基于改进堆叠自编码器的带钢力学性能预报模型  

Prediction model of mechanical properties of hot rolled strip based on improved stacked self-encoder

作　　者：宋勇[1,2] 李博 刘超[1,2] 李飞飞[1,2] SONG Yong;LI Bo;LIU Chao;LI Fei-fei(National Engineering Research Center for Advanced Rolling Technology,University of Science and Technology Beijing,Beijing 100083,China;National Engineering Technology Research Center of Flat Rolling Equipment,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]北京科技大学高效轧制国家工程研究中心,北京100083 [2]北京科技大学国家板带生产先进装备工程技术研究中心,北京100083

出　　处：《冶金自动化》2020年第6期2-10,16,共10页Metallurgical Industry Automation

基　　金：国家自然科学基金资助项目(51674028);广西创新驱动发展专项资金计划资助项目(GKAA17202008)。

摘　　要：钢铁工业在智能制造转型升级过程中对产品性能预报技术提出了越来越高的要求。针对目前带钢力学性能预报模型普遍存在精度和适应性不高的问题,提出了一种基于改进堆叠自编码器的力学性能预报深度学习模型。结合CSP热连轧实际工艺流程,模型中设置多个分别代表不同工序的自编码器进行堆叠,同时将各工序的过程参数逐步输入对应的自编码器,实现对带钢组织演变过程的数据建模,并利用降噪自编码器(DAE)和稀疏自编码器(SAE)解决数据噪声大、过拟合等问题,提高力学性能预测精度,改善模型的适应性。结果显示,这种基于堆叠自编码器的力学性能预报模型综合性能更好。The iron and steel industry has put forward higher and higher requirements for product mechanical properties prediction technology in the process of intelligent manufacturing transformation and upgrading.Aiming at the problem that the current prediction model of mechanical properties of strip steel is generally not accurate and adaptable,a deep learning model of mechanical properties prediction based on improved stacked self-encoder is proposed.Combined with the actual process of CSP hot rolling,a plurality of self-encoders representing different processes were stacked in the model,and the processing parameters of each process were gradually input into the corresponding self-encoder to realize the data construction of the strip steel microstructure evolution process.The module used the noise reduction self-encoder(DAE)and the sparse self-encoder(SAE)to solve the problems of large data noise and over-fitting,improve the prediction accuracy of the mechanical properties and the adaptability of the model.The results show that the mechanical properties prediction model based on stacked self-encoder has better comprehensive performance.

关 键 词：热轧带钢 力学性能预报 堆叠自编码器 逐层特征提取 深度学习 

分 类 号：TG142.1[一般工业技术—材料科学与工程] TP18[金属学及工艺—金属材料]