文献类型：期刊文献

英文题名：A quality-driven multi-attribute channel hybrid neural network for soft sensing in refining processes

作者：Li, Zhi[1];Xue, Kaige[1];Chen, Junfeng[1];Peng, Xin[1,2]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Engn Res Ctr Proc Syst Engn, Minist Educ, Shanghai 200237, Peoples R China

年份：2025

卷号：250

外文期刊名：MEASUREMENT

收录：;EI(收录号：20251017990172);WOS:【SCI-EXPANDED(收录号:WOS:001440678800001)】；

基金：This work was supported by National Key Research and Devel-opment Program of China (2023YFB3307800) , the Natural Science Foundation of Shanghai under Grant 24ZR1415900, Major Science and Technology Project of Xinjiang (No. 2022A01006-4) , State Key Labora-tory of Industrial Control Technology, China (Grant No. ICT2024A26, ICT2024A23) and Fundamental Research Funds for the Central Univer-sities.

语种：英文

外文关键词：Convolutional neural network; Long short-term memory network; Quality-driven model; Multi-attribute channel; Refining processes; Soft sensor

摘要：Soft sensing technology has been widely applied in quality prediction for refining production. The refining process involves multiple strongly coupled subsystems, resulting in process variables exhibiting localized patterns with different properties. Existing soft sensing methods often struggle to extract multi-attribute features from such complex coupled process data effectively. This paper proposes a multi-attribute channel hybrid neural network, integrated with a novel quality-driven learning mechanism to enhance the quality relevance of feature representation. Specifically, a multi-attribute channel feature learning module is designed to capture local spatio-temporal dependencies from process variables with varying properties, and the outputs from each channel are integrated to obtain a global feature representation of the variables. Additionally, a quality information extraction module is developed to account for fluctuations in quality data. This module employs wavelet transforms to analyze the overall trends and detailed information of quality variables, enabling the long short-term memory network to more effectively capture dynamic historical information. Finally, a feature interaction mechanism is introduced, where the results of the quality information extraction guide the feature learning of process variables to obtain more quality-relevant feature representations. The performance evaluation on two typical refining processes demonstrates the superiority of the proposed method compared to other modeling methods.