摘要
为提高神经网络预测精度,利用模拟退火算法对基于k-均值聚类的RBF(radical basis function)神经网络进行了结构优化.首先,以NUMISHEET 02翼子板冲压成形为研究对象,以6条等效拉延筋力作为输入变量,基于Spearman相关分析和拉丁超立方抽样抽取相关性系数较小的数据作为SA-RBF(simulated annealing-RBF)神经网络的训练样本;其次,将训练样本进行Dynaform数值仿真,以起皱缺陷和拉裂缺陷建立的成形质量评价函数为目标函数,通过SA-RBF神经网络建立等效拉延筋力与目标函数间的非线性映射关系;再次,利用NSGA-II算法对其进行求解得到Pareto最前沿,通过灰色关联分析理论确定最佳拉延筋力;第三,利用优化的拉延筋力对翼子板成形进行数值仿真分析,成形极限图结果表明,优化后的成形件起皱显著减少,而且塑性变形更加均匀,提高了成形质量.
The structure of a radial basis function （RBF） neural network based on the ^-means clustering algorithm was optimised by employing the simulated annealing algorithm for improving the prediction accuracy. The NUMISHEET 02 fender was considered as the object of research and six equivalent drawbead forces were used as input variables. Based on Spearman correlation analysis and Latin hypercube sampling, the data which had smaller correlation coefficient values were chosen as training samples for the simulated annealing-radial basis function （ SA-RBF） neural network . The numerical simulations of training samples were performed by employing the Dynaform software package. The evaluation functions of forming quality were established based on the wrinkling defects and crack defects. The nonlinear relationship between the equivalent drawbead force and the associated objective function was established by incorporating a SA-RBF neural network . NSGA-II algorithm was employed to achieve the Pareto frontier and the best equivalent drawbead forces were determined by applying grey correlation analysis theory . Finally, the numerical simulation of fender forming was performed based on the optimised drawbead forces. The resultant forming limit diagram （FLD） indicates decreased wrinkles in the optimised forming part and greater uniformity in the plastic deformation , thereby leading to improvement in the quality of fender forming .
出处
《西南交通大学学报》
EI
CSCD
北大核心
2017年第5期970-976,993,共8页
Journal of Southwest Jiaotong University
基金
国家自然科学基金资助项目(51005193)
国家大学生创新创业训练计划项目(201710613033)
