文献类型：期刊文献

英文题名：Mutual information and attention-based variable selection for soft sensing of industrial processes

作者：Yu, Zhenhua[1];Wang, Guan[2];Yan, Xuefeng[1];Jiang, Qingchao[1];Cao, Zhixing[1]

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, State Key Lab Bioreactor Engn, Shanghai 200237, Peoples R China

年份：2025

卷号：146

外文期刊名：JOURNAL OF PROCESS CONTROL

收录：;EI(收录号：20250217670591);WOS:【SCI-EXPANDED(收录号:WOS:001397445600001)】；

基金：This work was supported in part by the National Natural Science Foundation of China under Grant 62322309, Shanghai Rising-Star Program under Grant 21QA1402400, Shanghai Pilot Program for Basic Research under Grant 22TQ1400100-16, and Shanghai Science and Technology Innovation Action Plan under Grant 23S41900500.

语种：英文

外文关键词：Soft sensing; Variable selection; Attention mechanisms; Process modeling

摘要：This study introduces a novel method called mutual information (MI) and attention-based variable selection (MAVS) to address the challenges of irrelevant and redundant variables in industrial process soft sensing while providing interpretability in variable contribution analysis. First, irrelevant variables are eliminated based on low MI values with the quality variable. Second, attention scores are used to remove redundant variables, and the false discovery rate is used to determine the number of beneficial variables. Finally, this work provides an interpretable and accurate contribution of the selected variables by using kernelSHAP, a kernel-based Shapley analysis. Unlike traditional approaches, MAVS integrates MI with attention mechanisms to optimize variable selection dynamically and adaptively. MAVS obtains stronger robustness and higher accuracy than the existing state-of-the-art models through optimal variable selection. The former also obtains better superior generalization than the latter through adaptive adjustment of attention weights. The superiority of MAVS is demonstrated using two real-world datasets and one simulated dataset.