Fundamental Research On Novel Modeling Methods And Applications Of Near Infrared Spectroscopy

The quality control requirements for product are increasingly stringent in modern industries, however, the acquisitions of feedback information about quality of product are the basis of product quality control. Therefore the measurements of product quality are especially important in product quality control system. Near infrared (NIR) spectral analysis technology can realize quantitative/quality analysis of product quality and has many prominent advantages, including rapid speed, capacity of nondestructive and real-time online analysis and so on. Nowadays, NIR has been used in a broad range of areas. However, NIR spectral signal is relatively weak and spectral bands are wide and highly overlapping, these disadvantages make the quantitative/qualitative information about product quality cannot be acquired directly from the NIR spectral signal. In practice, NIR can only be used as an indirect analytical technology and the applications of NIR must rely on the effective quantitative/qualitative calibration models. Therefore, the modeling methods always are the core of NIR technology and have been the research emphasis and hot topic in the field of NIR during the past ten years. Although there have been great progresses in the theoretical system of NIR modeling method with respect to earlier stage, the research about this field is still in the development stage and it is still necessary to continue to explore and develop novel NIR modeling methods to further improve the precision of NIR technology and make contribution to the theoretical system of NIR modeling method. In addition, the research about the applications of NIR is also very important in the field of NIR technology and has attracted the attention of many scholars.Under the above background, this dissertation focus on exploring and developing novel NIR modeling methods and carrying out the research about the applications of NIR technology in real life. The main innovations and contributions in this dissertation are listed as follows:(1) Carry out research on qualitative analysis of NIR technology. In view of the fact that phenomenon of brand fakes of washing powder and type confusion and adulteration of polyacrylamide often happens on the market at present, explore the feasibility of NIR technology coupled with pattern recognition methods in the applications of brand discrimination of washing powder and type discrimination of polyacrylamide. In addition, carry out a preliminary study about the usage of NIR technology in discrimination of the adulterated polyacrylamide. The experimental results verify the feasibility of NIR technology coupled with pattern recognition methods in the discrimination of brand of washing powder and type of polyacrylamide. And NIR technology can also make accurate identification of adulterated polyacrylamide.(2) In NIR spectral quantitative analysis, too significant differences among samples always lead to serious nonlinearity between spectra and testing target of samples. Performance of commonly used partial least square (PLS) model will deteriorate in this case. To address this issue, three novel local PLS modeling methods are proposed. The first one is local PLS model based on net analyte signal, in this method, obtain the net analyte signal of all calibration samples and unknown sample, then the Euclidean distance between net analyte signal of the each unknown sample and net analyte signal of all calibration samples are calculated and utilized as similarity index to choose local calibration sets. Finally, a local PLS regression model was built on each local calibration sets for each unknown sample. The second one is local PLS modeling method based on spectral regression method, spectra of samples are compressed by spectral regression method firstly, then similarity index in this method is defined as the Euclidean distance between compressed spectra of the each unknown sample and compressed spectra of all calibration samples. The last one is local PLS modeling method based on spectral information divergence, this method introduces the spectral information divergence into local modeling methods for the first time to be used as a novel sample similarity criterion. The proposed methods are applied to a set of near infrared spectra of meat samples, experimental results show that the prediction precision and model complexity of the proposed methods are all superior to global PLS regression method and two conventional local modeling methods.(3) Propose an improved stacked PLS modeling method based on variable importance in the projection (VIP) in order to improve the performance of the existing stacked PLS method, and this method is named as VTP-SPLS. This method calculates VIP score of each wavelength based on conventional PLS model firstly, then sorts the spectral wavelengths according to the VIP score in descending order and splits the sorted spectra into several disjoint intervals with equal width, then builds one PLS sub-model on each interval and obtains the weight of each PLS sub-model based on cross validation, finally all these sub-models are combined together in the form of weighted average to obtain the stacked model. The proposed method is applied to two different sets of NIR spectra, experimental results indicate that the prediction precision of the proposed method is superior to conventional PLS method and existing stacked PLS modeling method.(4) Propose an improved extreme learning machine (ELM) modeling method to improve the performance of ELM model in NIR spectra quantitative analysis, and this method is named as iELM. In conventional ELM modeling method, all NIR wavelengths are used as inputs of ELM model and mapped into the output matrix H of hidden layer by the excitation function firstly, then Moore-Penrose generalize inverse method is used to build the regression model between the output matrix H and testing target. However, in NIR spectral quantitative analysis, ELM model always needs to incoprate too many hidden nodes due to the high dimension of NIR spectral data. Thus the problems of high dimension and high collinearity in the output matrix H are inevitable. In this case, the solutions of the regression model between H and testing target calculated by Moore-Penrose generalize inverse method should be ill-conditional thereby deteriorating the performance of ELM model. In iELM method, the newly proposed VIP-SPLS method is used to replace the Moore-Penrose generalize inverse method in conventional ELM model to build the regression model between H and testing target. The proposed method is applied to a commonly used benchmark NIR spectral data for evaluation. The results demonstrate that the precision of iELM model is superior to PLS model and conventional ELM model.(5) In Near Infrared spectroscopy quantitative analysis, model transfer is a technology which is commonly used to keep the universality of the built calibration model. Generally, model transfer can remove or decrease the spectral differences among instruments through various methods of spectral correction, thus the calibration model built on one instrument (called as master instrument) can apply to other instruments (called as slave instruments) with satisfactory prediction accuracy. Existing calibration transfer methods may be suffered from the redundant information in spectral data, and the prediction accuracy for spectral measured on slave instruments may also be affected under this circumstance. To solve this potential problem, this research combines two conventional calibration model transfer methods with recently proposed VIP-SPLS algorithm and proposes two new calibration model transfer methods. VIP-SPLS can overcome the negative effects of redundant information on calibration model and has the advantage of model fusion, the experimental results demonstrate that the prediction accuracy for spectra measured on slave instruments by the transferred calibration model built on master instrument can be further improved when conventional calibration model transfer methods are combined with VIP-SPLS algorithm.(6) The novel modeling methods proposed in this dissertation are applied to predict the octane number of gasoline by near infrared spectroscopy technology, in order to further improve the prediction accuracy of near infrared spectroscopy for octane number of gasoline and further evaluate the practicability and validity of the novel modeling methods in actual industrial production.