复杂结构可靠性的神经网络分析方法

Research on ANN-based Reliability Analysis of Complex Structures

将整体结构按拓扑关系划分为若干模块,在位移协调和对接力协调条件下,根据力的传递原理,利用单元刚度矩阵法求解各模块结构单元,用整体节点平衡方程求解整体结构单元。将模块结构信息作为输入,整体结构信息作为输出,构造样本集,利用神经网络的非线性映射能力,通过样本训练提取两种结构模型的特征和本质联系,以获得具有泛化能力的网络结构,从而以模块结构替代整体结构的重分析。算例以6层整体结构(包含6个模块结构)建立了整体结构和模块结构在等效载荷作用下基于神经网络的对应法则;以网络外推能力分析了8层整体结构单元的失效模式;以网络内插能力分析了4层整体结构单元的失效模式;与优化准则法的对比结果验证了所提出方法的有效性。

Integer structure was divided into some module structures according to topological connection. Comparing with integer structure, module structure was easily analyzed because of briefness structure, and the computation speed was improved. Module structure was computed on element stiffness matrix based on load transfer behavior and displacement compatibility conditions. Integer structure was computed on node balanced equation. Thus information of module structure was gained. Module structure information was considered as input data and integer structure information was considered as output data. Thus specimen was built on input data and output data. Using the nonlinear mapping capability, neural network structure with interpolation and extrapolation was gained by training specimen. So integer structure recomputed can be replaced by module structure computed. Rules of correspondence between integer structure and module structure under equivalent load were built on integer structure with six module structures as an instance. Integer structure with eight module structures was analyzed by neural network extrapolation. Integer structure with four module structures was analyzed by neural network interpolation. All of designs have perfect computing precision contrasting with optimization criteria.