基于BP神经网络和遗传算法的并行迭代优化研究
|
摘要
提出了一种将反向传播神经网络(BPNN)与遗传算法(GA)相结合的迭代优化方法。该方法首先使用较少的学习样本开发和训练BPNN模型,然后利用GA在可行域内求解训练好的BPNN模型以寻找模型的最优解。将基于该最优解进行验证的结果作为新样本添加到训练模式集中对BPNN模型重新训练。针对训练模式较少可能导致预测精度不足的问题,在BPNN模型的训练中分别采用贝叶斯正则化算法(BRA)和改进的Levenberg-Marquardt算法提高BPNN模型的泛化能力和收敛性,并结合精英策略将模拟退火算法(SAA)嵌入到GA中,以提高BPNN模型的局部搜索能力。将所提出的方法应用于汽车吹塑成型的聚丙烯波纹管的厚度优化。结果表明,在3次迭代后可以得到最优模具间隙,采用最佳模具间隙的波纹管在9个齿峰处的厚度落入期望的范围内(0.7±0.05mm),并且材料的使用量减少了22%。这种最佳间隙仅通过23次实验即可获得,远远少于实际模塑过程所需的实验次数。
An iterative optimization method which combines back-propagation neural network(BPNN)with genetic algorithm(GA)was proposed.Firstly,the BPNN model was developed and trained with fewer learning samples,and then the trained BPNN model was solved by GA in the feasible region to find the optimal solution of the model.The validation result based on the optimal solution was added to the training pattern set as a new sample to retrain the BPNN model.Aiming at the problem that less training modes may lead to inadequate prediction accuracy,Bayesian regularization algorithm(BRA)and improved Levenberg-Marquardt algorithm were adopted to improve the generalization ability and convergence of BPNN model in training BPNN model,and elite strategy was combined to embed simulated annealing algorithm(SAA)into GA to improve the local search ability of the high BPNN model.The proposed method was applied to the thickness optimization of polypropylene bellows for automobile blow molding.The results show that the optimum mold clearance can be obtained after three iterations.The thickness of bellows with the optimum mold clearance at the nine tooth peaks falls within the expected range(0.7 +0.05mm),and the material usage is reduced by 22%.The optimum clearance can be obtained by only 23 experiments,which is far less than the number of experiments needed in the actual molding process.
An iterative optimization method which combines back-propagation neural network(BPNN)with genetic algorithm(GA)was proposed.Firstly,the BPNN model was developed and trained with fewer learning samples,and then the trained BPNN model was solved by GA in the feasible region to find the optimal solution of the model.The validation result based on the optimal solution was added to the training pattern set as a new sample to retrain the BPNN model.Aiming at the problem that less training modes may lead to inadequate prediction accuracy,Bayesian regularization algorithm(BRA)and improved Levenberg-Marquardt algorithm were adopted to improve the generalization ability and convergence of BPNN model in training BPNN model,and elite strategy was combined to embed simulated annealing algorithm(SAA)into GA to improve the local search ability of the high BPNN model.The proposed method was applied to the thickness optimization of polypropylene bellows for automobile blow molding.The results show that the optimum mold clearance can be obtained after three iterations.The thickness of bellows with the optimum mold clearance at the nine tooth peaks falls within the expected range(0.7 +0.05mm),and the material usage is reduced by 22%.The optimum clearance can be obtained by only 23 experiments,which is far less than the number of experiments needed in the actual molding process.
引文
[1]邢铁辉,张元位,李倩.浅谈涉外工程化学水处理系统设计优化[J].电站系统工程,2018,34(3):81-82.
[2]刘帅,康建邨.大港垃圾焚烧工程控制二恶英排放的优化设计[J].再生资源与循环经济,2017,10(6):37-39.
[3]徐栋,蔡丽韫,李世龙,等.中亚天然气管道工程优化设计[J].油气储运,2016,35(8):882-887.
[4]刘漾,姜波,马晓元,等.上海某220kV地下变电站基坑工程优化设计[J].结构工程师,2015,31(4):199-205.
[5]顾雪,徐进力,王琦,等.响应面分析法在HR-ICP-MS仪器参数优化中的应用[J].物探化探计算技术,2017,39(6):841-847.
[6]林淼.一种改进的高铁桥沉降变形预报方法研究[J].城市勘测,2017(6):135-138.
[7]刘辉元,马金辉,黄琼.基于改进克里金插值的室内定位位置指纹库构建方法[J].重庆邮电大学学报(自然科学版),2017,29(6):751-757.
