文献类型：期刊文献

中文题名：基于注意力机制的Multi-head CNN-LSTM软测量建模

英文题名：Multi-head CNN-LSTM Soft Sensing Modeling Based on Attention Mechanism

作者：罗顺桦[1];王振雷[1];王昕[2]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237;[2]上海交通大学电工与电子技术中心,上海200240

年份：2022

卷号：29

期号：10

起止页码：1821

中文期刊名：控制工程

外文期刊名：Control Engineering of China

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：国家重点研发计划项目(2018YFB1701103);国家自然科学基金重点项目(61533003);国家杰出青年科学基金资助项目(61925305)。

语种：中文

中文关键词：软测量;卷积神经网络;长短期记忆网络;注意力机制

外文关键词：Soft sensing;convolution neural network;long short-term memory network;attention mechanism

摘要：软测量建模通过选取辅助变量,建立辅助变量与关键质量变量关系,能够高效地实现对关键质量变量的预测。然而当辅助变量维数较高,且对关键质量变量的影响程度不一时,网络预测误差将较大。针对这一问题,提出一种基于注意力机制的Multi-head CNNLSTM模型,首先根据辅助变量自身属性和特点将其切分成多组子变量后,使用多组独立并行工作的CNN-LSTM群对其子变量进行单独处理;再提取各组子变量上的特征向量,融合注意力机制,实现子变量特征向量的权重分配。所提算法不需提前根据工艺知识选择辅助变量,而是通过深度学习机制自动选择特征;最后,在乙烯精馏塔塔顶乙烷浓度软测量建模中进行应用,所提模型的预测精度优于LSTM以及CNN-LSTM软测量模型。
By selecting auxiliary variables and establishing the relationship between the auxiliary variables and key quality variables,soft sensing modeling can effectively predict the key quality variables.However,the network prediction error is large because of the high dimension of the auxiliary variables and its different infulence on the key quality variables.Aiming at this problem,a multi-head CNN-LSTM model based on attention mechanism is proposed in this paper.Firstly,the auxiliary variables are divided into multiple groups of sub-variables according to their attributes and characteristics and the sub-variables are processed separately by using multiple groups of CNN-LSTM that work independently and in parallel.Then,feature vectors of each sub-variable are extracted and combined with attention mechanism to achieve weight distribution of feature vectors of sub-variables.The algorithm proposed in this paper does not need to select auxiliary variables according to process knowledge in advance,but automatically selects features by a deep learning mechanism.Finally,it is applied in the soft sensing modeling of ethane concentration on the top of ethylene distillation tower.The results show that the prediction accuracy of the proposed model is better than that of the LSTM and CNN-LSTM soft sensing models.