电子测量与仪器学报

郭中宇,岳玉麒,陈娟.多视角对偶图注意力网络的软测量方法[J].电子测量与仪器学报,2026,40(1):191-200

多视角对偶图注意力网络的软测量方法

Soft sensing method based on multi view dual graph attention network

DOI：

英文关键词:soft sensing multi view learning graph attention network cosine similarity

基金项目:国家自然科学基金面上项目（61771034）资助

作者	单位
郭中宇	北京化工大学信息科学与技术学院北京100029
岳玉麒	北京化工大学信息科学与技术学院北京100029
陈娟	北京化工大学信息科学与技术学院北京100029

Author	Institution
Guo Zhongyu	College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China
Yue Yuqi	College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China
Chen Juan	College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China

摘要点击次数: 494

全文下载次数: 563

中文摘要:

在复杂的过程工业中,由于缺乏在线检测仪表或者由于恶劣的生产环境,导致一些生产过程中的关键变量难以测量,且无法实现在线测量,因此在过程工业中需要对这些变量进行软测量建模的研究。当前,基于深度学习的软测量建模大多关注的是单一视角下的特征建模,从而丢失了一些其他视角下有价值的信息,导致预测模型精度低。针对该问题,提出了一种多视角对偶图注意力网络(multi-view dual graph attention, Mv-DGAT)的工业软测量建模方法。该方法首先搭建多视角框架,构建基于最大信息系数的空间图注意力网络(spatial graph attention, SGAT),完成空间视角，并搭建基于多层次时序图结构和长短时记忆网络(long short-term memory, LSTM)的时间图注意力网络(temporal graph attention, TGAT),建立时间视角。其次,使用多头注意力机制进行时空特征融合预测。最后引入余弦相似度评估视角间的互补性,抑制冗余特征。所提出的方法在真实工业流程公开数据集上进行了实验,实验结果表明，所提出方法预测精度高,决定系数R2均达到085,均方根误差较对比模型降低10%以上。

英文摘要:

In the complex process industry, due to the lack of online detection instruments or harsh production environments, some key variables are difficult to measure and cannot be measured online. Therefore, research on soft sensing modeling of these variables is needed in process industries. Currently, deep learning based soft sensing modeling mostly focuses on feature modeling from a single perspective, neglecting valuable information from other perspectives, resulting in low accuracy of the prediction model. To address this issue, this paper proposes an industrial soft sensing modeling method based on a multi-view dual graph attention (Mv-DGAT) network. This method first constructs a multi-view framework, builds a spatial graph attention (SGAT) network based on the maximal information coefficient to complete the spatial perspective, and constructs a temporal graph attention (TGAT) network based on a multi-level temporal graph structure and the long short-term memory (LSTM)network to establish the temporal perspective. Secondly, the multi-head attention mechanism is used for spatiotemporal feature fusion prediction. Finally, the cosine similarity is introduced to evaluate the complementarity between views and suppress redundant features. The proposed method was tested on a publicly available dataset of real industrial processes. The experimental results showed that the proposed method has high prediction accuracy, with determination coefficients R2 reaching 0.85 and root mean square error reduced by more than 10% compared to the comparison model.

查看全文查看/发表评论下载PDF阅读器