Feature Extraction And Similarity Measure On Tobacco Near Infrared Spectra

In order to realize the individual evaluation and comprehensive evaluation oftobacco quality, the traditional assessment method for quality of raw material mainlyuse the chemical composition, physical characteristics of the original data, appearancequality of tobacco smoking, But there exist some drawbacks when this technology isused, such as Tobacco leaf quality evaluation lagging, not comprehensive and notunified. The rapid development of near infrared spectroscopy technology andcomputer application technology provide the advantage for tobacco leaf qualitydetection and analysis. Near infrared spectroscopy (NIRS) is a kind of absorptionspectrum in near infrared region, which is obtained through diffusing way by utilizingflood frequency vibration or rotation of C-H, N–H, O-H and C-C bonds in organicmatters. In tobacco leaf, there exist many chemical compositions, including total sugar,total nitrogen, reducing sugar, nicotine, chlorophyll and other substance. Thesesubstances contain abundant chemical bonds. So, near infrared spectroscopy can beused to mine the key feature of tobacco leaf spectrum and further to carry outqualitative and quantitative analysis. The many drawbacks of traditional chemicalmethods can be solved. In this dissertation, the application research of Near infraredspectroscopy in qualitative and quantitative analysis of tobacco leaf was conducted.The researches include tobacco near-infrared spectrum characteristics analysis, thedimension reduction of high dimensional spectral feature mapping, spectralwavelength points of variable selection, PLS modeling characteristics of componentextraction. The innovation achievements are as follows.1、From the Angle of molecular spectroscopy of tobacco leaves, the dissertationputs forward key chemical composition, style characteristic components in tobacconear-infrared spectra of effective expression. There are direct or indirectrelation between high aroma components or conventional chemical compositionand near infrared spectroscopy, these features on the expression of quantitativedetection of tobacco and varieties, location, grade, style characteristics plays an important fundamental role.2、When NIRS is applied in the spectrum feature extraction and similaritymeasure, some problems such as dimension disaster, information processing failure,and distance invalid can occur. This is because that near infrared spectrum ischaracteristic of high-dimensional, overlapping, and nonlinear and redundant. In thispaper, local maintaining projection method INLPP based on improved neighborhoodwas proposed. The design idea of the algorithm and improving process were discussedin detail. And it was compared with PCA, LDA, and LPP algorithms. The researchresults show that the algorithm can achieve dimension reduction and maintainnonlinear data structure.3、In NIRS, it is hard to build multivariate calibration model based on thespectrum of all wavelengths because too much poin ts in wavelength and seriousoverlap of bands. Especially, the strong absorption of moisture in the near infraredspectral band overlap of the region and the background of the other components, canreduce the performance of the model robustness and metastatic. In the face of all theseproblems, this dissertation puts forward wavelength variable selection method basedon CARS, and will choose the wavelength points for total sugar, total nitrogen andtotal nicotine index model. Compared with the full wavelength, the modelperformance has greatly improved.4、In near infrared quantitative modeling, PLS is one of the best and most widelyused modeling method. PLS method is mainly by extracting potential ingredients,containing multiple dependent variable and independent variable of data modelinganalysis. But how to extract the partial least squares ingredients and select somevaluable and meaningful component modeling as the final composition is one of themain problems that this paper plan to solve. Aiming at this problem, this dissertationimproved the PLS extracting principal component principle, to ensure that the extractingredients most powerful account of the dependent variable. At the same time, theexpression of the independent variable X has the best comprehensive ability. Thesimulation results show that the proposed improved PLS method on modelperformance has improved greatly. 5、In this paper, under LPP, the category information was added into the distancecalculation of k neighborhood, and the similarity measure of the different sampleswere obtained under the lower dimensional. The two groups of experimental datawere used for similarity measurement performance test, and all showed the goodperformance.6、The tobacco leaf quality rapid detection analysis networked system wasindependently researched and developed through the multiple key technology research.Through this system, the detection and the main chemical components in tobacco leafproduction areas discrimination of different styles of scent through a variety of dataanalysis methods of the system research on Shandong tobacco leaf quality division.The obtained results are same with that given by the experts in Shandong region. Thisalso laid the foundation.for the further application of NIRS in tobacco industry.