引用本文:吴修粮,赵磊,曹茂永,孙凯.面向空气分离过程的动态集成增量学习软测量算法研究[J].控制理论与应用,2026,43(4):843~852.[点击复制]
WU Xiu-liang,ZHAO Lei,CAO Mao-yong,SUN Kai.Research on dynamic ensemble incremental learning soft sensor algorithms for an air separation process[J].Control Theory & Applications,2026,43(4):843~852.[点击复制]
面向空气分离过程的动态集成增量学习软测量算法研究
Research on dynamic ensemble incremental learning soft sensor algorithms for an air separation process
摘要点击 161  全文点击 25  投稿时间：2024-04-29  修订日期：2025-04-25
查看全文  查看/发表评论  下载PDF阅读器   HTML
DOI编号  10.7641/CTA.2024.40246
  2026,43(4):843-852
中文关键词  循环神经网络  空气分离过程  Log-Cosh估计  L1正则化  增量学习  软测量
英文关键词  recurrent neural networks  air separation process  Log-Cosh estimation  L1 regularization  incremental learning  soft sensor
基金项目  山东省自然科学基金项目(ZR2021MF022), 山东省科技型中小企业创新能力提升工程项目(2023TSGC0399)资助.
作者单位E-mail
吴修粮 山东科技大学电气与自动化工程学院 qlwuxiuliang@sina.com 
赵磊 齐鲁工业大学(山东省科学院）信息与自动化学院  
曹茂永 山东科技大学电气与自动化工程学院  
孙凯* 齐鲁工业大学(山东省科学院）信息与自动化学院 sunkai79@qlu.edu.cn 
中文摘要
      针对空气分离过程多模态、多变量、大时延及存在测量离群点等问题, 本文提出了一种基于Lipschitz循环 神经网络(LRNN)的动态增量学习软测量算法. 首先, 利用易收敛、稳定性好的LRNN处理过程数据中存在的非线性 和时延问题. 其次, 设计了一种新颖的 LRNN 损失函数, 该函数结合了Log-Cosh估计对离群点的鲁棒性和L1正则化 的模型稀疏特性, 能够有效处理过程的测量离群点和模型冗余问题. 进一步, 对不同鲁棒LRNN基础模型进行动态 加权集成, 提高了模型对生产过程模态变化的适应性. 最后, 将所提出的算法应用于某钢厂实际空气分离过程中低 压塔出口O2浓度的软测量建模, 并通过与其他先进算法进行比较, 验证了所提出方法的有效性和优越性.
英文摘要
      To address the problems of multi-modality, multi-variable, large time delay and existence of measurement outliers in an actual air separation process, a dynamic incremental learning soft sensor algorithm based on a Lipschitz recurrent neural network (LRNN) is proposed. Firstly, the LRNN with easy convergence and excellent stability is used to deal with the nonlinearity and time delay in the process data. Secondly, a novel LRNN loss function is designed by combining the robustness to outliers of the Log-Cosh estimation and the sparsity of the L1 regularization, which is effective in dealing with the problems of measurement outliers and model redundancy of the process. Further, the adaptability to the modal changes of the production process is improved by the dynamic weighted ensemble of different robust LRNN base models. Finally, the proposed algorithm is applied to the soft sensor modeling of the O2 concentration at the outlet of the low-pressure tower in an actual air separation process in a steel plant, and the effectiveness and superiority of the proposed algorithm is verified by comparing with other state-of-the-art algorithms.