Condition Monitoring of Nuclear Equipment-Piping Systems Subjected to Normal Operating Loads Using Deep Neural Networks

Abstract

Various fields in engineering explore the applicability of deep learning within condition monitoring. With the resurgence of nuclear energy due to electricity and carbon-free power generation demand, ensuring safe operations at nuclear power plants is important. Nuclear safety systems can undergo vibrations due to operating loads such as pump operations, flow-induced, etc. Safety equipment-piping systems experience degradation over the course of time due to flow-accelerated erosion and corrosion. Undetected degradation at certain locations can be subjected to a buildup of cyclic fatigue due to operational vibrations and thermal cycles. A condition monitoring framework is required to avoid fatigue cracking and for early detection of degraded locations along with the severity of degradation. This study aims to propose a condition monitoring methodology for nuclear equipment-piping subject to pump-induced vibrations during normal operations by designing a novel feature extraction technique, exploring parameters and developing a deep neural network, incorporating uncertainty in degradation severity, conducting a thorough investigation of predicted results to analyze erroneous predictions, and proposing strategic recommendations for “safe” pump operating speeds, as per ASME design criteria. Even with nondestructive testing, the detection of fatigue in pipes continues to be a difficult problem. Thus, this novel strategic recommendation to the operator can be beneficial in avoiding fatigue in piping systems due to pump-induced vibrations. The effectiveness of the proposed framework is demonstrated on a Z-piping system connected to an auxiliary pump from the Experimental Breeder Reactor II nuclear reactor and a high prediction accuracy is achieved.