文献类型：期刊文献

英文题名：Efficient JITL framework for nonlinear industrial chemical engineering soft sensing based on adaptive multi-branch variable scale integrated convolutional neural networks

作者：Chen, Yifan[1];Li, Anlan[1];Li, Xiangyang[2];Xue, Dong[1];Long, Jian[1,3]

机构：[1]East China Univ Sci & Technol, Minist Educ, Key Lab Smart Mfg Energy Chem Proc, Shanghai 200237, Peoples R China;[2]Sinopec Jinan Refining & Chem Co, Jinan 250101, Shandong, Peoples R China;[3]Qingyuan Innovat Lab, Quanzhou 362801, Peoples R China

年份：2023

卷号：58

外文期刊名：ADVANCED ENGINEERING INFORMATICS

收录：;EI(收录号：20234314969705);WOS:【SCI-EXPANDED(收录号:WOS:001092178900001)】；

基金：This work was supported by National Natural Science Fund for Distinguished Young Scholars (61925305) , National Natural Science Foundation of China (62373155, 62173147,61973124) , and Major Program of Qingyuan Innovation Laboratory (Grant No. 00122002) .

语种：英文

外文关键词：Soft sensor; Convolutional neural network; Chemical process; Prediction; Just-in-time Learning

摘要：Just-in-time Learning (JITL) is a soft measurement method commonly used in industrial processes, which can update local models in real-time to solve the problem of inaccurate parameter measurement caused by dynamic changes in real chemical processes. At present, the accuracy of the similarity measurement of JITL is insufficient. The modeling is time consuming when there is too much historical data. Considering the applicability of different chemical processes, the local model needs to be further designed. Therefore, this study aims to address these limitations, an efficient JITL Framework based on Adaptive multi-branch variable scale integrated convolutional neural networks (EJITL-AMVs-ICNN) is proposed. Firstly, a mixed similarity measurement method with feature factors (MSM-FF) is proposed and used as the distance measurement method in K-means cluster to achieve data binning. Then, the most similar training samples are selected in the corresponding bin by the adaptive sample length selection method. Finally, the local model is established through AMVs-ICNN, which can establish the corresponding network structure according to the number of features of different chemical processes. Two chemical process datasets, the ebullated bed residue hydrogenation and physical separation unit were used to verify that the proposed method has higher prediction accuracy and lower elapsed time.