), 程兰, 陈泽华 太原理工大学 电气与动力工程学院，太原 030024
收稿日期:2019-12-31出版日期:2020-12-01发布日期:2020-12-31通讯作者:阎高伟E-mail:yangaowei@tyut.edu.cn作者简介:来颜博◎(1993-)，男，河南省商丘市人，硕士生，现主要从事机器学习、软测量与迁移学习的研究．基金资助:国家自然科学基金(61973226);山西省重点研发计划(201903D121143);山西省科技重大专项(20181102017);山西省自然科学基金(2015011052)Unsupervised Regression Model of Geodesic Flow Kernel Based on Dynamic Independent Component Analysis and Dynamic Principal Component Analysis
LAI Yanbo, YAN Gaowei(), CHENG Lan, CHEN Zehua College of Electrical and Power Engineering, Taiyuan University of Technology, Taiyuan 030024, China
Received:2019-12-31Online:2020-12-01Published:2020-12-31Contact:YAN Gaowei E-mail:yangaowei@tyut.edu.cn摘要/Abstract
摘要： 针对工业过程中工况改变时，传统软测量模型难以适应数据分布变化，易出现模型性能恶化的问题．引入一种基于测地线流式核的迁移学习方法，同时针对该方法难以解决工业过程中动态特性提取和数据不完全服从高斯分布问题进行优化．首先构建增广矩阵以应对过程中的动态特性，对处理后的数据进行独立成分分析和主成分分析，用以提取源域与目标域的非高斯信息和高斯信息，并在格拉斯曼流形空间下对源域的非高斯信息和高斯信息分别适配目标域，最后使用最大均值差异方法对适配后的源域与目标域进行分布度量，并为基于源域构建的模型加权．结果表明该方法不仅降低了源域和目标域的分布差异，而且解决了工业过程中的动态特性提取和其数据不完全服从高斯分布的问题．通过在田纳西伊斯曼数据上的实验，证明了模型的有效性和实用性．
关键词: 软测量, 测地线流式核, 动态特性, 动态独立成分分析, 动态主成分分析
Abstract: It is difficult to accurately measure parameters by using the traditional soft sensor algorithm when the working condition of industrial process is changed. Therefore, a transfer learning strategy is introduced based on geodesic flow kernel to solve this problem. At the same time, the method is optimized to solve the problems of dynamic characteristic extraction and incomplete Gaussian distribution in industrial process. The augmented matrix is first constructed to deal with the dynamic characteristics of the process. Independent component analysis and principal component analysis are performed on the processed data to extract the non-Gaussian and Gaussian information of the source domain and the target domain. Then, the non-Gaussian and Gaussian information of the source domain is adapted to the target domain respectively on the Grassmann manifold. Finally, the maximum mean discrepancy is used to measure the distribution between the source domain and the target domain after domain adaptation, and the calculated results are applied to construct the weight of the model based on the source domain after domain adaptation. The results show that the method not only reduces the difference of distribution between the source domain and the target domain, but also solves the problems of dynamic characteristic extraction and incomplete Gaussian distribution in industrial process. The validity and the practicability of the model are proved by experiments on Tennessee Eastman data.
Key words: soft sensor, geodesic flow kernel, dynamic characteristics, dynamic independent component analysis, dynamic principal component analysis
PDF全文下载地址:
点我下载PDF
