Research On Quantitative Modeling Technology Of Near Infrared Spectroscopy

With the progress of the times and the daily advancements in technology, especially the improvement of the manufacturing technology of near infrared spectrometer, near infrared reflectance spectroscopy (NIRS) analysis technique has developed rapidly. In the process of the product quality monitoring in many industries (like oil, tobacco, water, medicine), the applications of near infrared spectral analysis technology greatly shorten the time of the quality detection, improve working efficiency. However, compared with the traditional chemical analysis method, the method of the use of near infrared spectral analysis technology for establishing a quantitative model also has certain error. Its prediction accuracy needs to be improved.In the process of near infrared spectral quantitative modeling analysis, there are many factors influencing the accuracy of the analysis results, such as type of near infrared spectrum instrument, form of test samples, scanning environment, the selection of wavelength range, the selection of modeling algorithm, the selection of correction samples and interference of singular samples and other factors. The modeling algorithm directly determines the accuracy of the modeling analysis results. Additionally, the composition of samples in calibration set and the accuracy of its basic data determine the applicability and accuracy of the correction model among them. Therefore, in the quantitative modeling technologies of near infrared spectral, selection of modeling algorithm and eliminating singular sample in the calibration set are becoming hot topics in the modern research for scholars.In order to further improve the accuracy of quality detection, and to solve this discovery of the above hotspot issues, the thesis begins with analyzing the characteristics of near infrared spectral data. The thorough research on modeling technology was conducted. The research of this thesis includes two aspects. That is singular sample identification algorithm and quantitative modeling algorithm. The main research contents of this thesis are as follows:1) This dissertation describes background and significance of the research, and analyzes the present situation of research on modeling technology of near infrared spectroscopy quantitative analysis from two key aspects (singular sample recognition algorithm and quantitative modeling algorithm). This chapter points out the direction for the following research.2) This chapter introduces the physicochemical basis of near infrared spectroscopy analysis and the general process, and expounds the methods of near infrared singular sample recognition, the algorithms of quantitative modeling and the methods of model evaluation in common use. This chapter provides the basic reference for the later paragraphs about the optimization research of near infrared singular sample recognition method and quantitative modeling algorithm.3) Aiming at the low recognition accuracy problem for multiple singular samples in the calibration set, this chapter analyzes shortcomings of commonly used methods of recognition on near infrared singular sample. In order to improve the accuracy of multiple singular samples identification, this chapter improves the calculation strategy of leverage value based on leverage value method. To a certain extent, that reduces the dependence on the center of sample data set, and puts more distance between singular samples and normal samples. In addition, in order to avoid the unreasonable threshold set up by artificial according to experience, this chapter introduces the concept of jump degree in the field of Statistics to solve this problem. Combined with the new improved leverage value, an improved algorithm is proposed. That is Improved Algorithm of Automatic Threshold Value Based on the Leverage. Furthermore, the new algorithm is compared with some commonly used algorithms (such as Mahalanobis distance algorithm, Leverage-spectral residuals method) by theoretical analysis and quantitative modeling experiment. This comparison of experimental result verifies the validity of the improved algorithm. That provides quality guarantee of data in the calibration set for quantitative modeling of the subsequent chapters.4) Aiming at the nonlinear problem between the variable properties of near infrared spectrum and material concentration (content) in NIRS quantitative modeling, partial least squares (PLS) and artificial neural networks (ANN) method cannot obtain satisfactory results. In this chapter, the research optimizes the PLS algorithm. By drawing on previous research results, that increases a residual model part on the basis of the ANN algorithm. Based on the above work, this chapter proposes partial least squares residual neural networks (PLS-RNN) modeling algorithm and focus on the research of using PLS-RNN algorithm to make a quantitative modeling for the near infrared reflectance spectrum having nonlinear problem, and through theory and experiment respectively to compare the effects of the application of PLS-RNN^ PCA-ANN and PLS in quantitative modeling. The results show that, to a certain extent, PLS-RNN modeling algorithm in solving nonlinear problem is better than that of PLS and PCA-ANN modeling algorithm, but for the parameter design problem, it still need to do further research.5) The last chapter summarizes the innovation of this thesis. In the light of the deficiency of this thesis, a prospect was made for the research direction and the research emphasis in the future.