Research On Model Transfer Of Near-infrared Spectroscopy

Near-infrared spectroscopy(NIRs)can be used for nondestructive,rapid and real-time detection of samples,which has been widely used in many fields,such as petrochemical industry,food testing,agricultural product identification and drug supervision and so on.However,in the actual production,the poor universality of the model caused by the difference of measuring environment and instrument seriously restricts the development of NIR spectroscopy.In order to solve this problem,this paper studies the model transfer in spectrum preprocessing,best wavelength point matching,wavelength selection and transfer learning.(1)In order to solve the problem of poor universality of the model caused by noise and baseline drift in near-infrared spectrum,a model transfer method(WDTW)based on dynamic time warping algorithm of wavelet transform is proposed to realize the sharing of models among different instruments.Firstly,the spectrum is preprocessed by wavelet transform,then uses Dynamic time warping(DTW)algorithm to find the optimal correspondence between the wavelength points of the near-infrared spectrum and establishes the regression equation.In the experiment,the near-infrared drug spectrum data set and the gasoline data set are used to establish the transfer model,which verifies the effectiveness of the transfer method based on the WDTW.(2)In order to solve the problem of over fitting and model instability in the process of spectral processing,a model transfer method which combines random forest with direct orthogonal signal correction algorithm(RF-DOSC)is proposed.In this method,the random forest algorithm is used to select the wavelength points of near-infrared spectrum,and then the direct orthogonal signal correction method is used to process the spectrum and establish the regression equation.The PLS calculates the regression coefficient to obtain the model transfer matrix.In the experiment,the data set of maize near infrared spectrum measured by three spectrometers(S,S1,S2).Through comparative analysis with other algorithms,the model transfer method RF-DOSC can effectively eliminate the spectral noise,reduce the difference between the master and slave instruments,improve the stability of the model,and realize the sharing of models among different instruments.(3)In order to solve the problem of small sample size and poor transfer effect in the process of spectral model transfer,an improved Tr Adaboost transfer learning algorithm was applied to the study of model transfer.Firstly,the algorithm uses random forest to screen the wavelength,reduces the dimension of the sample,removes the noise in the spectrum,then initializes the weight according to the correlation between the auxiliary sample and the target sample,and reduces the problem that the number of iterations is difficult to determine due to the unreasonable initial weight setting of the algorithm.In the experiment,corn NIRs data set is used to establish the transfer model,and the experimental analysis is carried out in different proportion of training set and test set data.It is verified that this method can also achieve high accuracy and achieve the effect of model transfer under a small number of target sample data.