shot learning approach based on neural networks to develop an efficient anomalous event detector using sparse samples within NPPs. To achieve this, a neural network structure is employed for operational feature extraction to facilitate the transformation of input data into lower-dimensional vectors with embedded space. Then, a prototypical network is introduced to derive a prototype for each abnormal event class based on a few embedded vector training samples. In this way, the anomaly type in a new instance can be identified by comparing the similarity of its embedded vector to each prototype. Experimental results illustrate that the proposed approach achieves high abnormal event detection accuracy.