Study On The Error Of Tobacco Near Infrared Correction Model And Prediction Process

Near infrared spectroscopy analysis technique was the key to the use of chemometric methods to build robust mathematical models,so as to realize the unknown samples fast and accurate analysis.But in the processing of near infrared calibration and prediction,the stability of the instrument?sample collection?sample size uniformity?algorithm selection and other factors had a certain effect on the quality of the spectra and model,and would cause the prediction error in results.Therefore,the various influencing factors of in the processing of tobacco near infrared calibration and prediction was key to ensure the accuracy of the NIR analysis.According to the GB/T-29858 2013?Standard guidelines for molecular spectroscopy multivariate calibration quantitative analysis?four categories?nineteen kinds error?of quantitative correction,which based on the research station in recent years,the three aspects errors mainly from the of water?model selection and the difference on near infrared instrumentin in the processing of correction and prediction ware studied in detail.The main results were as follows:1?The study of the qualitative and quantitative error of the moisture content in cigarette near infrared was done.The main absorption interval of moisture content was in 6944cm^-11 and 5155cm^-1.With the increase of moisture content,the absorption intensity of the corresponding interval was bigger,and the difference between moisture content could not be effectively eliminated through the first derivative and other preprocessing methods.Moisture content also affected the distribution of the principal component space,along with the continuous increase of moisture content of cigarette samples,the difference in the distribution of the principal component distance was increases;The fixed sample position reduced the interference on the experimental results of uneven heating of microwave heating process,and compared with the traditional oven drying method found that the prediction accuracy by microwave drying 1?2min was better than the oven drying for 2h at 40?.The microwave drying method could not only improve the drying efficiency at the same time play a certain effect in sample prediction accuracy.2?According to different origins and grades,20 standard samples?10 samples from?-?areas,10 samples from?-?areas?were choosed to analyse the error of routine chemical composition models of redried lamina samples.The average relative deviation?ARD?and standard error of prediction?SEP?were evaluated of the model prediction results for different sample,and the conclusions were as follows:when the?-?areas model predicted samples in?-?areas,the accuracy of the prediction results was better than that the?-?areas model and the total model.The?-?areas model was better than the?-?areas model and the total model when predicting the samples in?-?areas.The total model predicted mixed samples that had the better prediction accuracy.Considering that the prediction results of?-?areas model and?-?areas model were not obviously advantage to the total model,when using the total model to predict the sample,not only the problem of model selection and repetitive modeling were avoided,but also the efficiency of the model could be effectively improved.Therefore,the total model also had certain advantages in the actual application process.3?The KS algorithm was used to select 60 standard samples?40 for the calibration set,20 for the validation set?.Based on the spectrum,principal component?PCA?and Mahalanobis distance?MD?,the rapid determination of the error between different instruments was realized;Then three kinds of model transfer method?bias,SBC and SST?were used for different brands of FT near infrared instruments.The bias correction and SBC were simple and convenient,which could meet the requirements of model transfer,but the correction of transfer capacity was limited.The transfer result of SST was the best,and the SEP and ARD after transfer were less than those of bias correction and SBC.It could not only achieve different brands of FT NIR model transfer requirements,but slso had high application and popularization value.4?A new model transfer algorithm based on prediction result correction was proposed,which can effectively reduce the amount of computation of SBC algorithm.The transfer results of the algorithm were compared with the three algorithms of bias correction,SBC and SST,and it was found that the model transfer result of the algorithm close to SBC.