Abstract: In order to achieve rapid on-site detection of common light fuel classification, five types of Raman spectra of 92# gasoline, 95# gasoline, 98# gasoline, 0# diesel and 3# jet fuel from different origins were collected. Normalization operator, Savitzky-Golay smooth convolution operator (SG) and adaptive iterative penalized least squares operator (airPLS) were successively used to preprocess the original spectrum. Then the sparse principal component analysis (SPCA) is applied to reduce the dimension of the preprocessed spectrum. Finally, the classification model is constructed by different methods to classify light fuel samples. The analysis results indicate that: the preprocessing of Raman spectrum can effectively eliminate the background noise interference and improve the accuracy of the classification model; the classification model constructed by sparse principal component analysis-support vector machine (SPCA-SVM) method has the best classification and identification effect on light fuel; the SPCA-SVM method is used to construct the classification model for distinguishing 95# gasoline from 92#gasoline + 95# gasoline blended gasoline, when the volume fraction of 92# gasoline in blended gasoline is more than 15 %, good recognition and classification results are achieved.

Key words: Raman spectra, spectrum preprocessing, qualitative analysis, sparse principal component analysis, support vector machine