The Research On Quality Analysis For Honey By Near Infrared Spectroscopy

Honey is a natural health product with high nutritional value. Quality assessment of honey includes following sections:quantitative analysis of physical and chemical parameters, classification of botanical and geographical origin and detection of adulteration. It is necessary to looking for fast and accurate methods to detect these indexes so as to assess honey quality more effectively. As a new analytical method, near infrared spectroscopy (NIRS) have advantages like highly efficient, fast, low cost and green environmental protection, and so on. Therefore, it is significant to develop NIRS methods to detect honey quality, which will benefit healthy development of honey industry and protect consumers’rights and interests. The dissertation mainly studies:(1) The different effects of three measurements with thin film transflectance, sample cup transflectance and cuvette transmission on the repeatability of spectra of honey sample were studied. The results showed that both sample cup transflectance and cuvette transmission had better repeatability than thin film transflectance. Further studies results showed that the transflectance sample cup performed better than cuvette transmission in measurement.(2) The different effects of scanning frequency and resolution on near-infrared spectra of honey samples and the prediction abilities of models were studied. For honey sample, the optimized parameters combination was32times of scanning frequency and8cm-1of resolution.(3) The quantitative analytic models of12different physical and chemical parameters, soluble solids content (SSC), moisture, reducing sugar (RS), pH-value, total acidity (TA), electrical conductivity (EC), fructose, glucose, sucrose, maltose, fructose/glucose ratio (F/G) and glucose/moisture ratio (G/M) in honey were developed by partial least squares (PLS) regression. Satisfying prediction accuracies were achieved for SSC, moisture, RS and F/G:Root mean square error of prediction (RMSEP) for each of the parameters was respectively0.1795,0.1696,1.5270and0.0344; Coefficient of determination in prediction (Rp) set was0.9989,0.9989,0.9191and0.9749, respectively. After rejection of the outliers by Monte-Carlo cross-validation (MCCV), competitive adaptive reweighted sampling (CARS) combined with PLS regressions were used to choose effective variables so as to optimize PLS models of8other parameters. The results showed that the prediction abilities of PLS models were obviously improved. Using the optimal PLS models, RMSEP of pH-value, TA, EC, fructose, glucose, sucrose, maltose and G/M in honey were0.1196,0.4674,2.6827,0.4955,0.5704,0.5711,0.2578and0.0394, respectively, and Rp were0.9058、0.9083,0.9679,0.9845,0.98790.9386,0.9586and0.9890, respectively.(4) Three models were developed for detection of five different botanical origin honey samples by Mahalanobis distance-discriminant analysis (MD-DA), partial least squares-discriminant analysis (PLSDA) and radical basis function neural networks (RBFNN) in the NIR region of4200～5400cm-1. The total prediction accuracy (PA) of both MD-DA model and RBFNN model can reach94.0%. However, PA of PLSDA model was only78.0%. Therefore, NIRS combined with MD-DA or RBFNN have a potential for quickly detecting botanical origin of honey.(5) The spectral data were compressed and de-noised using wavelet transform (WT). RBFNN and PLSDA were applied to develop classification models of the geographical origin of honey samples using either before or after the reconstructed signals, respectively. For apple honey samples, PLSDA, WT-PLSDA, RBFNN and WT-RBFNN produced a same total prediction accuracy of96.2%. For rape honey samples, PLSDA, WT-PLSDA, RBFNN and WT-RBFNN produced total prediction accuracy of86.4%,90.9%,81.8%and86.4%, respectively. The results showed that prediction accuracy varied widely in the different geographical origin honey samples when modeling method were the same. Linear WT-PLSDA model may be more suitable for geographical classification of honey samples than no-linear WT-RBFNN model. NIRS has a potential for quickly detecting geographical origin of honey samples.(6) Authentic honey samples were adulterated with high fructose corn syrup (HFCS), beet syrup (BS), maltose syrup (MS) and a blend of these syrups. Detection of honey adulteration was developed by distance match (DM), MD-DA and PLSDA. The best performances of classification of authentic and adulterated honeys were obtained by PLSDA:total accuracy for validation sets were82.4%,90.2%,90.2%and80.4%, respectively. The classification of types of adulterants was also developed using the above methods, and the highest accuracy of84.4%was obtained by both DM and MD-DA for MS adulterant. The quantitative analysis of adulterants by PLS regression gave satisfying results if adulterated honey samples were got from the same one authentic honey sample, otherwise it gave dissatisfying results for the adulterated samples from different botanical origins, except the adulteration is solo MS.(7) Honey samples were adulterated by blend of fructose and glucose. Near infrared spectrum of honeys were collected before and after adulteration. The spectral data were compressed and de-noised by WT. The radial basis function (RBF) was used as kernel function of WT-LSSVM model. By grid search, the selected optimal value of γ(the relative weight of the regression error) and σ2(the kernel parameter of the RBF kernel) was222.822and45.170respectively. The total accuracy of WT-LSSVM model was95.1%for the validation set, which was better than that of support vector machines (SVM), back propagation artificial neural networks (BPANN), K-nearest neighbor (KNN) and linear discriminant analysis (LDA). This study demonstrated that the superiority of LSSVM in building models with better generalization abilities than those obtained from SVM, BPANN, KNN and LDA for the problems studied. NIRS has the potential ability to quickly detect mixed sugar adulterants in honey.