Maintenance Methods Of Quality Detection Model For Different Varieties Of Pork Based On Hyperspectral Imaging Technology

China is a big consumer of pork, and pork quality is related to consumers’health and economic benefit. Whether the multivariate calibration model of pork quality by spectroscopy technology has a stable performance directly affects accuracy of analytical results. Since the changes of sample information (such as varieties, origin, particle size, etc.) or the changes of measure environment (such as temperature, humidity, etc.) may lead the calibration model to be unsuitable for new samples, the prediction performance of the spectral model of pork quality for other varieties of samples may deteriorate due to varieties’difference, and the model lack robustness and applicability. However, re-establish a stable and accurate model is a complicated procedure, which consumes a lot of manpower, material resources and time. Therefore, research on maintenance methods to improve applicability and reliability of spectral model of pork quality has important scientific and pratical value.In this paper, taking Shanhei pig and Linghao pig samples as object investigated, the research focused on pH value quatitatve detection method and tenderness qualitative detection of pork based on hyperspectral imaging technology. Effect of differrent spectra preprocessing and modeling motheds to the performance of the detection models were compared to determine the optimal ones. The applicability of detection models was studied and the maintenance effects of different model correction and transfer methods for pork quality detection models were compared. The main conclusions were as followed:1) The partition result and impact on the performance of model of different sample set partitioning methods for pork pH value quantitative analysis were studied. Sample set partitioning based on X-Y distances algorithm (SPXY) was the best method for pork pH value detection. Taking Shanhei pig samples for example, the data were respectively patitioned by random sampling algorithm (RS), Kennard-stone method(KS) and SPXY algorithm, then Partial Least Squares models were established. The results showed that RS method and KS method were not satisfied, the Rc and Rp of the model based on SPXY were both higher than the others. 2) The influence of different spetra preprocessing and modeling methods for pH value quantitative analysis of Shanhei pig samples were compared. The optimal preprocessing method was Normalize and the optimal modeling method was Partial Least Squares(PLS). Dividing Shanhei pig samples into calibration set and test set by SPXY, models based on Principal Componet Regression (PCR), Partial Least Squares (PLS) and Support Vector Machine (SVM) were established and analyzed. The results indicated that the performance of PLS model established with Normalize preprocessing spectra was the best, Rc, Rp, Root Mean Square Error of Cross Validation (RMSECV) and Root Mean Square Error of Predition (RMSEP) were0.885,0.864,0.1129and0.1059, respectively.3) The applicability of Shanhei pig pH value quantitative model was studied. PCA and Mahalonabis Distance were used to qualitatively test and Fisher value was firstly adopted to quantitatively test the applicability of Shanhei pig pH value detection model, then cross-test methods was adopted to verify. The results indicated that when Shanhei pig pH value model was directly used to predict Linghao pig samples, Prediction performance was very poor with only Rp of0.415, and RMSEP of0.1804. It could be concluded that Shanhei pig detection model could not performed well for Linghao pig samples directly, so Shanhei pig model needed to be corrected and transferred.4) Model updating method adding new samples to Shanhei pig detection model based on SPXY algorithm could improve the prediction performance of Shanhei pig model for Linghao pig samples. The inpact on model’s applicability by adding new Linghao pig samples to Shanhei pig detection model based on RS method, KS method and SPXY method respectively were compared. The results showed that the distribution of new Linghao pig samples based on different selection methods affected the correction result of model and SPXY algorithm was the best method. When14Linahao pig samples selected by SPXY algorithm was added to the calibration dataset, the prediction performance of correctional model for Linghao pig samples achieved optimal. Rp increased from0.415to0.797, improved92.05%, and RMSEP reduced from0.1804to0.1121, dropped37.86%.5) The maintenance effect of the Slope/Bias method for Shanhei pig pH value detection model was studied. The result indicated that the correction effect of the Slope/Bias method for pH value detection model was limited. The Slope/Bias method only could decrease RMSEP of pH value model from Linghao pig samples, and RMSEP reduced from0.1804to0.1343, only dropped25.54%.6) A calibration transfer algorithm based on spectral value correction was proposed to improve model’s applicability between different varieties. The algorithm did not change Shanhei pig model by eliminating the differences of the spectral values between varieties to improve the applicability of model. When the algorithm was adopted, Rp of the model for Linghao pig samples increased from0.415to0.837, rose101.69%, and RMSEP reduced from0.1804to0.0856, fell52.55%.7) The tenderness qualitative recognition model of Shanhei pig based on Hyperspetral imaging technology was established. Models based on K-Nearest Neighbor algorithm (KNN), Partial Least Square-Discriminant Analysis (PLS-DA) and Support Vector Machine-Discriminant Analysis (SVM-DA) with different spectra preprocessing methods were established and analyzed. The results showed that the recognition accuracy of SVM-DA model established with Normalize preprocessing spectra was the best, and the model’s calibration ser and test set both got the100%recognition rate.8) Establishing global model, model updating and spectra signal correction method were benefit to improve the recognition rate of Shanhei pig tenderness recognition model for Linghao pig samples. The recognition rate of the global model for different varieties of samples reached80%, and its applicability was better than models which only was established by one variety of samples. For model updating, when20Linahao pig samples was added to the Shanhei pig recognition model, the recognition rate for Linghao pig samples achieved optimal and increased from50%to76.7%. For spectra signal correction method, Recognition rate increased from50%to73.3%after correcting the difference of spectral signal between different varieties of pork samples.9) Spectral value correction algorithm could effectively improve the applicability of tenderness recognition model of Shanhei pig. The pectral transformation matrices of different tenderness grade of pork samples firstly were calculated, then spectral data of test samples of Linghao pig was reconstructed with the matrices. When the algorithm was adopted, recognition rate of original tenderness recognition model of Shanhei pig for Linghao pig samples increased from50%to100%.