文献类型：期刊文献

英文题名：Research on a physics-informed and genetic algorithm-based vine copula soft sensor modeling method

作者：Wang, Zhenyuan[1];Li, Shaojun[1];Liu, Mandan[1]

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

年份：2026

卷号：187

外文期刊名：JOURNAL OF THE TAIWAN INSTITUTE OF CHEMICAL ENGINEERS

收录：;EI(收录号：20263221241768);Scopus(收录号：2-s2.0-105046382974);WOS:【SCI-EXPANDED(收录号:WOS:001841715500001)】；

基金：This work was supported by the National Natural Science Foundation of China (grant number 22478111) and Shanghai Natural Science Foundation (grant number 24ZR1415700) .

语种：英文

外文关键词：Soft sensor; Vine copula; Genetic algorithm; Physics-informed models

摘要：Background: Soft sensor is an effective tool for real-time monitoring of key variables in industrial processes. Vine copula has been introduced into soft sensor modeling because it can portray nonlinear, non-Gaussian and autocorrelated characteristics. However, traditional hierarchical parameter optimization in vine copula tends to converge to local optima. In addition, labeled samples are often limited in industrial applications, which leads to overfitting in soft sensor models. Methods: To address these issues, this study proposes a physics-informed D-vine copula soft sensor model optimized by genetic algorithm (PI-DVC-GA). The genetic algorithm (GA) globally optimizes all bivariate copula parameters in the D-vine structure to avoid local optima. Latin hypercube sampling (LHS) is used to generate unlabeled virtual samples. These samples are used to evaluate whether the model output predictions violate physics-informed constraints among process variables. The degree of constraint violations is quantified and incorporated into the fitness function as a penalty term. Physics-informed constraints are thus embedded in the global optimization, which mitigates overfitting under limited labeled samples. A numerical example, an acetylene hydrogenation process and a Tennessee Eastman (TE) process are employed to validate the effectiveness of the proposed method. Significant Findings: The results show that the PI-DVC-GA method achieves superior prediction accuracy compared to conventional soft sensors. The model provides both point prediction and prediction interval at various confidence levels, which enhances interpretability and utility.