深層学習を用いた変位場予測における学習率の影響に関する検討

抄録

Evaluation of displacement and strain of structures and members is adequate for evaluating their soundness. Full-field measurement using digital images for this evaluation became widespread. On the other hand, the accuracy of the full-field measurement depends on the measurement environment and spatial resolution. Displacement field prediction using deep learning techniques can be a practical approach for alleviating this problem. To achieve this objective, we have developed a displacement field prediction method based on the theory of deep energy method. As a result of a previous study, the learning rate dramatically affects the improvement of prediction accuracy was confirmed. In this study, we investigated the effect of the learning rate on predicting the displacement field using the developed method. The initial and final values of the learning rate and the number of learning times, which are learning parameters, were adjusted. We discussed the prediction accuracy of the displacement field using these parameters.