| With a rich nutrition value, Pork has been one of the most important meat food sources for Chinese people. The pork qualities like p H and water content have a direct relationship to the storage time, the following production and processing and edible safety. The multivariate calibration model based on spectral technology can actualize the fast and non-destructive detection of pork, however, the model can only be applicable to a certain measurement condition(instrument) or specimen space range(species), but poorly performed on the other conditions. So it has an important science significant and good application prospect to study on the model transfer methods and build a robust, strong applicability and high prediction accuracy pork quality detection model.This study was researched on 3 species of duchangda, maojia and linghao pork samples, pork quality quantitative detection models based on hyperspectroscopy were built, and then the applicability of the model was tested. In order to solve the problem of poor applicability, pork quality detection model transfer methods were researched in this study.The main results are as follows:1) The influence of different partition methods of sample set was researched, and the optimal methods were chosen separately for the pork p H and water content detection model. 3 methods RS, K/S and SPXY were adopted to divide the pork samples into calibration set and test set, then a better method was chosen according to the parameters like range, mean and standard deviation of the calibration and test sets.The result indicated that SPXY is the best partition method for p H detection of the 3 species of pork samples. For water content detection, SPXY performed better for duchangda and maojia samples, and K/S is the optimal method for linghao samples.2) The influence of different spectral pre-processing methods on the capability of pork p H and water content model was studied, and the superior pre-processing methods were identified. Methods such as smooth, multiplicative scattering correction(MSC), standard normalized variate(SNV), mean center(MC) and their combination methods were used to pre-process the original spectra, and the original spectra and the pre-processed spectra were used to build the Partial Least Squares Regression(PLSR) model, then the better pre-processing method was confirmed based on the model prediction results.For pork p H detection model, duchangda original spectra could establish the best PLSR model with correlation coefficient of cross-validation(Rc) of 0.924, correlation coefficient of prediction(Rp) of 0.904, root mean squared error of cross validation(RMSECV) of 0.045, root mean squared error of prediction( RMSEP) of 0.046.The combination method Smooth+SNV+ Auto was the best method for maojia PLSR model with of Rc 0.903, Rp of 0.853, RMSECV of 0.090 and RMSEP 0.084. The combination method Smooth+MSC+Auto was the best method for linghao PLSR model with of Rc 0.906, Rp of 0.883, RMSECV of 0.133 and RMSEP 0.113. The results indicated that none pre-process, Smooth+SNV+ Auto and Smooth+MSC+ Auto were the best spectra pre-processing methods for duchangda, maojia and linghao pork samples separately.For pork water content detection model, mean center(MC) was the best pre-processing method for duchangda PLSR model with Rc of 0.940, Rp of 0.940, RMSECV of 0.279%, RMSEP of 0.237%. Autoscale was the best method for maojia and also linghao PLSR models with of Rc of 0.932 and 0.927 separately, Rp of 0.944 and 0.923, RMSECV of 0.512% and 0.512%, RMSEP of 0.395% and 0.382%. The results showed that MC, Autoscale and Autoscale were the best spectra pre-processing methods for duchangda, maojia and linghao pork samples separately.3) The hyperspectral detection model applicability was studied, and it identified that duchangda p H and water content hyperspectral model cannot be applied to predict the p H and water content of maojia and linghao samples. The applicability of duchangda model was tested by the mean spectrum, PCs scores spatial distribution, mahalanobis distance and model verification methods, the results indicated that there were big differences between duchangda and maojia, linghao samples, and the performance was very poor when duchangda p H and water content model was applied to predict maojia and linghao samples. The results of maojia and linghao samples predicted by duchangda p H detection model were Rp of 0.770 and 0.731, RMSEP of 0.111 and 0.209, RPD of 1.533 and 1.234 separately. The results of maojia and linghao samples predicted by duchangda water content detection model were Rp of 0.513 and 0.712, RMSEP of 1.151% and 0.857%, RPD of 1.000 and 1.214 separately.4) Model updating method was used to solve the problem of pork p H model transfer, and new model transfer methods of s ync correction of spectrum and prediction value(CSPV) and piecewise direct standardization combine with linear interpolation( PDS-LI) were