Explainable machine learning to identify flaws in supporting structures of fast reactor

太田 陽生*; 神田 侑奈*; 久持 陸也*; 矢田 浩基; 古谷 正裕*

no journal, , 

Since the core temperature of SFRs can be much higher than that of LWRs, when the operation is prolonged, it is desirable to confirm no scratches and cracks in the core supporting structure of SFRs. However, sodium is chemically active and opaque. Furthermore, it is difficult to extract sodium in the reactor vessel for inspection. That is why the method accessed inside the reactor vessel is difficult to conduct. Non-destructive testing is one of the effective methods to detect welding defects inside the reactor. Ultrasonic testing (UT) can be applied to this detection without damaging specimens. UT was conducted to detect welding defects from the outer of the reactor vessel at this time. Therefore, the distance between the probe and the flaw is far. If there is a welding defect with noise, the flaw is too small, or the angle of the flaw is parallel to the direction in which the ultrasound is traveling, the welding defect is difficult to detect, even for skilled engineers. Machine learning (ML) is a valid method to resolve this problem. ML enables us to judge whether welding defects exist or not in supporting structures. In this study, we acquired PA images from 0.35 m specimens and confirmed validation accuracy for the classification of whether welding defects exist or not using ML, the basis of classification using explainable AI, and collected PA images from 1.0 m specimens.