摘要
在研究深度神经网络预测轧制力的基础上,针对极限学习机-自编码器的回归问题,提出一种自增删网络结构优化算法。利用自编码器进行原始数据的特征提取,为模型提供有效的高阶特征。极限学习机的学习速度快且泛化能力强,监督阶段时使用极限学习机回归轧制力,并使用隐层节点增删策略调节极限学习机的网络结构,解决了极限学习机-自编码器的结构设计问题。该方法用于轧制力回归,采用深层网络结合大量数据保证模型的回归精度的同时,实现了轧制数据的特征提取和网络结构的自增删。结果显示,该深层结构自增删网络具有很好的模型收敛和参数回归能力,在训练速度与精度方面均优于弹性RBF和稀疏自编码器神经网络算法。
On the basis of studying the prediction of rolling force by deep neural network, a self-addition and deletion network structure optimization algorithm was proposed for the regression problem of extreme learning machine-autoencoder. Then, the features of the original data were extracted by the autoencoder to provide effective high-order features for the model. However, the learning speed of the extreme learning machine was fast and the generalization ability was strong, and the rolling force was regressed by the extreme learning machine in the supervision stage. Therefore, the network structure of the extreme learning machine was adjusted by adding and deleting the hidden layer nodes to solve the structural design issues of the extreme learning machine-autoencoder. Furthermore, the method was applied to rolling force regression, and the regression accuracy of the model was ensured by the deep network combining with a large amount of data. At the same time, the feature extraction of the rolling data and the self-addition and deletion of the network structure were realized. The result proves that the self-addition and deletion network of the deep structure has good model convergence and parameter regression ability, and it is superior to flexible RBF and sparse autoencoder neural network algorithm in training speed and accuracy.
作者
张志强
尚猛
张宇
Zhang Zhiqiang;Shang Meng;Zhang Yu(Institute of Digital Engineering,Zhejiang Dongfang Polytechnic,Wenzhou 325000,China;Institute of Flight,College of Anyang Technology,Anyang 455000,China;SMS Siemag Technology(Beijing)Co.,Ltd.,Beijing 100020,China)
出处
《锻压技术》
CAS
CSCD
北大核心
2019年第12期192-197,共6页
Forging & Stamping Technology
基金
国家留学基金资助项目(201708410230)
关键词
带钢冷轧
轧制力
神经网络结构
极限学习机
自编码器
cold rolling of strip steel
rolling force
neural network structure
extreme learning machine
autoencoder
