冷连轧轧制力深度神经网络模型泛化能力并行优化  

Parallel Optimization of Generalization Capability of Rolling Force Deep Neural Network Model in Tandem Cold Rolling Mill

作　　者：吴爽 闫奕 李爽 李峰 WU Shuang;YAN Yi;LI Shuang;LI Feng(School of Computer and Information Engineering,Nanyang Vocational College,He'nan Nanyang 474550,China;School of Mechanical and Automotive Engineering,Nanyang Vocational College,He'nan Nanyang 474550,China;School of Mechanical and Power Engineering,He'nan Polytechnic University,He'nan Jiaozuo 450000,China)

机构地区：[1]南阳职业学院计算机与信息工程学院,河南南阳474550 [2]南阳职业学院,机械与汽车工程学院,河南南阳474550 [3]河南理工大学机械与动力工程学院,河南焦作450000

出　　处：《机械设计与制造》2023年第8期171-174,共4页Machinery Design & Manufacture

基　　金：2021年河南省社会科学界联合会调研课题(SKL-2021-1513)。

摘　　要：为了更好调控冷连轧板厚参数,设计了一种冷连轧轧制力深度神经网络模型,增强了冷连轧模型的控制效果。选择2030冷连轧结构进行研究,对多输入多输出(MIMO)深度神经网络(DNN)进行预处理,针对多线程CPU与GPU实施了优化,对比了神经网络模型和冷连轧系统Siemens模型误差。研究结果表明:L-M算法表现出了更优的收敛稳定性、测试和验证性能、梯度下降趋势,并且收敛速度也更快。以随机方式选择200个数据并测定泛化性能测试得到,L-M算法获得了比SCG算法更大的相关系数。都是随着隐含层数的增加,获得了性能更优的神经网络模型,并且都会增加训练时间。从各项模型指标分析,L-M算法都比SCG算法的性能更优。构建神经网络轧制力模型总共包含二个隐含层、节点数介于17~30、通过L-M算法进行训练。采用神经网络轧制力模型得到的结果与实测值之间的误差比Siemens机理模型和测试值的误差更低。In order to better control the plate thickness parameters of tandem cold rolling,a neural network model of rollingforce depth in tandem cold rolling was designed to enhance the control effect of the model.The structure of 2030 tandem cold rolling was selected to study,MIMO deep neural network was preprocessed,multi-threaded CPU and GPU were optimized,and the errors of neural network model and Siemens model of tandem cold rolling system were compared.The results show that the L-M algorithm has better convergence stability,test and verification performance,gradient descent trend,and faster convergence speed.By randomly selecting 200 data and measuring the generalization performance test,L-M algorithm obtained a larger correlation coefficient than SCG algorithm.As the number of hidden layers increases,the neural network model with better performance is obtained,and the training time is increased.From the analysis of various model indexes,the performance of L-M algorithm is better than that of SCG algorithm.The neural network rollingforce model constructed in this paper consists of two hidden layers,the number of nodes is between 17 and 30,and is trained by L-M algorithm.The error between the results obtained by using the neural network rolling force model and the measured value is lower than that of Siemens mechanism model and the measured value.

关 键 词：深度神经网络模型 L-M算法 SCG算法 并行优化 轧制力模型 

分 类 号：TH16[机械工程—机械制造及自动化] TP273[自动化与计算机技术—检测技术与自动化装置]