Hyperspectral Feature Wavelength Selection And Quantitative Analysis Of Fat Content In Milk

The fat in milk has a high nutritional value and is one of the most important nutrients for human health.In today’s market there are strict requirements for milk quality testing,and the rapid and accurate determination of milk fat content is particularly important.The development and application of spectroscopy provides a new way to detect the fat content of milk.Hyperspectral data has a very high spectral resolution and can reflect the spectral information of different components in milk.However,hyperspectral data have many dimensions and a large amount of data,especially between adjacent or close wavelengths,and there is serious information redundancy,which not only reduces the accuracy of the prediction model,but also increases the computing time of the model.To address this problem,this thesis takes different brands of pure milk as the research object,uses hyperspectral imager to obtain spectral data of milk samples,studies the feature wavelength selection method suitable for milk fat content detection,and constructs the best prediction model for milk fat content.The main work carried out in this thesis is as follows:(1)Hyperspectral imaging data of different brands of pure milk samples were collected by hyperspectral imaging technique and the spectral reflectance of milk samples were extracted by ENVI software.Eight single algorithms as well as eight combined algorithms were selected from four directions of scale scaling,scattering correction,baseline correction and smoothing processing to pre-process the raw spectral information,build PLSR and SVR correction models respectively,and evaluate them using cross-validation.When PLSR is the correction model,VN preprocessing is the best;when SVR is the correction model,SNV preprocessing is the best.(2)The linear and non-linear relationships between milk spectral reflectance and milk fat composition were investigated using the correlation coefficient method and the swarm intelligence optimization algorithm combined with linear and non-linear models to compare the differences in the selection and modelling of feature wavelengths using different relationships.The experimental results show that nearly half of the wavelength points are completely different using linear and non-linear relationships to filter the characteristic wavelengths respectively.The non-linear model has a higher prediction accuracy,R_P~2 and RMSEP of 0.9869 and 0.1458,respectively,but the linear model runs faster and has guaranteed prediction accuracy.(3)A new method of feature wavelength selection(PLSRC-ACO)based on an ant colony algorithm combined with PLSR model regression coefficients is proposed.To address the shortcomings of the initial pheromone of the ACO,the PLSR model regression coefficients were used as the main basis for evaluating the importance of wavelengths,which were used as the heuristic information of the ACO to conduct an intelligent search to screen the best wavelength combinations suitable for milk spectra to predict fat content.Finally,PLSRC-ACO is compared with classical wavelength selection methods such as SPA,UVE and CARS.Experimental results show that the PLSRC-ACO algorithm not only improves the accuracy of the model,but also reduces the number of wavelengths required for modelling.(4)A new method of feature wavelength selection by PLSRC-ACO combined with crossover operator(PLSRC-ACO-CO)is proposed.A global threshold is set and the outstanding individuals generated in the PLSRC-ACO iteration are screened by population dissimilarity as the initial population of the PLSRC-ACO-CO algorithm.Observe the frequency of selected outstanding individual wavelengths in PLSRC-ACO to generate a number of coarse interval points,perform coarse crossover operation at coarse interval points,generate a number of fine crossover points randomly within coarse interval points,and perform fine crossover operation within coarse interval to exploit the potential of PLSRC-ACO outstanding individuals combining with each other as much as possible.Compared with PLSRC-ACO,the number of wavelengths is selected more flexibly,the model is more stable and the prediction accuracy is higher.(5)Three linear models,MLR,RR and PLSR,and three non-linear models,SVR,RFR and XGBoost,were used to predict milk fat content.Through optimization analysis and comparison,the VN-(PLSRC-ACO-CO)-MLR milk fat content prediction model was constructed based on the linear relationship between spectral reflectance and fat composition,requiring a wavelength number of 44 and a running time of only 0.02(s),R_P~2and RMSEP of 0.9842 and 0.1606;the SNV-(ACO-SVR)-SVR milk fat content prediction model was constructed based on the non-linear relationship between spectral reflectance and fat composition,with the required number of wavelengths of 55,a run time of 11.27(s),R_P~2 and RMSEP of 0.9869 and 0.1458.The combined SNV-(ACO-SVR)-SVR model has the highest prediction accuracy,but the VN-(PLSRC-ACO-CO)-MLR runs most efficiently,with a running time of only 2‰of the SNV-(ACO-SVR)-SVR,and has a higher prediction accuracy,which can meet the demand for accurate detection.Under comprehensive consideration,VN-(PLSRC-ACO-CO)-MLR was selected as the milk fat content prediction model.