| In this paper, after the introduction of the fundamental principles, routine testing methods and steps of analysis of the Near-Infrared (NIR) spectroscopy, and the current status of the technique and equipment of on-line non-destructive determination of the internal quality of agricultural products by visible/Near-Infrared (Vis-NIR) spectroscopy, non-destructive determination of fruit sugar content (SC) using Vis-NIR spectroscopy technology and chemometrics knowledge was studied, the research contents mainly including: research of sugar content determination of fruit with simple and complex sample temperature variation, research on non-destructive prediction of the surface color and SC of fruit synchronously, study on the prediction of SC using the combined spectral and different colors information of fruit, study a variety of multivariate calibration methods based on variable selection, research of the optimization of real-time detection system., development of the real-time non-destructive inspecting system for fruit sugar content, development of fruit internal quality evaluation and real-time classification software using Visual C++6.0 software and OmniDriver 2nd development package, and study of model updating and model correction for industry application.The main results and conclusions were:(1) The fundamental principles, routine testing methods and analysis steps of Visible-NIR spectroscopy were introduced, and the key issues-the model robustness in Visible-NIR industrial application were plagued.(2) The influence of sample temperature on the prediction accuracy of fruit sugar content with simple temperature variation was studied. The optimal prediction results of sugar content of honey peaches mixed with room temperature and low-temperature samples, root mean square standard error of cross-validation (RMSECV) 0.67°Brix, and 0.532°Brix, were obtained using partial least squares (PLS) regression and stepwise multiple linear regression (SMLR), respectively. Results showed that SMLR was better than PLS to predict SC of samples with simple temperature variation by establishment of analytical model with robust wavelength selection, which was not sensitive to the temperature. In addition, we found that the single-temperature calibration results were better than the mixed, RMSECV=0.26°Brix for samples with room temperature, RMSECV=0.31°Brix for samples with low-temperature. This shows that the impact of temperature can induce significant errors in NIR predictions.(3) Two-dimensional correlation analysis between sample temperature and spectroscopy. By two-dimensional correlation spectroscopy analysis, it was found that the group bands around 850 nm,790 nm and 690nm peaks were sensitive to disturbances of temperature, and the OH group was more sensitive to the temperature perturbation than the CH group, and this evidence on the two-dimensional correlation spectroscopy of samples with high SC was stronger than that of samples with low SC.(4) The prediction of fruit SC with complex sample temperature variation was studied. By compared with 5 calibration model:SMLR, PLS, genetic algorithm-partial least squares (GA-PLS), least squares-support vector machines (LS-SVM 1) (only 30-band spectra information obtained by SMLR analysis inputted as X variable) and LS-SVM2 (both 30-band spectra and temperature inputted as X variable), the results showed that the prediction performance was in the order of LS-SVM2, LS-SVM1, GA-PLS, PLS and SMLR. It was concluded that SMLR approach was become inadequate for complete temperature compensation under complex temperature changes, GA-PLS model established on selected band using genetic algorithm (GA) obtained a good prediction than the full-band PLS model, while the LS-SVM as a nonlinear model can effectively compensate the non-linear effects on spectrum caused by temperature, and LS-SVM2 including the temperature into X-matrix, to a certain extent, compensate the effects of temperature and increase the prediction accuracy.(5) Non-destructive determination of fruit surface color and SC using Vis-NIR spectroscopy synchronously was studied. SC and color values of CIE L*, a*, b* and C* were predicted by SMLR model built on 20 wavelengths acquired by a portable spectrometer, the best results were:coefficient of determination of calibration Rcal2=0.937, RMSEC=0.294°Brix, coefficient of determination of cross-validation Rcv2=0.896, RMSECV=0.3794°Brix, and Ratio performance deviation (RPD)=3.2 for SC; Rcal2=0.968, RMSEC=0.472, Rcv2=0.948, RMSECV=0.603, and RPD=4.5 for color index L*; Rcal2=0.939, RMSEC=0.593, Rcv2=0.899, RMSECV=0.765, and RPD=3.3 for color index a*; Rcal2=0.950, RMSEC=0.573, Rcv2=0.918, RMSECV=0.738, and RPD=3.5 for color index b*; Rcal2=0.958, RMSEC=0.548, Rcv2=0.930, RMSECV=0.708, and RPD=3.8 for color index C*. The RPD values of these 5 quality indexes were bigger than 3 that indicated the models are good enough to performance.(6) Vis-NIR based technique coupled with color features for identification SC of pear fruits was studied. The results showed that the predicting results of LS-SVM regression with both the 15-band spectrum information (first input) and color information (second input) as the input X variables was higher than that of only using the same 15-band spectrum information as input by SMLR regression and LS-SVM regression methods, and coupled with all color information (L*+a*+b*), the best result was obtained, RMSEP= 0.561°Brix, and others in order were (a*+b*)> L*. The results suggested that it was technically feasible to detect fruit SC and improve its prediction accuracy by color and spectral information fusion. In addition, this can be realized in industrial applications.(7) A variety of feature variable selection based modeling methods were studied. SMLR, GA-PLS, continuous projection algorithm-multiple linear regression (SPA-MLR) and synergy interval PLS (siPLS) were used to on-line analysis of fruit SC. SMLR model based on 18 wavelengths obtained the optimal prediction results of R2=0.849, RMSEP=0.346°Brix, in which the selected bands was 2% of the full spectral range between 520 nm and 920 nm within 901 spectral points. GA-PLS model built on selected 225 wavelengths achieved the optimal prediction results of R2=0.873, RMSEP=0.315°Brix, in which the selected bands was 25% of full spectral range. SPA-MLR model built on selected 11 wavelengths obtained the optimal prediction results of R2=0.808, RMSEP=0.391°Brix, in which the selected bands was 1.22% of full spectral range. The siPLS method divided full spectral range (520 nm-920 nm) into 20 intervals, and the model established with [7,10,12,15] intervals achieved optimal prediction results of R2= 0.832, RMSEP= 0.362°Brix, in which the selected spectral range was 20% of full spectral range. The results showed that although the GA-PLS predicted the best results, its modeling used maximum spectral information of 25%, and required the longest time, however, SPA-MLR model used only 1.22% spectral information, but its prediction accuracy was relatively poor, so according to the prediction accuracy, the required time and the spectral information used, SMLR performance can meet industrial applications.(8) Optimized the real-time detection system was studied, including the comparisons and selections of different reference materials, different spectrometers, and different fruit cups, to ensure the reliability of detection system and model. The results showed that ND0.7 reference material and spectrometer type A (slit size=100μm) are good enough for detection system and modeling, and in this detecting system the influence of different fruit cups can be ignored.(9) Research and development of on-line detection system for fruit SC, and model updating and correction methods for industrial application were studied. The results showed that the model updating method was superior to model correction method to reduce the predicting error caused by the sample changes.
|
PDF Full Text Request