PCA结合马氏距离法剔除近红外异常样品

Eliminating outlier samples in near-infrared model by method of PCA-mahalanobis distance

采用PCA(principal component analysis)结合马氏距离法对近红外校正样品集中的异常样品进行剔除,从校正集的60个食醋样品中剔除了12个异常样,用剩下的48个样品建立了总酸、挥发酸的校正模型,并对预测集的15个食醋样品进行预测分析,以相关系数(R)、预测均方差(RMSEP)、平均相对误差(Er)为预测模型的评价指标.预测集R分别为0.9759,0.9383;RMSEP分别为0.0938,0.1635;Er分别为1.34%,2.80%.与原始校正集所建模型相比,校正模型的预测精度和稳定性得到显著提高.

Outlier samples often strongly influence the stability of the model in near-infrared（NIR） spectroscopy. By using the method of PCA-mahalanobis distance, 12 outliers were detected and eliminated from 60 vinegar samples in calibration set. The calibration model with the remained 48 samples was established and used to estimate the contents of total-acid and volatile acid of the other 15 vinegar samples in predicted set. The results showed that the relative coefficients（R） of the predicted set were 0. 975 9 and 0. 938 3 ; the root mean square errors of prediction were 0. 093 8 and 0. 613 5 ; and the mean relative errors of prediction were 1.34% and 2.80%. Compared to the original calibration model,the predicted accuracy and stability of this model were greatly impoved.