进化神经网络响应面法在结构可靠性中应用  

Evolutionary Neural Network Response Surfaces for Structural Reliability Analysis

作　　者：徐元铭[1] 陈丽华[1] 彭兴林[1] 

机构地区：[1]北京航空航天大学航空科学与工程学院,北京100191

出　　处：《飞机设计》2008年第5期33-38,共6页Aircraft Design

基　　金：航空科学基金(05B51043);国家自然科学基金(10572012)项目资助

摘　　要：针对传统可靠性分析方法的缺点,研究了网络拓扑结构和权值自适应调整的进化神经网络响应面,以用于实现隐式功能函数的全局映射。给出了基于遗传算法的神经网络响应面的构建步骤和样本点的选取方法。用训练好的神经网络代替可靠性分析中的有限元计算,大大提高了计算效率。通过工程实例与蒙特卡罗法、传统响应面法进行了对比分析,证明了该方法的有效性和实用性,为大型复杂结构的可靠性分析提供了一条高效的途径。To remedy flaws of traditional structural reliability analysis, the evolutionary neural network response surfaces （ENN）were proposed to project actual unknown performance functions. The networks are of self-adaptive topological structures and associated connection weights. And they can be trained to replace the finite element analysis during actual reliability analysis. The key issues of optimizing ENN＇s structure and weights by Genetic Algorithm（ GA）, as well as sampling by central composite design are detailed. The approach can greatly raise the computation efficiency. Through the practical example on engine compressive rotary table, comparisons of accuracies of ENN with Monto-Carlo and conventional response surfaces were given. And results concluded that ENN response surface offers an effective and less expensive approach for reliability analysis of large complex structures.

关 键 词：结构可靠性分析 响应面 进化神经网络 遗传算法 

分 类 号：V215[航空宇航科学与技术—航空宇航推进理论与工程] TP183[自动化与计算机技术—控制理论与控制工程]