| In this paper,the near-infrared spectrometer was used to measure the adulterated animal hydrolysate protein of camel milk with different contents to obtain the original spectral matrix,and multiple linear regression and machine learning were used to conduct modeling analysis of the near-infrared original spectral matrix.Interval partial least squares method(IPLS)was used for multiple linear regression.In this method,different spectral pretreatment methods were used for the pretreatment of near infrared original spectral matrix of camel milk adulterated hydrolyzed animal protein.The effects of different spectral pretreatment methods on model parameters were compared and the spectral pretreatment methods were compared and determined.After selecting the pretreatment method,the spectral interval selection was further optimized,and the optimal IPLS model was finally determined.In machine learning,Iterative Variable Subset Optimization(IVSO)+ Extreme Learning Machine(ELM)was used to extract and model the features of the original spectral matrix.The two modeling ideas and models were compared and analyzed,and the conclusions were drawn as follows:(1)Compared with the original spectrum,the first derivative,SNV,SG convolution smoothing method,the first derivative +SG smoothing method,the first derivative +SNV,SG+SNV concentrated spectral pretreatment methods,the spectrum was preprocessed,and the 1 principal component regression models of the global spectrum were taken as the evaluation criteria.It was found that the original spectral correlation coefficient and RMSECV value were 0.8455 and 0.3220,respectively,which were better than other and treatment methods.Therefore,the original spectrum was used to model the adulteration of camel milk.(2)RMSECV values of 10,20,30 and 40 principal components were calculated to observe the change of RMSECV values.In the case of calculating 10,20,30 and 40 global principal components,the minimum value of RMSECV is all 1.With the expansion of the calculation scale,its RMSECV value becomes significantly higher when it reaches 29,and then when it increases 10 scales,its value is significantly higher than 29.Further expansion of the principal component calculation scale will only lead to the increase of the model redundancy calculation.It can be seen from the calculation of the global maximum principal component number that the RMSECV value between 2 and 20 basically does not change.In order to reduce the redundancy of the model calculation,the global maximum principal component calculation scale is set as 10.(3)The number of intervals was set as 20 in advance.IPLS modeling was carried out on the original spectral matrix to obtain the regression model with the principal component number of 7,the wave number of the eighth interval ranging from 7887.8 to 77590.87cm-1,the correlation coefficient of 0.9413,and the RMSECV value of 0.2081.(4)The interval setting of 15,25,30,35 and 40 is adopted.The number of available intervals is all 1,and the correlation coefficients are: The RMSECV values of 0.9172,0.9413,0.9290,0.9451,0.9393 and 0.9285 were 0.2511,0.2081,0.2271,0.2001,0.2117 and 0.2288,respectively.By investigating the influence of interval variation on the model,the optimal band bisecting interval 30 can be obtained to obtain the optimal subinterval model.(5)By optimizing the parameters of IPLS,the principal component number is 6,the bisection interval is 30,the band selection is 7787.56?7590.87,the correlation coefficient is 0.9451,RMSECV value is 0.2001,and the model is the best model of the system.(6)Iterative variable subset optimization(IVSO)method was used to extract the features of the original spectral matrix.The dimension of the matrix was reduced from the original 43×1555 to 43×53,and the data scale was greatly reduced.(7)Based on the matrix reduced by subset optimization of substitution variables(IVSO),the correlation coefficient of regression model of camel milk adulterated hydrolyzed animal protein of ELM was 0.91448,and the prediction accuracy was93.3333%. |
PDF Full Text Request