Similarity based spectral data fusion physical parameter regression modeling method

Abstract

Spectroscopy are widely used in routine concentration measurement. However, when spectral measurement is carried out in process industry, the measurement environment often changes thus the prediction accuracy of the regression model is spoiled. Existing studies regard the measurement environment change as noise, but in fact, the measurement environment also contains useful information. In this paper, a modeling method is proposed to augment the measured environmental parameters (physical quantities) into the calibration modeling to improve the prediction accuracy. To solve the problem of physical quantity parameters being overridden caused by direct variable extension method, we use the data fusion method based on sample similarity. Gaussian kernel function is used to calculate the similarity matrix of spectral and physical quantities respectively. Then fusion matrix is obtained by weighting combination. Finally, the regression model of fusion matrix and concentration is established by standard PLS modeling method. A regression model is established for the data collected during the fermentation process. The results showed that the prediction performance of the model could be improved by nearly 10 % by adding physical quantity information.