Online fault monitoring based on deep neural network & sliding window technique

Abstract

Nuclear power plants have proved their worth in energy sector by providing clean and uninterrupted power over decades. However, a Nuclear Power Plant (NPP) is a complex, dynamic system with potential radioactive release risk which makes it crucial to achieve highest standards of safety. Specially, in preview of massive monitoring data received in modern NPPs which makes it difficult for operators to extract vital information about actual plant state in a timely and accurate manner. On the other hand, advancements in latest machine learning methods have made it possible to process such massive data for operators to act accordingly. However, current machine learning approaches cited for this field, fall short of required capabilities needed for such safety critical industry. In manuscript, an online fault monitoring system is proposed which utilizes deep neural networks and sliding window technique. The proposed model not only fulfills the requirement of validity but also encompass all necessary diagnosis functions like detection, identification, assessment and robustness. The model allows for a fault to be identified and assessed in different plant states and then validate the predicted results through online correlation of simulation vs original data. The study was conducted for IP-200 NPP utilizing RELAP5 thermal-hydraulic code. The proposed model was verified by inducing 04 different faults for different states and severities. The results were found to be conducive for improving reliability and accuracy of next generation fault monitoring systems of Nuclear Power Plants.