転移学習による建屋内機器の地震時損傷判定に関する基礎的検討

抄録

Following a major earthquake leading to shut down of a nuclear reactor, it is crucial to rapidly assess the damage state of equipment to ensure a quick restart. In this study, we have developed a rapid method utilizing transfer learning to assess the degree of damage of components. The training images were made using capacity spectrum and demand spectrum. Only 80 images were used for this study. We tested four models (GoogleNet, ResNet, VGG16, and AlexNet), and have achieved estimation accuracy exceeding 90% for all models. Furthermore, a comparison with deep learning-based numerical estimation of ductility factor revealed that the transfer learning method is better than the numerical estimation by the deep learning. Therefore, we conclude that the transfer learning is effective for estimating seismic damage with a small amount of data. Moving forward, we intend to investigate the impact of training data volume and image data to the transfer learning.