轧机轧制力的改进训练策略深度神经网络预测  

Rolling Force Prediction Based on Improved Training Strategy Deep Neutral Network

作　　者：于飞[1] 于博 YU Fei;YU Bo(Department of Mechanical and Electrical Engineering,Liaoyuan Vocational and Technical College,Jilin Liaoyuan 136200,China;Changchun Institute of Technology,Mechanical and Electrical Engineering College,Jilin Changchun 130012,China)

机构地区：[1]辽源职业技术学院机电工程学院,吉林辽源136200 [2]长春工程学院机电工程学院,吉林长春130012

出　　处：《机械设计与制造》2023年第1期96-100,共5页Machinery Design & Manufacture

基　　金：吉林省教育厅“十三五”科学技术项目(JJKH20191248KJ)。

摘　　要：为了提高双机架炉卷轧机的轧制力预测精度,提出了具有快速而高效训练策略的深度神经网络预测方法。介绍了双机架炉卷轧机的工作原理,分析了轧制力影响参数。在深度神经网络基础上,使用随机小批量的样本选取法,提高深度神经网络训练速度;提出自适应矩估计梯度优化算法,用于解决传统训练方法陷入局部极值的问题,从而给出了改进训练策略的深度神经网络轧制力预测方法。经轧制实验验证,改进深度神经网络的训练时间为226.15s,而传统网络的训练时间为862.93s;改进网络的预测误差绝大部分控制在3%以内,而传统网络的预测误差绝大部分控制在5%以内。以上数据表明,改进深度神经网络的训练速度和预测精度均远优于传统深度神经网络。In order to improve double-stand Steckel mill rolling force prediction accuracy,prediction method based on improved training strategy deep neutral network is proposed. On the basis of deep neutral network,stochastic mini-batch gradient descent is used to improve training speed of deep neutral network. Adaptive moment estimation gradient optimization algorithm is put forward to solve the falling into local optimal problem of traditional training method,so that rolling force prediction method based on improved training strategy deep neutral network is provided. It is clarified by rolling experiment that training time of improved neutral network is 226.15s,and 862.93s of traditional deep neutral network. Almost all the prediction error is within 3% by improved network,and within 5% by traditional network. The above data indicates that training speed and prediction accuracy of improved deep neutral network are both optimal to traditional deep neutral network.

关 键 词：深度神经网络 轧制力预测 自适应矩估计梯度优化 随机小批量梯度下降法 

分 类 号：TH16[机械工程—机械制造及自动化] TP183[自动化与计算机技术—控制理论与控制工程]