| The requirement of consumer to pork quality is increasing with the improvement of living standards. Pork freshness is one of the most important indicators of pork quality and safety. The process of traditional detection methods is cumbersome, time-consuming, so it is not possible to achieve rapid and non-destructive testing. A rapid non-destructive testing method makes sense to improve people’s life quality. Spectral analysis techniques had been used in detecting the pork freshness, however the results were seriously impacted by the difference of pork breeds.Hubei white pig, Du changda and the Enshi mountain boars had been selected as samples in this paper, and near-infrared spectroscopy data and hyperspectral data were used to establish freshness index detection model. By comparing the different de-noising methods, spectral preprocessing methods and model transfer methods to determine the best spectrum detection methods and model transfer method of pork freshness. These will provide further theoretical and technical support to carry out more in-depth studies of pork freshness. The main results and conclusions were listed as follows:1) S-G smoothing in different modes and different decomposition scale wavelet decomposition reconstruction methods were studied in the spectral signal de-noising to determine the optimal de-noising method. The result shows that, for near-infrared spectroscopy and hyperspectral data, different modes of SG smooth de-noising made a small change of model performance, and result was unstable, so it is difficult to evaluate which is the optimal noise cancellation method; Wavelet decomposition reconstruction method with decomposition scale of7db3established a better model performance. Chose Signal-to-noise ratio as the evaluation indicator, compared different modes of the SG smoothing methods and different wavelet decomposition and reconstruction methods, the results show that, for the near-infrared spectroscopy and hyperspectral data, decomposition scale of7db3wavelet decomposition reconstruction method is the best de-noising method.2) The method combined partial least squares with Monte Carlo had been used to test abnormal samples in different varieties of pork samples based on Near-infrared spectral data and high hyperspectral data. For the Near-infrared spectral data, three samples numbered17,43,82in Hubei white pig sample set, three samples numbered50,84,94in Du changda sample set and two samples numbered87,89in Enshi mountain boar sample set are abnormal. For the high spectral data, three samples numbered4,18,59 in Hubei white pig sample set, three samples numbered73,80,81in Du changda sample set and three samples numbered29,63,94in Enshi mountain boar sample set are abnormal.3) The influence of different pretreatment methods was compared on modeling of pork samples. The modeling effect after preprocessing methods of variable center of centralized, standardized, multiplicative scatter correction, normalization, vector normalization, first derivative and orthogonal signal correction were compared, for Hubei White Pig, Du changda and Enshi mountain boar samples, the optimal pretreatments were None, None, Autoscale respectively based on near infrared spectroscopy and Autoscale, Mean Center, Normalize based on hyperspectral data.4) The feasibility of generic model created based on three varieties of mixed pork samples was studied. Partial least squares regression (PLSR) algorithm was used to create spectrum quantitative analysis models, for generic model of Near Infrared Spectroscopy, the root mean square error of prediction of Hubei white pig samples, Du changda samples and Enshi mountain boar samples are3.533,4.918and3.728, for generic model of hyperspectral model, the root mean square error of prediction of Hubei white pig samples, Du changda samples and Enshi mountain boar samples are1.732,2.584and2.944.The results show that the generic model established using three varieties of mixed samples has a better prediction for each species pork sample.5) An improved model transfer method based on S/B algorithm was proposed on the basis of model transfer between the instruments, and the model transfer effects of different varieties of pork by this method were studied. After different varieties of pork models were transferred by this method, R2of the models were decreased, and the results were stable. Among them, after Du changda model was transferred by Hubei White NIR model, RMSECV was dropped from23.465to3.167, model transfer effect was ideal.6) An improved model transfer method based on PDS algorithm was proposed on the basis of model transfer between the instruments, and the model transfer effects of different varieties of pork by this method were studied. The results showed that:the result of this improved model transfer method was unstable. Among which, after Du changda model was transferred by Enshi mountain boar model, RMSEP of model was significantly increased, which indicated that this method was not applicable to transfer Du changda model by Enshi mountain boar model.7) A model transfer method of wavelength screening based on regression coefficients was proposed, and the model transfer effects of different varieties of pork by this method were studied. The results showed that:the results were stable. After Hubei White Pig model and Enshi mountain boar model transferred by Du changda NIR model, and Hubei White Pig model and Du changda model transferred by Enshi mountain boar NIR model, RMSEP of the models were respectively dropped from18.215,5.947,5.749and5.695to3.989,2.963,3.891and3.701, R2were increased from0.666,0.312,0.394and0.075to0.701,0.827,0.720and0.516; For hyperspectral data, the RMSEP of Du changda model transferred by Hubei White Pig model, Hubei White Pig model and Enshi mountain boar model transferred by Du changda model, and Hubei White Pig model and Du changda model transferred by Enshi mountain boar model were decreased from14.024,11.039,4.624,11.482and5.993to1.876,1.733,2.365,2.046and2.117, R2were increased from0.007,0.461,0.528,0.132,0.003to0.776,0.943,0.876,0.927and0.708. These results were similar to the prediction effect of generic model, which indicated that this method could be used in different varieties of pork model transfer. |
PDF Full Text Request