基于动态注意力深度迁移网络的高炉铁水硅含量在线预测方法  

Online Prediction Method for Silicon Content of Molten Iron in Blast Furnace Based on Dynamic Attention Deep Transfer Network

作　　者：蒋珂 蒋朝辉[1,2] 谢永芳 潘冬[1] 桂卫华[1] JIANG Ke;JIANG Zhao-Hui;XIE Yong-Fang;PAN Dong;GUI Wei-Hua(School of Automation,Central South University,Changsha 410000;Peng Cheng Laboratory,Shenzhen 518000)

机构地区：[1]中南大学自动化学院,长沙410000 [2]鹏城实验室,深圳518000

出　　处：《自动化学报》2023年第5期949-963,共15页Acta Automatica Sinica

基　　金：国家自然科学基金(61773406,61725306,61290325);国家重大科研仪器研制项目(61927803);中南大学研究生自主探索创新项目(2021zzts0183);湖南省研究生科研创新项目(CX20210242)资助。

摘　　要：铁水硅含量是反映高炉冶炼过程中热状态变化的灵敏指示剂,但无法实时在线检测,造成铁水质量调控盲目.为此,提出一种基于动态注意力深度迁移网络(Attention deep transfer network, ADTNet)的高炉铁水硅含量在线预测方法.首先,针对传统深度网络静态建模思路无法准确描述过程变量与铁水硅含量之间的关系,提出一种基于注意力机制模块的输入过程变量与输出硅含量之间的动态关系描述方法;其次,为降低硅含量预测模型训练时对标签数据的依赖,考虑到铁水温度与硅含量数据之间的正相关性,利用小时级硅含量标签数据微调基于分钟级铁水温度数据预训练好的深度模型的结构,进而提高基于动态注意力深度迁移网络的硅含量预测精度;同时,为增强预测网络的可解释性,实时给出了基于动态注意力机制模块计算的每个样本各过程变量对铁水硅含量的贡献度;最后,基于某钢铁厂2号高炉的工业实验,验证了该方法的准确性、有效性和先进性.The molten iron silicon content,which can reflect the thermal state in blast furnace hearth during the ironmaking process,is difficult to detect in real time.This will cause blind adjustment to the quality of the molten iron.Hence,this paper proposes a data-driven model for the online prediction of the silicon content based on dynamic attention deep transfer network(ADTNet).First,considering that the deep network cannot accurately describe the relationship between process variables and the silicon content based on static modeling process,a dynamic attention module is designed to describe the dynamic relationship between the inputs and outputs.Then,to reduce the dependence of labeled silicon content samples during model training process,an online prediction model is established based on minute-level temperature data after considering the positive correlation between molten iron temperature and silicon content data.Subsequently,using the labeled silicon content data to fine tune the parameters of the well-trained deep model and improve the silicon content prediction performance based on dynamic attention deep transfer network.Furthermore,to enhance the interpret-ability of the deep black box model,the contribution of process variables of each sample to the silicon content is presented based on the dynamic attention module.Finally,industrial experiments of No.2 blast furnace in a steel plant verify the accuracy,effectiveness and advancement of the proposed method.

关 键 词：高炉炼铁 铁水硅含量 深度网络 迁移学习 动态注意力机制 预测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TF54[自动化与计算机技术—控制科学与工程]