[8]刘彬,肖守讷,朱涛,等.自由外翻式组合吸能装置的设计及优化[J].中国机械工程,2018,29(17):2052-2059.
[9]陈士华,马益平,方健美.遗传算法在智能配电网故障定位应用分析[J].能源与环保,2017,39(12):219-222.
[10]刘吉超,陈阳舟.基于GA-PSO混合优化的BPNN车速预测方法[J].交通运输系统工程与信息,2017,17(6):40-47.
[11]段明旭,邱冬炜,李婉,等.改进灰色人工神经网络模型的超高层建筑变形预测[J].测绘科学,2017,42(4):141-146,183.
[12]南敬昌,胡婷婷,盛爽爽,等.Doherty功放的贝叶斯正则化神经网络逆向建模研究[J].计算机工程与科学,2018,40(8):1496-1502.
[13]刘志敏,杜守强,王瑞莹.求解线性互补问题的Levenberg-Marquardt型算法[J].应用数学学报,2018,41(3):403-419.
[14]姜东,唐秋华,李梓响,等.多目标模拟退火算法求解混装线平衡与排序[J].机械设计与制造,2018(9):189-192.
[15]黄瑾,张涛,郭阳,等.基于精英策略的多目标规划问题的粒子群算法[J].长江大学学报(自科版),2018,15(13):1-6,101.
[16]高俊龙,袁如意,易建强,等.基于一型模糊规则自主构建二型TSK神经模糊系统方法设计[J].控制理论与应用,2016,33(12):1614-1629.
[17]王英晓,杜守强.求解一类广义随机线性互补问题的投影Levenberg-Marquardt方法[J].上海工程技术大学学报,2017,31(4):326-331.
[18]邱建成.多重壁中空吹塑成型技术[J].塑料包装,2017,27(6):49-50,61.
[2]刘帅,康建邨.大港垃圾焚烧工程控制二恶英排放的优化设计[J].再生资源与循环经济,2017,10(6):37-39.
[3]徐栋,蔡丽韫,李世龙,等.中亚天然气管道工程优化设计[J].油气储运,2016,35(8):882-887.
[4]刘漾,姜波,马晓元,等.上海某220kV地下变电站基坑工程优化设计[J].结构工程师,2015,31(4):199-205.
[5]顾雪,徐进力,王琦,等.响应面分析法在HR-ICP-MS仪器参数优化中的应用[J].物探化探计算技术,2017,39(6):841-847.
[6]林淼.一种改进的高铁桥沉降变形预报方法研究[J].城市勘测,2017(6):135-138.
[7]刘辉元,马金辉,黄琼.基于改进克里金插值的室内定位位置指纹库构建方法[J].重庆邮电大学学报(自然科学版),2017,29(6):751-757.
[8]刘彬,肖守讷,朱涛,等.自由外翻式组合吸能装置的设计及优化[J].中国机械工程,2018,29(17):2052-2059.
[9]陈士华,马益平,方健美.遗传算法在智能配电网故障定位应用分析[J].能源与环保,2017,39(12):219-222.
[10]刘吉超,陈阳舟.基于GA-PSO混合优化的BPNN车速预测方法[J].交通运输系统工程与信息,2017,17(6):40-47.
[11]段明旭,邱冬炜,李婉,等.改进灰色人工神经网络模型的超高层建筑变形预测[J].测绘科学,2017,42(4):141-146,183.
[12]南敬昌,胡婷婷,盛爽爽,等.Doherty功放的贝叶斯正则化神经网络逆向建模研究[J].计算机工程与科学,2018,40(8):1496-1502.
[13]刘志敏,杜守强,王瑞莹.求解线性互补问题的Levenberg-Marquardt型算法[J].应用数学学报,2018,41(3):403-419.
[14]姜东,唐秋华,李梓响,等.多目标模拟退火算法求解混装线平衡与排序[J].机械设计与制造,2018(9):189-192.
[15]黄瑾,张涛,郭阳,等.基于精英策略的多目标规划问题的粒子群算法[J].长江大学学报(自科版),2018,15(13):1-6,101.
[16]高俊龙,袁如意,易建强,等.基于一型模糊规则自主构建二型TSK神经模糊系统方法设计[J].控制理论与应用,2016,33(12):1614-1629.
[17]王英晓,杜守强.求解一类广义随机线性互补问题的投影Levenberg-Marquardt方法[J].上海工程技术大学学报,2017,31(4):326-331.
[18]邱建成.多重壁中空吹塑成型技术[J].塑料包装,2017,27(6):49-50,61.