基于非线性慢特征回归的动态软测量建模  

Dynamic soft sensor modeling based on nonlinear slow feature analysis

作　　者：狄克松 王宇红[1] 尚超[2] 黄德先[2] 

机构地区：[1]中国石油大学(华东)信息与控制工程学院,山东青岛266580 [2]清华大学自动化系,北京100084

出　　处：《计算机与应用化学》2016年第11期1160-1164,共5页Computers and Applied Chemistry

基　　金：山东省自然科学基金项目(2013ZRE28089)

摘　　要：隐变量模型如部分最小二乘已经被广泛用于建立低维子空间,并以此建立回归模型用于质量预测。然而,它们都是基于工业过程的静态假设,一般实际的工业过程都是动态的。本文提出一种非线性慢特征回归模型,用作动态软测量模型。首先,对线性慢特征分析进行非线性扩展,然后非线性的慢特征作为隐变量通过扩展后的慢特征分析从过程数据中被提取出来。不同于传统的隐变量模型,慢特征分析假设隐变量具有缓慢变化的动态特性。由于工业过程明显的动态变化,慢特性可以被看作有效的先验知识加以利用。最后,利用提取的慢特征建立回归模型并用于产品质量的预测。实验结果表明,基于非线性慢特征的软测量模型要比传统的软测量模型预测精度高。Latent variable model have been widely used to derive low-dimensional subspaces and build regression models. However, they are based on the assumption of static industrial process, on the contrary, the actual industrial is generally dynamic. In this paper, a nonlinear slow feature regression （SFR） model is proposed, which is used as a dynamic soft sensor model. Firstly, the linear slow feature analysis （SFA） is expanded into nonlinear SFA through the second order polynomial. Then, the nonlinear slow features as hidden vari- ables are extracted though the nonlinear SFA from process data. Different from the classical I,V models, SFA assumes the LVs have slowly dynamic characteristics. Due to obvious dynamics of the industrial process, slow feature can be regarded as the prior knowledge. Finally, the regression model is established based on the extracted nonlinear slow features and used to predict the product quality. Ap- plication in the TE process verify the effectiveness of the model. The experimental validation shows that the SFR model is more accu- racy than the traditional method, for example, partial least squares （PLS） and support vector machine（SVM）. The test error is reduced 3.5% than PLS model and reduced 13.6% than SVM model. The SFR model has higher prediction accuracy.

关 键 词：动态软测量 慢特征回归 慢特征分析 慢特征 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]