Information Extraction in Frequency Domain Based on Entropy Theory and Genetic Algorithm in Near-Infrared Spectra

In order to efficiently extract the analyte information in near infrared (NIR) spectra, a new algorithm, which is named as GA-FD-SEL, was proposed by signal processing in frequency domain. At first, to take advantages of multiscale property of NIR spectra, wavelet packet transform (WPT) algorithm was adopted to split the raw spectra into a series of frequency components in wavelet domain. The optimum decomposition level was determined based on the entropy contribution of each components. Then the set of WPT components was initially selected according to entropy contribution again, and further selected using genetic algorithm (GA). At last, based on the set of filtered-components, the multivariate regression models were built using the partial least squares (PLS) regression method. To validate the GA-FD-SEL algorithm, it was applied to construct the prediction model for measuring the oil content of corn. Experimental results showed that the GA-FD-SEL algorithm can greatly improve the prediction ability of NIR regression models and the prediction errors could decrease by up to 50.2%, indicating that it could efficient find the useful WPT components where analyte information resides.