| Meat occupies an important position in Chinese dietary structure, while pork is dominant in the meat consumption for a long time. Therefore, pork safety is a crucial issue in people’s livelihood, and study on pork quality detection has important realistic meaning for pork quality safety control and residents’ food safety safeguard. At present, meat quality detection methods are mostly sensory evaluation, physical and chemical analysis which are not conductive to rapid detection of pork products in circulation. In recent years, hyperspectral imaging technology provides an alternative methodology in agricultural and livestock product nondestructive detection field, because it has the advantages of high resolution, no need of sample pretreatment, easy operation, nondestructive and so on. Some achievements have been made in meat quality nondestructive detection research by utilizing hyperspectral imaging technology, however further research is still needed. Firstly, hyperspectral imaging can obtain the internal spectral information of samples while detecting spatial signals. These signals are kept in huge data cube which may slow down the data processing speed. Secondly, data processing methods may influence the accuracy of the multivariate model. Thirdly, changes of some factors, such as sample variety and so on, may weaken the adaptability of the multivariate model. Therefore, study on nondestructive detection methods can improve detection accuracy and speed, and enhance the adaptability of the multivariate model while hyperspectral image technology is utilized to detect pork quality. The research shows important scientific significance and broad application prospect for the further practical promotion of pork quality nondestructive detection by utilizing hyperspectral image technology, for the development of multispectrum online detection system and portable detection equipment, and for the promotion of the industrialization process of pork production.In this paper, cold fresh pork was taken as the research object, and its water content, p H value and TVB-N content were taken as the evaluation indices. Data processing methods, such as outlier detection, sample set partition, spectral preprocessing, modeling, optimum wavelength selection, and model maintenance, were studied by utilizing chemometrics, mathematical statistics, machine learning theory and computer technology comprehensively. The study contents and conclusions are as follows:1) The outlier Monte Carlo second detection method was proposed. The results of cold fresh pork outlier detection by three methods, such as the leverage-studentized residual T test, the Monte Carlo method and the outlier Monte Carlo second detection method, were compared. Then the outlier Monte Carlo second detection method was determined as the best method for cold fresh pork outlier detection.2) The results of cold fresh pork sample set partition by five methods, such as KS, SPXY, duplex, RS and CG, were compared. Then CG was determined as the best method for the No.0 indigenous pork water content sample set and enshi mountain pork TVB-N content sample set, SPXY was determined as the best method for the No.0 indigenous pork p H value sample set.3) A fast method for choosing the best spectral pretreatment was proposed. Then "MSC+FD+AS" was determined as the best preprocessing method for the No.0 indigenous pork water content PLSR model, and "normalize+OSC+MC" was determined as the best preprocessing method for the No.0 indigenous pork p H value PLSR model, and MC was determined as the best preprocessing method for the enshi mountain pork TVB-N content PLSR model.4) The performances of four kinds of model, such as MLR, PCR, PLSR, and SVR, of cold fresh pork water content, p H value, TVB-N content were compared, respectively. Then PLSR was determined as the most suitable model for cold fresh pork quality detection.5) The abilities of selecting the important wavelengths by five algorithms, such as PCA, UVE, SPA, GA and CARS, were compared. Then CARS was determined as the best important wavelength selection method6) The applicability of model updating method to different detection indices of cold fresh pork was studied. Current research on model updating method mostly emphasizes the improvement of prediction ability of updating model on the new samples measured in new conditions, such as new temperature, new instrument etc. However, further research is needed on the degree of the improvement, and the effection of model updating on the original samples. In this paper, the updating results of models of cold fresh pork water content, p H value and TVB-N content were compared, and the applicability of model updating method to different detection indices of cold fresh pork was determined. Tests show that the updating results of different detection index models vary greatly. Model updating method has very good maintenance ability for TVB-N content model, but it has very poor maintenance ability for water content model and p H value model.Therefore, the following conclusions can be drawn that model updating method is suitable for the maintenance of TVB-N content model, but not suitable for the maintenance of water content model and p H value model.7) A new model transfer algorithm called variety sensitive wavelength selection combining with piecewise direct standardization(VSWS-PDS) was proposed to improve the applicability of cold fresh pork water content and p H value models. At present, researches on model transfer algorithm are mostly about near infrared spectroscopy and among different instruments, while research on hyperspectral model transfer algorithm between different varieties is relatively less.In this paper, VSWS-PDS model transfer algorithm is put forward to transfer the cold fresh pork hyperspectral model between different varieties with fully considering the influence of variety difference on cold fresh pork quantitative analysis model.After the water content model was transferred from the master variety(No.0 indigenous pork) to the slave variety(enshi mountain pork), the Rp2 was enhanced to 0.8570 from 0.3795, and the RPD was enhanced to 2.67 from 1.05.And after the p H value model was transferred from the master variety(No.0 indigenous pork) to the slave variety(enshi mountain pork), the Rp2 was enhanced to 0.6821 from 0.3833, and the RPD was enhanced to 1.53 from 0.87. The results show that VSWS-PDS model transfer algorithm can effectively eliminate the influence of variety difference on the cold fresh pork water content and p H value models. |
PDF Full Text Request