进化神经网络在复合材料格栅结构设计中的应用

Application of evolutionary neural networks to grid-stiffened composite structure design

根据Kolmogorov多层神经网络映射存在定理,利用进化神经网络来实现结构设计参数(输入)与结构响应参数(输出)的全局非线性映射关系,以此来代替实际结构优化过程中存在的大量有限元计算,从而提高优化效率。以遗传算法为优化求解器,神经网络屈曲稳定性响应面为主要约束,对复合材料格栅加筋结构的优化问题进行了分析研究。算例表明,在相同(有限元)样本数据的情况下,进化神经网络通过自适应调节网络结构和权值,可获得比BP神经网络更高精度的映射模型,具有很强的泛化能力。该方法可为解决大型复合材料结构优化设计提供一条高效途径。

Based on Kolmogorov theorem, the global nonlinear mapping relationship between structural design parameters （input） and structural response parameters（output） was realized by using evolutionary neural networks （ENN）, which can replace massive finite element calculation during actual optimization process so as to improve optimization efficiency. Taking genetic algorithm （GA） as the optimization procedure and the neural network buckling response surface as main constraints, the optimal design of grid-stiffened composite panel under axial compressive loads was investigated. The results show that with the same FEM sample data, evolutionary neural networks can get more accurate mapping model than traditional BP neural network through self-adaptive adjustment grid structure and weight value. The ENN-GA algorithm provides an efficient approach to the structure optimization design of large complex composite.