摘要
为了解决动态变化过程数据随设备参数变化而服从不同分布的问题,提出了一种基于深层化工机理和自适应方法的多工况深层机理特征自适应(deep mechanism feature adaptation,DMFA)与深度学习的异常识别方法。首先,建立了工艺的物料衡算和热量衡算方程。其次,卷积操作抽取了工艺的机理特征表示。然后,自适应了源域(训练数据)和目标域(测试数据)的深层机理特征分布。最后,基于深度学习实现了无监督目标域的异常识别。由于DMFA在学习工艺机理时,使模型得到了紧致的机理特征表示,适配了工艺过程的设备参数变化,所以DMFA可以有效解决欠拟合、过拟合以及欠适配问题。脱丙烷精馏过程的应用表明,该方法能够有效识别精馏过程多工况,其异常识别的平均F1分数达到了99.87%。
A multi-conditions deep mechanism feature adaptation(DMFA)and deep learning based anomaly identification method was proposed for data treatment from dynamic process because its distributions depend on equipment parameters.The convolution operation extracted the mechanism characteristic representation of the process variables based on the material balance and heat balance equations of chemical process.Then the deep mechanism feature distribution adaptive for the source domain(training data)and the target domain(testing data)was realized,and deep learning-based anomaly identification was achieved for the unsupervised target domain.The process mechanism learnt by DMFA can represent mechanism feature compactly and adapt to the process parameter changes automatically,so DMFA can effectively address the under-fitting,over-fitting,and owing adaptation problems.The identification in the depropanization distillation process shows that the average F1 score of this method is 99.87%in identifying the distillation process anomaly under multi-conditions.
作者
贾旭清
杨霞
田文德
李传坤
刘福胜
罗忠军
王辉
JIA Xu-qing;YANG Xia;TIAN Wen-de;LI Chuan-kun;LIU Fu-sheng;LUO Zhong-jun;WANG Hui(College of Chemical Engineering,Qingdao University of Science&Technology,Qingdao 266042,China;State Key Laboratory of Safety and Control for Chemicals,SINOPEC Qingdao Research Institute of Safety Engineering,Qingdao 266071,China;Shandong Qiwangda Group Petrochemical Co.Ltd.,Linzi 255400,China)
出处
《高校化学工程学报》
EI
CAS
CSCD
北大核心
2020年第4期1026-1033,共8页
Journal of Chemical Engineering of Chinese Universities
基金
国家自然科学基金(21576143,21706291)
山东省重点研发计划(2018YFJH0802)。
关键词
机理知识
特征自适应
深度学习
化工过程
异常识别
mechanism knowledge
feature adaptation
deep learning
chemical process
anomaly identification
作者简介
贾旭清(1994-),男,山东烟台人,青岛科技大学硕士生;通讯联系人:田文德,E-mail:tianwd@qust.edu.cn。
