(A) study on diagnosis of NPP's pipe thinning status by using accelerometer and machine learning algorithms

Abstract

The operating condition of secondary loop of nuclear power plant has the characteristics that are vulnerable to flow accelerated corrosion (FAC) phenomena. FAC phenomena induces wall-thinning effect and it may leads secondary loop tube rupture events. From 1970 to 2012, in the world, 1987 number of events were occurred because of FAC effect.

Nuclear power plant utilities try to estimate the FAC induced wall thinning effect by using CHECWORKS code [1]. However, the code analysis method requires the pipes' empirical result. In reality, extract the whole test result from secondary system is almost impossible.

To overcome this issue Kyung Ha Ryu [9] tried to estimate wall thinning by using equipotential switching direct current and Kwae Hwan Yoo [12] tried to estimate wall thinning by using infrared tomography. However, these methods requires additional complex equipment. Also, Jung Taek Kim [11] focused on the change of pipe's vibration characteristic due to wall thinning effect. To analyze vibration characteristic, Jung Taek Kim [11] used Fourier Transform (FT). However, pipes' vibration change is too tiny to be recognized from FT.

In this study, several machine-learning techniques (Support Vector Machine, Convolution Neural Network and Long-Short Term Memory Network) are used to estimate pipe's thinning condition. Pipes' vibration signal is used to train each machine learning techniques. As a results, for the simple classification problem (normal pipe, 1.0mm grinded pipe and 1.5mm grinded pipe classification), all three machine learning techniques (SVM, CNN, LSTM) shows good performance. However, for the date classification problem, only the LSTM net-work successfully classify and distinguish the pipe's thinning condition. From the result, by combining vibration data and LSTM network, pipe's thinning condition can be successfully diagnosed.