LAUSR

Abstract As a potential convenient and lower cost analytical technique, ATR-FTIR spectra can be used to identify graded asphalt for overcoming the drawback of traditional physical and chemical methods. However, due to the existence of many uninformative and interferential wavenumbers, it is necessary to select discriminating wavenumber to represent significant features of graded asphalt spectra before developing classification model. To get the discriminating wavenumbers, a wavenumber selection strategy is put forward in this paper. In the strategy, full-spectrum, four spectral wavenumber selection methods (competitive adaptive reweighted sampling, CARS; successive projections algorithm, SPA; Monte Carlo uninformative variable elimination, MC-UVE and moving window Partial Least Squares, mwPLS) coupling with Support Vector Machines (SVM) classifier (Full-SVM, CARS-SVM, SPA-SVM, MC-UVE-SVM, mwPLS-SVM) are applied respectively to investigate the wavenumber selection method, and then the wavenumbers selected by the best wavenumber selection method are regarded as the discriminating wavenumbers correspondingly. Experiment results show that the CARS-SVM is the best wavenumber selection method among CARS-SVM, SPA-SVM, MC-UVE-SVM and mwPLS-SVM methods, and 42 discriminating wavenumbers are selected by CARS. Furthermore, the stability of these 42 discriminating wavenumbers has been verified by new graded asphalt from different source. Hence these discriminating wavenumbers of ATR-FTIR spectra coupling with SVM classifier can be used as a fast graded asphalt identification method in road construction sites.