| 引用本文: | 王 珠,刘佳璇.MOb-GRU神经网络工业软测量建模方法与输出预测[J].控制理论与应用,2022,39(9):1758~1768.[点击复制] |
| WANG Zhu,LIU Jia-xuan.MOb–GRU neural network for industrial soft sensor modeling method and output prediction[J].Control Theory and Technology,2022,39(9):1758~1768.[点击复制] |
|
| MOb-GRU神经网络工业软测量建模方法与输出预测 |
| MOb–GRU neural network for industrial soft sensor modeling method and output prediction |
| 摘要点击 1736 全文点击 582 投稿时间：2021-06-21 修订日期：2022-07-01 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2022.10533 |
| 2022,39(9):1758-1768 |
| 中文关键词 软测量技术 MOb–GRU 非线性动态 自适应学习率 神经网络 |
| 英文关键词 soft sensing technology MOb–GRU nonlinear dynamic adaptive learning rate neural networks |
| 基金项目 国家自然科学基金项目(61703434), 中国石油大学(北京)科研基金项目(2462020YXZZ023)资助. |
|
| 中文摘要 |
| 由于工业过程具有强非线性、动态特性与慢时变性, 其完整性建模相对较难. 针对工业过程的现有软测量技术并未综合考虑过程的非线性和动态特性, 本文提出了一种依赖模型阶次的GRU(MOb–GRU)神经网络软测量模型, 针对非线性动态过程进行全动态建模. 首先, 在MOb–GRU的结构选择上, 本文根据所研究实际对象的动态特性复杂程度确定网络的总模块数. 另外, MOb–GRU能灵活设置反向更新的单元数, 这种设置打破了传统GRU只能从第1个模块开始输出的限制. 其次, 为使记忆网络以较快的速率收敛到最优, 本文分别设计了基于自适应学习率和学习率矩阵的网络训练算法. 接着, 仿真实验分别选取了典型的单变量与多变量非线性动态过程, 并采用MOb–GRU神经网络对其进行建模和预测. 最后, 仿真结果证实了MOb–GRU网络结构的合理性以及训练算法的高效性. |
| 英文摘要 |
| Modeling the integrity of industrial process is a relatively difficult task due to its strong nonlinearity, dynamic characteristics and slow time variability. Though there exist some soft sensing technologies for industrial process, they fail to consider the nonlinear and dynamic characteristics comprehensively of the process. Therefore, this paper proposes a model order based gated recurrent unit (MOb–GRU) neural network soft sensor model for fully-dynamic modeling of nonlinear dynamic process. Specifically, firstly, in terms of the MOb–GRU structure selection, this paper determines the total module number of the network according to the complexity of dynamic characteristics of the actual object. Moreover, the MOb–GRU can flexibly set the number of units for reverse update, which breaks the limitation that the traditional GRU can only output from the first module. Secondly, in order to make the memory network converge to the optimal at a faster rate, this paper designs the network training algorithms based on the adaptive learning rate and the learning rate matrix, respectively. Then, the typical univariate and multivariable nonlinear dynamic processes are selected in the simulation experiment, and the MOb–GRU neural network is used to model and predict them. Finally, the rationality of MOb–GRU network architecture as well as the high efficiency of the training algorithms is demonstrated through the simulation results. |
|
|
|
|
|