The Research Of Svm Model Based On C-H Group Of Near Infrared Spectroscopy

Near infrared spectroscopy(NIRS)analysis is a fast,reliable,non-destructive and environment-friendly method.NIRS analysis has been widely applied in many areas,such as agricultural products test,petroleum products component analysis,etc.Building chemometric model with strong robustness and accuracy is a main part of NIRS analysis technology.Support vector machine(SVM)is a learning machine based on VC dimension theory and structural risk minimization principle.SVM has the advantage in solving problems with small quantity of samples,and has superiority in solving nonlinear modeling problems.Least squares support vector machine(LSSVM)is the improvement of traditional SVM.Quadratic programming problem in SVM has been converted into solving linear equations in LSSVM.Computational complexity in SVM can be reduced in this way.Aiming at parameters optimization in LSSVM,an improved fruit fly optimization algorithm(DAFOA)is proposed to determine the penalty factor and the parameter in kernel function.In order to improve the ability of traditional fruit fly optimization algorithm(FOA)to get rid of local minimum,DAFOA can tolerate suboptimal flies and adopt dynamic step.The experiments of test function and NIRS data of diesel show that,DAFOA can improve the ability of parameter optimization,and then improve the forecast precision of LSSVM.Aiming at wavelength selection in NIRS,an improved recursive feature elimination(IRFE)method is proposed.Generalization error bound in LSSVM is adopted as criterion function to determine the significance of each wavelength.Improved sequential floating backward selection(ISFS)is adopted to select key wavelengths.It demonstrates experimentally by NIRS data of diesel and corn,that IRFE method can improve the efficiency of wavelength selection,and obtain the simpler and better wavelength set.In order to simplify LSSVM model of the substance made up of C-H groups,a method that combining IRFE with C-H groups absorption bands is proposed.Frist,wavelength bands are defined according to the center absorption wavelength of each C-H group.And then,wavelengths in these bands can be selected by IRFE.This method can avoid selecting invalid wavelengths,and reduce the complexity of wavelength selection.Compare with IRFE based on full spectrum,this method can reduce operation time and obtain a simpler wavelength set.The generalization ability of LSSVM model can be improved.