基于MHSA-LSTM的软测量建模及其在化工过程中的应用  

Soft sensor modeling based on MHSA-LSTM and its application in chemical process

作　　者：李文华 叶洪涛[1,2,3] 罗文广 刘乙奇[4,5] LI Wenhua;YE Hongtao;LUO Wenguang;LIU Yiqi(Guangxi University of Science and Technology,School of Automation,Liuzhou 545616,Guangxi,China;Guangxi Key Laboratory of Automatic Detection Technology and Instruments(Guilin University of Electronic Technology),Guilin 541004,Guangxi,China;Guangxi Key Laboratory of Automotive Components and Vehicle Technology(Guangxi University of Science and Technology),Liuzhou 545616,Guangxi,China;South China University of Technology,School of Automation Science and Engineering,Guangzhou 510640,Guangdong,China;National Engineering Center for Paper Making and Pollution,Guangzhou 510640,Guangdong,China)

机构地区：[1]广西科技大学自动化学院,广西柳州545616 [2]广西自动检测技术与仪器重点实验室(桂林电子科技大学),广西桂林541004 [3]广西汽车零部件与整车技术重点实验室(广西科技大学),广西柳州545616 [4]华南理工大学自动化科学与工程学院,广东广州510640 [5]造纸与污染国家工程中心,广东广州510640

出　　处：《化工学报》2024年第12期4654-4665,共12页CIESC Journal

基　　金：国家自然科学基金项目(62273151,62073145);广东省基础与应用基础研究项目(2021B1515420003);2023年广西汽车零部件与整车技术重点实验室自主研究课题项目(2023GKLACVTZZ07)。

摘　　要：化工过程数据的动态性和非线性等特性常使传统的软测量方法难以准确提取数据的动态和非线性特征,从而影响关键质量变量的预测精度和系统整体的控制优化。因此,提出了一种融合多头自注意力机制的长短期记忆网络(multi-head self-attention mechanism long short-term memory network,MHSA-LSTM)的软测量建模方法。首先,利用LSTM充分挖掘数据的时序特征,以便提取化工过程数据的动态变化信息;其次,使用多头自注意力机制对LSTM隐藏层的输出特征进行加权,可有效地捕捉不同尺度特征向量的长期相关性,且能提高模型的长期记忆能力;然后,将提取的特征向量与其对应的特征权重相乘得出加权结果输入全连接层,可有效地提高关键质量变量预测的精度。对所提方法在脱丁烷塔过程和硫回收单元进行仿真验证,结果表明所建模型的预测精度优于门控循环单元、LSTM以及融合自注意力机制的LSTM软测量模型。The dynamic and nonlinear characteristics of chemical process in data often make it difficult even impossible for traditional soft sensing methods to accurately extract the dynamics and nonlinearity,which affects the prediction accuracy of key quality variables negatively and the overall control optimization of the system.Therefore,this paper proposes a soft sensor model,termed as the multi-head self-attention mechanism long short-term memory network(MHSA-LSTM).First,the LSTM is used to fully exploit the temporal characteristics of the data in order to extract the dynamic change information of the chemical process data.Second,a multi-head self-attention mechanism is used to weight the output features of LSTM hidden layer,and effectively capture the long-term correlation of feature vectors with different scales and improve the long-term memory ability of the model.Furthermore,the weighted results obtained by multiplying the extracted feature vector and its corresponding feature weight are input to the full connection layer.It can effectively improve the accuracy of prediction of key quality variables.Finally,the proposed method is simulated and verified in the debutanizer column process and sulfur recovery unit.The results indicate that the prediction accuracy of the constructed model is superior to gated recurrent unit,LSTM and self-attention LSTM soft sensing models.

关 键 词：软测量 化工过程 神经网络 多头自注意力机制 预测 实验验证 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]