proposed, and the superior methods were identified according to the performance of the 3 methods to ducahngda p H model.When model updating method was used for duchangda p H model transfer, the results were different between maojia and linghao samples. For maojia samples, with 11 representative samples were added to duchangda model, the optimal result was Rp increased from 0.770 before transfer to 0.869; RPD increased from 1.533 to 1.934, which met the condition of Rp ≥0.837, RPD ≥1.9, indicated that model updating could be applied to duchangda model transfer to maojia samples. For linghao samples, with 9 representative samples were added to duchangda model, the optimal result was Rp increased from 0.731 before transfer to 0.845; RPD increased from 1.234 to 1.804, which could not meet the condition of Rp ≥0.837, RPD ≥1.9, indicated that model updating was not applicable to duchangda model transfer to lilnghao samples.To solve the problem of model transfer between different species, a new method, sync correction of spectrum and prediction value(CSPV) was proposed, and was proved to be effective for duchangda p H model transfer. For maojia samples, with only 9 standard samples, duchangda p H model could be applied to predict maojia samples with the optimal result of Rp of 0.889, increased by 15% over the result before transfer, RPD of 2.071, increased by 35%; For linghao samples, with only 10 standard samples, duchangda p H model could be applied to predict linghao samples with the optimal result of Rp of 0.900, increased by 23%, RPD of 2.213, increased by 79%.To solve the problem of model transfer between different species, a new method, piecewise direct standardization combine with linear interpolation( PDS-LI) was proposed, and was proved to be effective for duchangda p H model transfer. For maojia samples, when the standard samples raised to 29, duchangda p H model could be applied to predict maojia samples with the optimal result of Rp of 0.895, increased by 16% over the result before transfer, RPD of 2.179, increased by 42%; For linghao samples, when the standard samples improved to 22, duchangda p H model could be applied to predict linghao samples with the optimal result of Rp of 0.892, increased by 22%, RPD of 2.005, increased by 62%.The model transfer results showed that the prediction of the 2 methods CSPV and PDS-LI was similar, and both of them were better than model updating method, so it identified that CSPV and PDS-LI methods were the superior methods for pork p H hyperspectral model transfer.5) The problem of pork water content hyperspectral model transfer was researched in this study. Model updating, CSPV and PDS-LI methods were applied to transfer duchangda pork water content model, and the superior methods were identified according to the performance of the 3 methods.Duchangda model could be used to predict maojia and linghao samples by the transfer of model updating method. For maojia samples, with 21 representative samples were added to duchangda model, the optimal result was Rp of 0.891, increased by 74% over the result before transfer, RPD increased to 2.166, increased by 216%; For linghao samples, with 19 representative samples were added to duchangda model, the optimal result was Rp increased to 0.853, increased by 20%, RPD increased to 2.076, increased by 71%.CSPV model transfer method was proved to be effective for duchangda water content model transfer. For maojia samples, with only 20 standard samples, duchangda water content model could be applied to predict maojia samples with the optimal result of Rp of 0.918, increased by 79% over the result before transfer, RPD of 2.460, increased by 246%; For linghao samples, with only 18 standard samples, duchangda water content model could be applied to predict linghao samples with the optimal result of Rp of 0.925, increased by 30%, RPD of 2.379, increased by 96%.PDS-LI model transfer method was proved to be effective for duchangda water content model transfer. For maojia samples, with only 32 standard samples, duchangda water content model could be applied to predict maojia samples with the op timal result of Rp of 0.912, increased by 78% over the result before transfer, RPD of 2.447, increased by 245%; For linghao samples, with only 22 standard samples, duchangda water content model could be applied to predict linghao samples with the optimal result of Rp of 0.921, increased by 29%, RPD of 2.364, increased by 95%.Results showed that, the 3 methods could improve the predict precision of duchangda model to maojia and linghao samples in different extent, and achieved the duchangda water content model transfer, it showed that both CSPV and PDS-LI methods were overmatch the model updating method, and with the similar transfer results, CSPV and PDS-LI methods were identified to be the superior methods for pork water content hyperspectral model transfer. |
PDF Full Text Request