Sensor Anomaly Detection in Nuclear Power Plant Using Deep LSTM Denoising Autoencoder and Isolation Forest

Abstract

Industrial health monitoring is essential for managing and maintaining infrastructures in a process industry where the primary goals are reducing downtime, improving health, and ensuring safety performance. On the contrary, unplanned downtimes caused by regular maintenance often result in financial losses. This scenario calls for automated fault diagnosis that facilitates online health monitoring to predict faults before irreversible damage occurs. This letter proposes a deep learning approach based on long short-term memory denoising autoencoder (LSTM-DAE) combined with Isolation Forest (IF) for early detection of sensor anomalies in nuclear power plants. Residual signals from LSTM-DAE are fed to IF to generate anomaly scores for early fault detection. The proposed approach is validated using the dataset obtained from KALBR-SIM, a full scope operator training replica simulator, which replicates the Prototype Fast Breeder Reactor at Indira Gandhi Centre for Atomic Research. Results demonstrate that the proposed approach detects faults much earlier than the state-of-the-art approaches, with an accuracy of 98.2%.