Abstract: Aiming at the low global precision and uncertain sample size for static surrogate model in sheet metal forming optimization,a dynamic Genetic Algorithm Back Propagation neural network (GABP) modeling method that used multi-stage supplementing samples to gradually improve the prediction accuracy of the surrogate model was proposed.The method increased the global precision by adding the sample according to maximum minimum distance criterion,and improved the local precision by supplementing the optimized solution as a sample,which could achieve a reduction in sample size and an improvement in computational efficiency.Based on the grey system theory,a new iterative formulation of grey relativity was derived.The multi-objective problem was transformed into single-objective optimization with maximum grey relativity,and the optimization framework of joint genetic algorithm and dynamic GABP model was established.The function example showed that the dynamic equivalent could cut down the sample size by about 20% and had better prediction accuracy by comparing with the static GABP model,and the new grey relativity formulation could promote the stable convergence of optimization iteration.The framework was used to conduct the forming process optimization of NUMISHEET 93 square box.Compared with the initial design,the thinning index and the wrinkle index of the optimized design was reduced by 16.62% and 8.26%respectively,which effectively improved the forming quality of the square box parts.

Key words: deep drawing, dynamic genetic algorithm back propagation neural network model, grey relativity, multi-objective optimization

摘要： 针对板料成形优化中构造静态代理模型过度设计样本量、预测精度不易控制等问题,提出分步增补样本逐渐提高代理模型预测精度的动态遗传BP神经网络(GABP)建模方法。该方法依据模型精度按最大最小距离准则增添样本来提高全局精度,根据优化解精度将优化解增补为样本以改进局部精度,从而减少样本量,提高计算效率;基于灰色系统理论推导了灰色关联度的迭代计算新格式,将多目标问题转化为最大化关联度的单目标优化,并建立联合遗传算法与动态遗传BP神经网络模型的优化流程框架。通过函数算例表明,与静态遗传BP神经网络模型相比,动态遗传BP神经网络模型能减少约20%的样本量,且预测精度更好,关联度迭代新格式实现了迭代过程平稳收敛;采用该优化流程完成了NUMISHEET 93方盒件的成形工艺优化,与初始设计方案相比,优化方案的减薄指标和起皱指标分别降低了16.62%和8.26%,有效改善了方盒件的成形质量。

关键词: 拉深成形, 动态遗传BP神经网络模型, 灰色关联度, 多目标优化