基于偏最小二乘算法的高维谱数据特征选择  

Feature Selection Approach of High Dimension Spectral Data Based-on Partial Least Squares Algorithm

作　　者：汤健[1] 贾美英[1] 刘卓[2] 乔勇 赵立杰[4] 

机构地区：[1]北京图形研究所,北京100029 [2]东北大学自动化研究中心,沈阳110004 [3]辽宁金叶纸业有限公司,沈阳110300 [4]沈阳化工大学信息工程学院,沈阳110142

出　　处：《控制工程》2015年第6期1127-1130,共4页Control Engineering of China

基　　金：国家自然科学基金(61573364;61273177);国家863计划项目(2015AA043802);中国博士后基金(2013M532118;2015T81082);辽宁省教育厅科学研究一般项目(L2013272)

摘　　要：基于高维谱数据全部谱变量建立的软测量模型不但存在模型学习速度慢、泛化性和可解释性差等问题,并且难以揭示软测量模型所蕴含的物理含义和进行合理的物理解释等问题。对高维谱数据进行变量选择,降低输入变量维数一直是特征选择领域的热点问题之一。针对这些问题,提出了一种基于偏最小二乘(PLS)算法的高维谱数据特征选择方法。该方法首先分析了基于偏最小二乘算法的潜变量特征提取方法,然后,采用PLS算法分析了原始未标定谱数据的不同谱变量的灵敏度,计算从谱数据提取的不同潜在变量系数的平方和,最后结合球域准则和软测量模型精度基于优化求解的思路进行谱变量选择,进而实现高维谱数据特征的全局优化选择。采用近红外谱数据验证了所提方法的有效性。Modeling with the high dimensional full spectrum, the soft sensor has the characters of lower generation performance and poor interpretation. Moreover, the physical interesting contained in the model cannot be found clearly. One of the open issue problems in feature selection domain is how to select some more useful variables in the high dimensional data. Aiming at these problems, a new feature selection approach of high dimension spectral data based on partial least squares （PLS） algorithm is proposed in this paper. At first, partial least squares based latent variable extraction method is analyzed. Then, the PLS algorithm is used to analyze the sensitivity of the spectral variables based on the unscaled original spectral data. Then, the square sums of different latent variable coefficients are calculated. At last, sphere domain criterion combined with prediction accuracy of the soft sensor model is used to select the most useful spectral variables, which are solved by using the optimization aspect. Thus, the global optimal feature selection for the high dimensional spectral data is realized. The near infrared spectral data is used to validate the oroDosed method

关 键 词：特征选择 高维谱数据 球域准则 偏最小二乘算法 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]