The Research And Analysis Of Leaf Samples Modeling And Model Transfer Based On Different Physical States

In the process of making tobacco, testing of key chemical indicators such assugar, nicotine or nitrogen in tobacco leaves is usually needed.Traditional laboratoryanalysis and testing methods are not suitable for real-time online analysis.Near-infrared (NIR) spectroscopy is a rapid non-destructive analysis techniques forreal-time analysis of the quality of tobacco samples online. In this paper, a hybridmodel of the spectrum is built using near-infrared technology, which can accuratelypredict the concentration of the components of tobacco. In order to achieve the goalof the design, this paper has done the following work:1. By reading a lot of literature,it introduces the tobacco research and model evaluation method at home and abroadand it does a more complete elaboration to the near infrared spectroscopy analysis ofthe development and application of technology in various areas.2. It introducesthe pretreatment method and data standardization processing method of spectralanalysis in the near infrared spectra, according to the root mean square error ofprediction and the relative standard deviation of the minimum principle, researchingthe pretreatment method of tobacco in different physical conditions. The results showthat: in the total sugar as an example, the pretreatment method of the most suitabletobacco spectrum is the root mean square average with two orders Savitzky-Golayderivative smoothing; the most suitable pretreatment method tobacco sheet spectrumis arithmetic average combination of first order Savitzky-Golay derivative smoothing.3. Because the prediction errors on the tobacco smoke plate spectral model is toolarge, so it reduces the prediction errors from the perspective of prediction model. Amixed model is established and the end of the cigarette smoke uses the partial leastsquares method, using the model evaluation methods to estimate error modelof predicting the improving effect of smoke. The results show that: the training set iscomposed of187tobacco powder spectra and50smoke models is established for theoptimal spectral mixture model, the root mean square error of tobacco leaf spectralprediction is reduced from1.592to1.29, the relative standard deviation is reducedfrom5.494%to4.242%, the correlation coefficient increased from0.895to0.915.4.Based on the leaves and a mixture model tobacco powder spectrum is introduced onthe model transfer algorithm, compared the shenk’s algorithm and slope/intercept algorithm model in smoke and smoke at the end of two kinds of physical state ofthe transfer effect, shenk’s algorithm for model transfer from spectral angle,slope/intercept model transfer algorithm in predicting value perspective. Theexperimental results show that: shenk’s algorithm does not realize model sharing inthe two kinds of physical condition; slope/intercept method achieves better sharingmixed model under two kinds of physical condition, the prediction results ofthe relative standard deviation is reduced from4.242%to3.057%. At the sametime, the hybrid model has little effect on the smoke at the end of the forecasterror; the relative standard deviation is increased from2.99%to3.052%. Therefore, itcan set up a cigarette and tobacco spectral mixture model to predict the smoke andtobacco powder spectra of the component concentration.5. The paper is summarizedfrom the analysis results and innovation. It puts forward that the next work willfocus on the establishment of a hybrid model of tobacco leaves, tobacco and smoke atthe end of three kinds of physical condition of the spectrum, and it will graduallyrealizes the near infrared spectral analysis technology based on real-time detection oftobacco quality.