深度神经网络轧制力建模及其并行优化研究  

Research on Rolling Force Modeling and Parallel Optimization of Deep Neural Network

作　　者：刘翰培 汪宇轩 王亚琴 罗小川[3] LIU han-pei;WANG Yu-xuan;WANG Ya-qin;LUO Xiao-chuan(School of Computer Science and Engineering,Northeastern University,Shenyang 110819,China;School of Materials Science and Engineering,Northeastern University,Shenyang 110819,China;School of Information Science and Engineering,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学计算机科学与工程学院,辽宁沈阳110819 [2]材料科学与工程学院,辽宁沈阳110819 [3]东北大学信息科学与工程学院,辽宁沈阳110819

出　　处：《控制工程》2022年第8期1379-1386,共8页Control Engineering of China

基　　金：2019年国家重点研发计划项目(2019YFB1705002)。

摘　　要：冷连轧过程控制的轧制力模型是整个轧制过程计算机控制的基础。为提高5机架2030冷连轧系统轧制力模型的精度和适用性,提出了多输入多输出深度神经网络轧制力模型的数据预处理、建模和并行优化方法。对含有不同隐含层数和节点数的神经网络,采用不同训练算法(SCG算法和L-M算法)与不同优化方法(多线程CPU、单GPU和多线程CPU+GPU),研究了神经网络结构、训练算法和优化方法对神经网络轧制力模型的性能、训练时长、线性相关系数的影响。研究结果表明:含有2个隐含层、采用L-M算法和多线程CPU优化方法可获得综合性能最优的神经网络轧制力模型;神经网络轧制力模型的计算误差远小于在线使用的Siemens轧制力模型的计算误差。The rolling force model of cold continuous rolling process control is the basis of computer control for the entire rolling process.In order to improve the accuracy and applicability of rolling force model of 5-stand 2030 cold continuous rolling system,methods for data preprocessing,modeling and parallel optimization of rolling force model of deep neural network with multi-input and multi-output are proposed.For neural networks with different numbers of hidden layers and nodes,different training algorithms(SCG algorithm and L-M algorithm) and different optimization methods(multi-threaded CPU,single-GPU and multi-threaded CPU+GPU) are used to study the effects of the neural network structure,training algorithm and optimization method on performance,training duration,linear correlation coefficient of the neural network model.The results show that the rolling force model of neural network with two hidden layers,trained by L-M algorithm and optimized by multi-thread CPU,can obtain the best comprehensive performance.The calculation error of the rolling force model of neural network is much smaller than that of Siemens rolling force model used online.

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

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]