基于半监督集成即时学习的橡胶混炼过程门尼黏度软测量研究  

Study on semi-supervised ensemble just-in-time learning based soft sensing of Mooney viscosity in rubber mixing process

作　　者：金怀平 张燕[1] 董守龙 杨彪[1] 钱斌[1] 陈祥光 JIN Huai-ping;ZHANG Yan;DONG Shou-long;YANG Biao;QIAN Bin;CHEN Xiang-guang(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;School of Chemistry and Chemical Engineering,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500 [2]北京理工大学化学与化工学院,北京100081

出　　处：《高校化学工程学报》2022年第4期586-596,共11页Journal of Chemical Engineering of Chinese Universities

基　　金：国家自然科学基金(62163019,61763020,61863020);云南省应用基础研究计划(202101AT070096)。

摘　　要：针对工业橡胶混炼过程中门尼黏度标记数据有限,导致模型预测性能受限的问题,提出了一种半监督(SS)集成即时学习(EJIT)高斯过程回归(GPR)软测量方法,称为SSEJITGPR。当查询样本到来时,该方法通过在线迭代学习的方式获取高置信度伪标记样本,其中使用集成后的即时学习高斯过程回归(JITGPR)模型对非标记样本进行预测,并以集成预测方差作为置信度评价准则。随后,基于伪标记样本扩充后的建模数据库构建多样性的半监督JITGPR基模型。最后,采用有限混合机制实现基模型的自适应集成。与传统门尼黏度软测量方法相比,SSEJITGPR在处理局部过程特征、克服标记样本不足、预测可靠度不高等问题上表现出显著优势,其有效性和优越性通过工业案例进行了验证。Traditional soft sensors for Mooney viscosity estimation in industrial rubber mixing process often encounter the scarcity of labeled data,thus leading to great difficulties in obtaining accurate estimations.Therefore,a semi-supervised(SS)ensemble just-in-time(EJIT)learning based Gaussian process regression(GPR)method referred to as SSEJITGPR was proposed.When a query sample comes,a set of diverse just-in-time Gaussian process regression(JITGPR)base models is constructed and combined to predict unlabeled samples for providing high-confidence pseudo-labeled samples through iterative learning,where the ensemble prediction variance is used for confidence evaluation.Then,a group of diverse semi-supervised JITGPR base models is built from the modeling database extended by the selected pseudo-labeled data.Finally,a finite mixture mechanism is used to realize the adaptive combination of the base models.Compared with the traditional methods for Mooney viscosity estimation,SSEJITGPR showed significant advantages in dealing with local process characteristics,overcoming scarcity of labeled data and low prediction reliability.The effectiveness and superiority of SSEJITGPR has been verified by an industrial application case.

关 键 词：软测量 即时学习 半监督学习 集成学习 高斯过程回归 门尼黏度 橡胶混炼 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]