データ拡張技術を用いた変位場およびひずみ場の予測精度向上に関する研究

抄録

In generating a surrogate model of CAE results using deep learning techniques, it is possible to predict the strain field using an automatic differentiation function from the displacement field. This function makes it possible to reduce the number of learning targets. It is necessary to improve the prediction accuracy of the displacement field to utilize this advantage. One of the causes of the decrease in the prediction accuracy of the displacement field is a property of imbalance in the distribution of nodal value used for learning. Improving prediction accuracy by increasing the learning data is expected by data augmentation techniques. Since the input data for learning CAE node information is a one-dimensional array, only the data augmentation method by adding noise can be adopted. However, the current data augmentation method of adding noise to one-dimensional learning data is less effective in addressing data imbalance, and a more effective solution is needed. We proposed a relative frequency equalized data augmentation method to solve this problem. In this paper, we examined the improvement of the prediction accuracy of the displacement field using the proposed method and predicted the strain field using the automatic differentiation function.