文献类型：期刊文献

中文题名：偏最小二乘近红外光谱模型中潜变量个数对模型传递性能的影响

英文题名：Effect of Number of Latent Variables for Partial Least Square Model Based on Near Infrared Spectroscopy on Models Transfer Performance

作者：李永琪[1];洪士军[1];黄雯[2];张立国[1];葛炯[2];栾绍嵘[1];倪力军[1]

机构：[1]华东理工大学化学与分子工程学院,上海200237;[2]上海烟草集团有限责任公司技术中心理化实验室,上海200082

年份：2020

卷号：39

期号：10

起止页码：1231

中文期刊名：分析测试学报

外文期刊名：Journal of Instrumental Analysis

收录：CSTPCD;;北大核心:【北大核心2017】；CSCD:【CSCD2019_2020】；

基金：国家烟草专卖局卷烟烟气重点实验室开放性课题(K2018-156P)。

语种：中文

中文关键词：近红外光谱模型传递;偏最小二乘;潜变量个数;玉米;烟叶

外文关键词：near infrared spectroscopy model transfer;partial least square;number of latent variables;corn;tobacco

摘要：以玉米中水分、蛋白质、脂肪和淀粉4种主要成分含量以及烟叶总植物碱的偏最小二乘近红外光谱(PLS-NIRs)模型传递为例,考察了模型中潜变量个数(nLVs)对模型传递误差的影响。研究发现,根据累积贡献率大于99.9%确定的玉米、烟叶样品PLS-NIRs模型的nLVs分别为1和13,nLVs=1时建立的玉米模型对两台从机样品4个成分的预测值和主机预测值的重现性指标均满足国标要求;nLVs=13时建立的烟叶总植物碱模型经分段直接校正(PDS)后,可使4台从机样品的平均相对预测误差(MRE)小于6%。采用留一交叉验证或四折交叉验证确定的玉米、烟叶PLS-NIRs模型的nLVs分别为5~10,16与19,在这些nLVs下建立的玉米PLS-NIRs模型对从机样品的预测误差显著增大,超过许可的误差范围,且模型即使经PDS校正后,从机样品预测值与主机样品预测值的重现性指标大多不满足国标要求;nLVs>13时所建烟叶总植物碱PLS-NIRs模型的转移误差随nLVs增大而增大,且PDS校正后不能保证模型对所有从机样品的MRE小于6%。根据累积贡献率大于99.9%或接近99.9%为准则选取nLVs,可有效避免过拟合,提高NIRs模型的传递性能。
Using the calibration model transfer of PLS-NIRs models for predicting contents of moisture,protein,fat and starch in corn,as well as total alkaloids in tobacco leaves as an example,effect of number of latent variables(nLVs)on the transfer errors of the models were investigated in this paper.It was found that the nLVs in PLS-NIRs models for corn and tobacco leaves selected by cumulative contribution rate greater than 99.9%were 1 and 13,respectively.The prediction reproducibilities for the four ingredients in corn between master and slave samples predicted by the PLS-NIRs models with one latent variable all satisfied the requirements of national standards.When the PLS-NIRs model predicting total alkaloids content built on the master with 13 latent variables was transferred to four slaves,mean of relative prediction errors(MRE)of tobacco leaves tested on the four slaves were all lower than 6%after piecewise direct standardization(PDS)correction.While the nLVs in PLS-NIRs models for corn and tobacco leaves determined by leaving one sample in turn as cross validation set or fourth-fold cross validation method were 5-10,16 and 19,respectively.The prediction errors for the slave corn samples derived from the models with nLVs greater than 5 were significantly increased and exceeded the allowable error level.Even after being corrected by PDS method,most indices of prediction reproducibility for the four ingredients in corn between master and slave samples given by these models could not satisfy the requirements of national standards.The transfer errors of PLS-NIRs models for total alkaloids in tobacco leaves by selecting nLVs greater than 13 increased with the increase of nLVs,while PDS correction cannot guarantee the MRE for all slave instruments given by these models lower than 6%.Results indicated that selecting nLVs for PLS-NIRs models based on the principle of accumulative contribution rate greater than 99.9%or near to 99.9%could effectively avoid over-fitting and improve the transfer performance of the models.