A methodology for diagnosing FAC induced pipe thinning using accelerometers and deep learning models

Abstract

The flow accelerated corrosion (FAC) phenomenon is critical phenomenon that undermine the safety of nuclear power plant. The problems of FAC are not only induces pipe's thinning but occurs widely in nuclear power plants. Therefore, precise diagnosis of FAC phenomenon is important but the FAC phenomenon has a complicated mechanism. As a result, accurately diagnosing the FAC phenomenon is difficult (Kain, 2014). To overcome these difficulties, we proposed a methodology utilizing vibration data and deep learning to diagnose FAC induced thinning. To minimize the effect of outliers, Cook's distance and the Hilbert transform were used. When there was a significant difference in the pipes' thickness, the support vector machine (SVM), convolutional neural network (CNN), and long-short term memory (LSTM) network all showed good results. However, when the difference in pipes' thickness was subtle and thinning was non-uniform, only LSTM successfully diagnosed the pipe's condition.