文献类型：期刊文献

英文题名：Enhanced meta-learning based multi-scale fusion network for silicone monomer synthesis process prediction under data sparsity

作者：Li, Zhongmei[1,2,3];Sun, Qinhao[2];Zhang, Bing[2,4];Lou, Jionghao[2];Wu, Shiming[2];Xiao, Jiyang[2];Feng, Enbo[2];Du, Wenli[2]

机构：[1]East China Univ Sci & Technol, State Key Lab Ind Control Technol, Shanghai, Peoples R China;[2]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai, Peoples R China;[3]Huzhou Inst Ind Control Technol, Huzhou 313099, Peoples R China;[4]Yunnan Nengtou Silicon Mat Technol Dev Co, Jinjiang, Peoples R China

年份：2026

卷号：321

外文期刊名：CHEMICAL ENGINEERING SCIENCE

收录：;EI(收录号：20254419436631);WOS:【SCI-EXPANDED(收录号:WOS:001621422700001)】；

基金：This work was supported by National Natural Science Foundation of China (62394343, 62394345) , Shanghai Rising-Star Program (24QA2706100) 'and the Programme of Introducing Talents of Discipline to Universities (the 111 Project) under Grant B17017 and the Fundamental Research Funds for the Central Universities and the State Key Laboratory of Industrial Control Technology, China (ICT2024A01) .

语种：英文

外文关键词：Industrial informatics; Meta-learning; Few-shot learning; Silicone monomer synthesis; Time series prediction

摘要：Accurate prediction of Dimethyldichlorosilane (M2) selectivity in silicone monomer synthesis is hindered by complex reaction mechanisms, diverse operating conditions, and sparse quality-labeled data due to offline detection. These challenges severely limit the performance of conventional data-driven models, particularly under real-time, few-shot scenarios. To address this, a Multi-Scale Fusion Network with Enhanced Meta-Learning (MSF-EML) is proposed for robust prediction in data-scarce industrial environments. The framework employs Bi-LSTM to extract dynamic features, followed by three complementary temporal encoders: Windowed Time-Decay Attention for short-term fluctuations, Dynamic Multi-Scale Convolution for mid-term patterns, and Transformer Encoders for long-range dependencies. These are adaptively fused through a hierarchical mechanism combining channel attention and gating to enhance representation and reduce overfitting. On the algorithmic side, this method integrate an Enhanced Model-Agnostic Meta-Learning (EnhancedMAML) strategy to enable cross-condition generalization. A dual-modal task similarity measure based on Dynamic Time Warping (DTW) and meta-gradient analysis guides task reweighting and clustering. Furthermore, Projecting Conflicting Gradients (PCGrad) is used to mitigate gradient conflicts during meta-optimization, improving convergence stability. Experimental validation on real-world data (2023-2024) from a silicone monomer synthesis plant shows that MSF-EML outperforms existing methods in predictive accuracy, convergence, and generalization, especially under conditions with fewer than 150 samples per working condition. In addition, the method has been applied to the modeling process of the real-time optimization (RTO) system in the plant, effectively improving production efficiency in silicone monomer synthesis.