Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : November 16, 2022 - November 18, 2022
The structural strength evaluation of crash boxes is predicted by machine learning in this study. The training data was obtained from the dynamic elastic plastic analysis of the crash box. The input physical quantities are barrier angle, box thickness, material properties and mass equivalent to vehicle weight. The output physical quantity is the reaction force. Buckling occurs in the analysis and different directions of corruptions are one of the most interesting phenomenon from a point of engineering view. Physically meaningful features that take into account physical laws, physical properties, shape, and so on were added. As a result, we showed that learning by CNN is possible with higher accuracy. In addition, data design and data augmentation that takes physical phenomena into account are necessary to deal with large outlier. We would like to propose an adaptive method for machine learning in structural evaluation that can be used for a wide range of structural evaluations.
