Detection And Identification Of Pork Freshness Based On Near Infrared Spectroscopy

Traditional detection methods of pork freshness are complex, time-consuming, difficult to detect a large scale of samples quickly and in a real time. Inaddition, the problem that frozen-then-thawed pork replaced the fresh pork happened occasionally. Therefore, this research established quantitative prediction models of total volatile basic nitrogen (TVB-N) and total viable count (TVC) of bacteria which present the pork freshness quality, and qualitative indetification model for frozen-then-thawed pork. The specific conclusions are as follows:1. Three partial least square (PLS) models were established by three kinds of sample's pretreatment methods as being intact, grounded and exuded juice. They were optimized by many spectral data preprocessing methods and got their best model respectively. By comparing the three best models, the conclusion was drawed:the best sample's pretreatment was grounded and the best model was based on the ground pork's original spectrum by standard normal variable transformation (SNV) preprocessing. Its correlation coefficient and root mean square error in calibration set was0.85412,2.65, and in prediction set were0.82626,2.79.2. Multisensor's information of Near infrared (NIR) spectral data, colour data and pH data were fused by radial basis function networks (RBFN) in original data levels for quantitative detection TVB content in pork. By comparing the detection results by any single sensor' information or any combination of their information, we found the best model was achieved by three information fused. Its correlation coefficient and root mean square error in training set was0.98,1.565, and in prediction set were0.945and2.722.3. The quantitative detection model of total viable count of bacteria was established and optimized by multiple data processing methods. Among these, the best spectral data preprocessing mothed was multeiplicative scatter correction (MSC); the correlation coefficient method was used to extract the characteristic spectral regions as9759～8878cm-1,8756-8230cm-1,7610-7212cm-1,6766-6606cm-1,4860～4582cm-1, the first, fourth and fifth regions are similar as absorption regions of the known protein'N-H group, so we can infer that the TVC content in pork correlated to the variation of the pork protein; at last, the characteristic waveband were input to the back propagation artificial neural networks (BP-ANN) and got the best detection model. Its correlation coefficient and root mean square error in training set and prediction set were both about0.98and0.23.4. The principal component analysis composed with mahalanobis distance discriminant model and probabilistic neural network (PNN) model were established to discriminate the frozen-then-thawed pork samples storaged in-18℃at0,10,20,30days. The latter model was better and its recognition rate and reject rate were higher than93%. With the preparation of the different period prok tissue section, their tissue structure was compared, and the frozen-then-thawed samples storaged in-80℃at40days were input to the two model to discriminate, we primarily got the mechanism that NIR diffuse reflection spcetroscopy discriminate the frozen-then-thawed pork at different storaged period was the spectroscopy got the pork tissue structure information to discriminate them.