改进的蒙特卡洛方法用于烟丝近红外光谱定量分析中奇异样本的识别

Improved Monte Carlo method for outlier detection in quantitative analysis of cigarette cut samples using near infrared spectroscopy

【目的】提高烟草及烟草制品的化学成分近红外光谱定量分析校正模型的准确性。【方法】通过改进蒙特卡洛奇异样本检测方法构建不同的定量模型,获得正常样本和可疑样本的预测误差分布,依据分布间的差异识别正常样本与奇异样本。【结果】改进后的烟丝中化学成分的近红外光谱定量分析模型的相关系数和校正均方根误差得到明显提高和改善;总植物碱、总糖、还原糖、总氮、钾和氯的含量模型的预测均方根误差(RMSEP)分别由0.061%、0.799%、0.926%、0.054%、0.115%、0.076%降低到0.059%、0.786%、0.817%、0.048%、0.107%、0.058%;模型稳健性的评价参数SEP/SEC的值均小于1.2。【结论】该方法剔除奇异样本后所得模型具有较好的稳健性。

[Objective]This study aims to improve the accuracy of calibration models for quantitative analysis of chemical compositions in tobacco and tobacco products using near-infrared spectroscopy.[Methods]By using an improved Monte Carlo outlier detection(IMCOD)method,a pool of quantitative models were established.On this basis,a distribution of the prediction errors of normal and suspicious samples was achieved.Based on the distribution differences,normal and outlier samples were identified.[Results]By applying IMCOD to the NIR quantification of the concerned compositions in cigarette cut samples,it can be known that the correlation coefficients of resultant models were not only enhanced significantly but also the root mean square errors of prediction(RMSEP)were improved.The RMSEP values of independent validation samples decreased from 0.061%,0.799%,0.926%,0.054%,0.115%and 0.076%to 0.059%,0.786%,0.817%,0.048%,0.107%and 0.058%in content of total alkaloids,total sugars,reducing sugars,total nitrogen,potassium and chlorine,respectively.All SEP-to-SEC ratio values obtained were less than 1.2.[Conclusion]This indicates the calibration models excluding outlier samples have good robustness and predictive performance.