Optimization of Characteristic Wavelength Variables of Near Infrared Spectroscopy for Detecting Contents of Cellulose and Hemicellulose in Corn Stover
Liu Jin-ming,Chu Xiao-dong,Wang Zhi,Xu Yong-hua,Li Wen-zhe,Sun Yong
DOI: https://doi.org/10.3964/j.issn.1000-0593(2019)03-0743-08
2019-01-01
Spectroscopy and spectral analysis
Abstract:Pretreatment is an effective way to improve the utilization efficiency of the corn stover biotransformation. The conversion rate is directly related to contents of the cellulose and hemicellulose in corn stover during the bio-refinery conversion to biofuels. To achieve an effective control for the corn stover bio-refining process after the pretreatment, the near infrared spectroscopy (NIRS) was used to quickly detect contents of the cellulose and hemicellulose, solving the problems of being time consuming and high-cost in the traditional chemical analysis method. To improve the efficiency and precision of the NIRS detection, the genetic simulated annealing algorithm (GSA) based on genetic algorithm (GA) combined with simulated annealing algorithm (SA) was presented for optimizing the characteristic wavelength variables of NIRS. In the GSA, firstly, the number of the NIRS wavelengths was used as the code length for binary coding; secondly, the root mean square error of cross-validation (RMSECV) of the partial least squares (PLS) regression model was used as the objective function; thirdly, the fitness function was designed combining with the temperature parameter; and last, the selective replication of the perturbation solution was realized based on the Metropolis criterion. Therefore, GSA can effectively improve the search efficiency at the later stage of evolution while avoiding premature convergence. 120 samples of corn stover were prepared by using the pretreatments of alkaline, biology, and the combination of alkaline and biology. The contents of cellulose and hemicellulose were measured using the wet chemistry methods. The NIRS were collected using the Nicolet Antaris II Fourier near infrared spectrometer. The spectrum was pretreated by 7 points Savitzky-Golay smoothing combining with multivariate scattering correction and standard normal variate transformation. The samples were divided into correction set and validation set by using Kennard-Stone algorithm at a ratio of 3:1. The GSA is used for the characteristic wavelength variables optimizations of the NIRS whole wavelengths (Full-GSA), the synergy interval partial least squares selected spectral region (SiPLS-GSA), and the backward interval partial least squares selected spectral region (BiPLS-GSA), respectively. And then, the optimized results of the characteristic wavelength variables were evaluated by the PLS regressive model with the validation set. In Full-GSA, 1 557 wavelength points were used as chromosome genes in whole wavelengths, 118 cellulose characteristic wavelength points and 164 hemicellulose characteristic wavelength points were selected after 16 executions. In SiPLS-GSA, the cellulose and hemicellulose wavelength points of spectral region optimized by Si-PLS were 388 and 160, respectively, and 157 cellulose characteristic wavelength points and 148 hemicellulose characteristic wavelength points were gotten after the further optimization by GSA. In BiPLS-GSA, the cellulose and hemicellulose wavelength points of spectral region optimized by BiPLS were 358 and 180, respectively, and 130 cellulose characteristic wavelength points and 153 hemicellulose characteristic wavelength points were selected after the further optimization by GSA. It was shown that not only the number of wavelengths was significantly decreased after the optimization, but also the performance of regressive model was obviously better than that of the whole wavelengths. The best performance of regressive model for cellulose characteristic wavelengths was obtained by Full-GSA, and the best performance for hemicellulose characteristic wavelengths was obtained by SiPLS-GSA. The mean relative error (MRE) values of validation set for cellulose and hemicellulose in the best model were 1.752 4% and 2.020 8%, which were decreased by 13.636 6% and 25.368 4% compared with the whole wavelengths, respectively. The GSA combining with temperature parameters to design the fitness function is suitable for the NIRS characteristic wavelength selection of the cellulose and hemicellulose contents in corn stover, and has a good global search capability. The encoding scheme of GSA using each wavelength point in whole wavelengths as chromosome gene is suitable for the characteristic wavelength selection of NIRS whole spectrum. GSA is also suitable for the characteristic wavelength selection of the spectral region optimized by SiPLS and BiPLS, and the selection of wavelength points in the optimized spectral region can also be achieved effectively.