用于近红外校正模型的训练样本选择  

Training Sample Selection Used in NIR Calibration Model

作　　者：张其可[1] 戴连奎[1] 

机构地区：[1]浙江大学智能系统与决策研究所,杭州310027

出　　处：《传感技术学报》2006年第4期1190-1194,共5页Chinese Journal of Sensors and Actuators

基　　金：国家高技术研究发展计划项目资助(2003AA412110)

摘　　要：近红外(near-infrared,NIR)校正模型建立时传统的训练样本选择方法只考虑光谱欧氏距离的同类就近选取,不考虑光谱特征异常的训练样本的影响。基于主元分析(principal component analysis,PCA)残差,在同类就近取样的基础上引入异常光谱剔除技术进行训练样本的二次提取,用于建立偏最小二乘(partial least squares,PLS)回归模型。实验结果表明,该方法比传统方法的预测精度有较明显的提高。When building a near-infrared spectroscopy （NIR） calibration model, traditional training sample selection method only considers selecting nearest neighbour samples by Euclidean distance, without considering the influence of training samples with abnormal spectrum feature. Based on principal component analysis （PCA） residue, this paper introduces abnormal sample eliminating technique to selected nearest neighbour samples. These improved sample set is then applied to partial least-squares regression （PLSR） modeling. Experiment results show, introduction of abnormal sample elimination improves prediction precision of NIR model obviously.

关 键 词：主元分析 近红外光谱 校正模型 训练样本选择 汽油辛烷值 

分 类 号：TP212[自动化与计算机技术—检测技术与自动化装置] O659.32[自动化与计算机技术—控制科学与工程]