Study On Application Of Near Infrared Spectrum Based On Kernel Method

Near Infrared Spectrum (NIRS) has been widely applied in many fields since its extraordinary advantage, such as high-efficiency analysis, lower cost, and fast detection, etc. Near-infrared spectroscopy measures the composition of samples according to absorption characteristics of the near-infrared spectral region. Therefore, the focus of near-infrared spectroscopy approach is the establishment of a function relationship between the chemical compositions and the data of absorption spectra. This modeling approach has become the key point of the research. This paper presents a near-infrared spectroscopy regression modeling method, i.e. partial least squares principle, and by this approach, we established a calibration model of the test sample. However, the quantitative function relationship between the chemical composition and absorption spectra is always nonlinear, and traditional linear modeling methods are affected by this nonlinear relationship which will reduce predictive capability for the model. To this end, this paper proposed a nonlinear modeling method for supporting vector regression based on a kernel method. This paper also used the grid search method and genetic algorithm-based method for determining how to choose parameters. In addition, we established the calibration model using the same test sample, and compared the predictive capabilities of two models. Finally, a conclusion can be drawn that the nonlinear modeling method for supporting vector regression based on a kernel method has the features of high accuracy and popular promotion.