Comparison of Least Square Support Vector Machine-Based Calibration Methods in Diesel Property Analysis by Near Infrared Spectroscopy

Abstract

Partial least square (PLS) and least square support vector machine (LSSVM) are widely applied in NIR calibration modeling. The aim of this paper is to compare three kinds of LSSVM based calibration methods in diesel property by NIR analysis, which are regular LSSVM, LSSVM with feature extraction by principal component analysis (PCA) and PLS. Thirty nine diesel samples with known properties which include calculated cetane index, density, total sulfur and T50 (boiling point at 50 % recovery), are collected from a refinery in China. Their near infrared spectra are measured by a spectrometer with the wavelength range of 900-1700 nm. They are divided into a calibration subset with 29 samples and a validation subset with 10 samples. The above LSSVM based calibration methods as well as PLS are employed to build models with the calibration samples and tested with the validation samples. Experimental results show that the LSSVM with PLS feature extraction presents the best performances in all of the calibration methods. Its root mean squared error of prediction (RMSEP) of diesel calculated cetane index, density, total sulfur and T50 are 0.45, 3.19, 0.0482 and 3.90, respectively and the corresponding multiple correlation coefficients of prediction (R²) are 0.953, 0.924, 0.974 and 0.954, respectively.