A novel spectral multivariate calibration approach based on a multiple fitting method.

This paper introduces a novel multivariate regression approach based on a multiple fitting algorithm that combines fitting functions to accordingly configure different regression models for the quantitative analysis of spectra data. The novel regression method employs multivariate fitting functions to characterize the potential functional relationship of spectral information and sample information with the corresponding attributes, and uses single fitting function elements as the independent variables for biased parameters to determine the amplitude of each fitting function. The peak width of the different fitting functions and the biased parameters are optimized by a simulated annealing algorithm. After parameter optimization, the fitting functions are superposed together to form a functional super surface, and a multiple fitting regression model is then used to characterize any functional relationship among the spectral variable information, sample information, and the corresponding analyte information. In this paper, a common fitting function, namely the Gaussian function, is used to create multiple fitting regression models. The simulated dataset and two real near infrared spectral datasets were used to validate the multiple fitting regression model. The results are compared to those obtained using partial least squares regression and least squares support vector regression. It is shown that the proposed multiple fitting regression algorithm achieved an accuracy comparable to the two conventional methods. Therefore, the multiple fitting regression is demonstrated to be a useful tool for spectra multivariate regression analysis and may also be suitable for linear and nonlinear multivariate calibration.