Spectral Analysis Of The Mixed Solution And The Fusion Modeling Research

Due to the characteristics of non-destructive,on-line detection and high efficiency,infrared spectrometry has been applied to the quantitative and qualitative analysis of agricultural products,food,petrochemical,pharmaceutical and other fields.However,the measured spectra often contains noise,baseline drift and interference information,these factors seriously affect the final results of spectral analysis.The suitable preprocessing,band selection and modeling method are the key to overcome the interference and improve the performance of the model.Derivative spectra has been widely used as common pretreatment method,can not only eliminate the effect of baseline drift,but also distinguish aliasing peaks and spectral information of the depth of excavation to improve the spectral resolution.In this sense,this thesis presents a derivative spectra modeling method based on singular perturbation technique,termed ensemble derivative spectra Partial Least Squares(EDSPLS).The core idea of this method is that derivative spectrum information mining ability,the optimization ability of band selection are used to achieve the purpose of improving the performance of the model.In order to obtain the high quality derivative spectral information,the weighted fusion modeling of the same order spectral information space is realized by using the interval partial least squares.On this basis,the integration of different order spectral information is further integrated to form the final model.In order to verify the validity of the algorithm,the beer and wine samples are tested for the experiment.In the experiment,root mean square error of cross validation(RMSECV)and root mean square error of prediction(RMSEP)of beer original spectral partial least squares(PLS)model are 1.3775,1.2475,RMSECV and RMSEP of stacked PLS(SPLS)model are 0.0945,0.1161,RMSECV and RMSEP of EDSPLS model are 0.0664 and 0.0854,respectively.Compared with PLS and SPLS,RMSECV and RMSEP are increased by 95.18%,29.76%and 90.75%,26.44%.In the experiment of wine,compared with the EDSPLS model,the RMSECV and RMSEP of the PLS model and SPLS model are improved by 28.33%,36.09%and 28.91%,16.95%,respectively.Experimental results show that the proposed method of ensemble derivative spectra partial least squares method is superior to the existing modeling method.