Human-cyber-physical system for operation in nuclear reactor possessing asymmetric multi-task learning-based predicting framework

Abstract

This work proposes a novel architecture of a human-cyber-physical system (HCPS) to operate nuclear reactors in the context of next-generation artificial intelligence (NGAI) technologies. Developing the generation IV nuclear power plant requires NGAI to fulfill intelligent operation and thereby enhance reliability and safety. Nevertheless, the NGAI based on machine computing is limited in auto-cognition whereas humans are adept in information processing. Operation states prediction is served as a specific scene to investigate the key enabling technologies (KETs) in HCPS, where humans are elevated to a supervisory level for NGAI technologies. Humans interact with NGAI in two-stage which comprise the initial construction and later upgradation of an intelligent prediction framework. In this paper, asymmetric multitask learning (AMTL) is elaborated as a KET which is essential to implementing intelligent prediction in HCPS. AMTL simultaneously trains multi-tasks and permits asymmetric information transfer between tasks. The amount of transfer varies with inter-task relatedness and individual task losses, thereby preventing negative transfer and enhancing accuracy. Furthermore, the improved AMTL is adapted for the small and imbalanced samples of tasks. A sparse and directed regularisation graph is introduced to visualise the correlation between tasks, reflecting the physical mechanism as well. Finally, the improved AMTL is utilized to predict reactivity of a specific pressurised water reactor (PWR), which validates the efficiency and accuracy.