| 引用本文: | 隋璘,马君霞,熊伟丽.基于注意力绞杀的门控循环单元网络及其工业软测量应用[J].控制理论与应用,2026,43(2):423~435.[点击复制] |
| SUI Lin,MA Jun-xia,XIONG Wei-li.Gated recurrent unit network based on attention garrote and its application for industrial soft sensors[J].Control Theory & Applications,2026,43(2):423~435.[点击复制] |
|
| 基于注意力绞杀的门控循环单元网络及其工业软测量应用 |
| Gated recurrent unit network based on attention garrote and its application for industrial soft sensors |
| 摘要点击 127 全文点击 17 投稿时间：2023-08-18 修订日期：2025-03-06 |
| 查看全文 查看/发表评论 下载PDF阅读器 HTML |
| DOI编号 10.7641/CTA.2024.30565 |
| 2026,43(2):423-435 |
| 中文关键词 软测量 门控循环单元 注意力机制 非负绞杀 变量选择 稀疏优化 |
| 英文关键词 soft sensor gated recurrent unit attention mechanism nonnegative garrote variable selection sparse optimization |
| 基金项目 国家自然科学基金项目(61773182), 国家重点研发计划子课题项目(2018YFC1603405–03), 江苏省研究生科研实践创新计划项目(KYCX2 3 2447)资助. |
|
| 中文摘要 |
| 复杂工业过程非线性、动态性及变量冗余会导致建模难度增加与模型性能降低, 因此, 本文提出一种基于
注意力机制与非负绞杀(NNG)估计的门控循环单元(GRU)网络, 并应用于实际工业过程软测量建模. 首先, 将时序
注意力引入GRU网络, 根据不同时刻隐含层间的时序相关性, 自适应分配注意力权重, 提高模型时序特征表征能力;
其次, 设计了一种针对过程变量的注意力权重向量, 并嵌入NNG算法约束, 以实现其系数的近似无偏估计; 进一步
采用带变量注意力的NNG算法对GRU网络进行稀疏优化, 降低模型复杂度, 提高其可解释性并防止过拟合, 通过数
值仿真验证了算法的有效性和优越性; 最后, 将算法应用于某燃煤电厂烟气脱硫过程排放净烟气SO2浓度的软测
量, 实验结果表明所提算法具有优于其它先进对比算法的性能, 能够在有效剔除冗余变量并简化模型结构的同时提
高其预测性能. |
| 英文摘要 |
| The nonlinearity, dynamics and variable redundancy of complex industrial processes lead to increase modeling difficulty and reduce model performance. Therefore, a gated recurrent unit (GRU) network based on attention mechanism and nonnegative garrote (NNG) estimation is proposed and applied to actual industrial process soft sensor modeling.
Firstly, the temporal attention is introduced into the GRU network, and the attention weights are adaptively assigned according to the temporal correlation between the implied layers at different moments to improve the model of temporal
feature characterization capability. Secondly, an attentional weight vector for process variables is designed and embedded with NNG algorithm constraints to approximate unbiased estimates of its coefficients. Then the NNG algorithm with
variable attention is used to perform sparse optimization of the GRU network to reduce model complexity, improve its
interpretability and prevent overfitting. The effectiveness and superiority of the algorithm are verified by numerical simulation. Finally, the proposed algorithm is applied to the soft sensor of SO2 concentration in net flue gas emissions from a
coal-fired power plant desulphurization process. The experimental results show that the proposed algorithm outperforms
other advanced comparative algorithms and improves its predictive performance while effectively eliminating redundant
variables and simplifying the model structure. |
|
|
|
|
|