目的関数のニューラルネットワーク指定による衝撃最適設計法の開発 : (第2報, 車両部品構造への適用)

概要

論文の詳細を見る
For the optimization problem of large structure, especially of crash worthiness analysis, finite element method is often used to calculate the structure response. However, the crash worthiness optimization problem is characterized by the non-linearity and noisy behavior (numerical inaccuracies) of response, it leads to tremendous computation cost and will fail to find the correct optimal solution when the sensitivity based optimization strategies are used. In this paper, we propose a new function approximation method based on response surface methodology, to smooth and simulate the structure response, by using Holographic Neural Network (HNN), to solve the optimal design. Sequential training scheme is applied initially to increase the accuracy of function approximation. Then we discuss the initial designs and the added designs selection to effectively reduce the analysis times (function calls). This optimization approach is illustrated for two numerical examples comparing with the Sequential Quadratic Programming (SQP) and shows an effective procedure of reducing the function calls. Finally, an industrial crash worthiness optimization design problem is carried out.
一般社団法人日本機械学会の論文
1998-10-25