复合泡沫塑料力学行为模拟的ANN方法  

Method for simulating mechanical behavior of syntactic foam plastics by artificial neural networks

作　　者：邹波[1] 卢子兴[1] 

机构地区：[1]北京航空航天大学航空科学与工程学院,北京100083

出　　处：《北京航空航天大学学报》2007年第7期860-864,共5页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金资助项目(10572013);国家自然科学联合基金资助项目(NASF10276004)

摘　　要：对人工神经网络(ANN)方法在复合泡沫塑料力学行为模拟中的应用进行了研究.首先,选取影响材料力学行为的因素和所需模拟、预测的力学性能作为输入、输出量;然后,利用反向传播算法建立了四层神经网络模型,对复合泡沫塑料的力学性能和本构关系进行了模拟和预测.数值结果表明,训练后的神经网络模型能较好地模拟、预测材料的模量、屈服强度和不同应变率及不同温度下的压缩应力-应变曲线.此外,3种不同改进训练方法的比较说明,Bayesian规则化法的泛化能力最好,LM法收敛最快,而自适应梯度下降动量法则需要较长的迭代时间才能达到相同的精度.Application of artificial neural networks （ANN） method on the mechanical behavior simulation of syntactic foam plastics was discussed. Firstly, factors influencing on the mechanical behavior and mechanical properties simulated and predicted were separately taken as input and output quantities. Secondly, Fourlayer neural networks model was established to simulate and predict the mechanical properties and constitutive relationship of syntactic foam plastics by means of back-propagation algorithm. The numerical results show that the trained ANN model can preferably simulate and predict the mechanical behavior of material, such as Young＇s modulus, yield strength and stress-strain curves under different strain rates or temperatures. Additionally, by comparison among three different modified training methods, it is found that Bayesian regularization back-propagation has the best capacity of improving network generalization, Levenberg-Marquardt（LM） backpropagation would converge fastest, and gradient descent momentum ＆ adaptive learning rate back-propagation need long-end iterative process before the same precision in calculation is achieved.

关 键 词：人工神经网络 复合泡沫塑料 力学性能 

分 类 号：O341[理学—固体力学]