Research On Multi-Model Fusion Modeling Method Based On Near Infrared Spectral Data

When performing multivariate calibration analysis on near-infrared spectral data,the quantitative regression model is usually established using a partial least squares algorithm(PLS).Usually only a single calibration model was established in actual calculation.However,the single-model modeling method is easily leaded to model over-fitting when the number of samples is small but the number of variables is large.Therefore,it is a challenge to establish a stable and accurate model for the complex spectral data.To solve the problem of instability of single model,this paper proposed to combine the multi-model modeling method with the sparse algorithm Lasso,and implemented it to modeling the regression model for near infrared spectral data.Additionally,this paper discussed the multivariable data processing methods.The main research contents are as follows:In the first chapter of the paper,the background information and research purpose of this paper were introduced.And the principle and characteristics of near infrared spectroscopy technology were described.In addition,the regression modeling methods for near infrared spectroscopy in this paper were briefly explained.Finally,the development history and advantages of fusion model were summarized.This paper combines multi-member models modeling method with sparse algorithm based on the idea of multi-model fusion modeling then the fusion process of the model was studied from three aspects of partial least squares regression,ridge regression and error weight.At last,three fusion algorithms of partial least squares regression coefficient fusion(PLSRCF),ridge regression coefficient fusion(RRCF)and error weight fusion(EWF)were established.Using above three modeling algorithms,the near infrared spectral data of corn,apple and marzipan were employed to build the quantitative models to verify the effectiveness of the three fusion models.Finally,the results of the fusion models were compared with the single PLS model.It was found that the root mean square error values of PLSRCF,RRCF and EWF were lower than that of single PLS model.And the prediction accuracy of the fusion model in corn data is even improved by 80%.The results show that these three fusion algorithms can make full use of the information of spectral data and achieve accurate analysis on components of spectral data.