Study On Maintenance Methods Of Freshness Detection Model For Beef Based On Spectrum Technology

As the basis of spectral quantitative analysis, whether a multivate calibration model has a stable prediction performance directly affects the precision and accuracy of spectral analysis. Generally, a calibration model contains lots of information about the characteristics and preparation conditions of samples, or even the status information of instruments. When a calibration model is used, it may be unavailable to new samples due to the change of instrument status or samples, which will have a great influence on the adaptability and reliability of the model, or even leads to erroneous results. Since there are great differences in spectra for beef samples of different origin or species and the lack of versatility and universality, the beef quality detection model may have different results when test on samples of different species. So it is of great importance to study on the maintenance methods of beef quality detection model.The thesis takes three different species of beef samples (Huangpi cattle, Enshi buffalo, Simmental cows)as the research object, established the quantitative model for beef freshness (TVB-N content) detection based on near infrared spectroscopy and hyperspectral technology. Compared the prediction results of models for the three species of beef samples, and study on the suitability evaluation of the models. Compared the maintenance effects of different model updating and transfer methods on models for beef freshness detection based on spectra. The main conclusions are as follows:(1) Compared the sample partition results of Random method, Kennard-Stone and SPXY algorithm, and also the impact on model prediction performance for each method, with the result that the most suitable method for sample partition of beef freshness detection was SPXY algorithm. Take the freshness detection of Huangpi cattle based on NIR for example, the RCV and Rp of the detection model, which was established by the same spectra preprocessing and modeling method, were both higher than the others, and reached0.8247and0.8122, respectively.(2) Different spectra preprocessing and modeling methods were compared, and also the impacts on the prediction performances of the freshness detection model as well. The results were as follows:For freshness detection based on NIR, the best preprocessing methods for Huangpi cattle, Enshi buffalo and Simmental cows were Poisson scaling, Autoscale and Poisson scaling, respectively. And for the hyperspectral detection, the three best methods were MSC, Autoscale and Poisson scaling. The most suitable modeling method for beef freshness detection was PLS regression method. (3) The changing trend of freshness and spectra information of beef samples were obviously variable with species of beef samples, which leaded to that the model established by a single specie of beef samples were not suitable for the freshness detection of another one. In order to study on the suitability evaluation of the models, PCA, Mahalonabis Distance and Cross test methods have been adopted. Take the NIR freshness detection model for Huangpi cattle for example, the Rp of the model were0.7310and0.7065for freshness detection of Enshi buffalo and Simmental cows, and the RMSEP reached up to17.3271and19.4359. In one word, a model can only performed well for the samples used to established it. The prediction performance of the model decreased sharply when applied for the detection of another variety of samples directly.(4) Model updating based on global model and adding new typical samples selected by SPXY algorithm can contribute to improving the prediction performance of freshness detection models for different species of beef samples. The prediction results of the global model for three species of beef samples were rather stable. The RCV of the model based on NIR and hyperspectral were up to0.8143and0.9155, with the RMSEP reached3.9075and2.6386, respectively. Since23typical Simmental cows samples selected by SPXY algorithm were added to the NIR model of Enshi buffalo, the model’s Rp for Simmental cows increased from0.6432to0.8111, with the RMSEP from5.244to3.5668. For the Hyperspectral model of Huangpi cattle, its Rp for Enshi buffalo increased from0.021to0.896when18typical Enshi buffalo samples were added to the calibration set, with the RMSEP from8.347to2.848.(5) The transfer effect of the slope/bias correction method for freshness detection model was studied. And a spectra correction method based on absorbance modification was also proposed to improve the prediction performance of the freshness detection models. The results showed that the slope/bias correction method was not satisfied with the model transfer, which could not improve the prediction correlation coefficients and only can make the RMSEP slightly decreased. But when the spectra correction method was adopted, the Rp of the NIR model established by Enshi buffalo increased to0.806for the freshness prediction of Simmental cows. And for the Hyperspectral model of Huangpi cattle, when applied for the freshness prediction of Enshi buffalo, the two parameters were improved to0.851and2.872, respectively.