A total of 87 fuel petrol samples were collected. An oil filter (TY-II/1000) was dedicated to evaporate samples until fully dewatering. The dewatering samples were considered as pure petrol. The octane number of the pure petrol samples can be detected by using SKY2102-VII Gasoline Octane Number Tester (Shanghai Shenkai Petroleum Instrument Co. Ltd.). The 87 values range from 83.4 to 89.6. The arranged pure petrol samples were placed in an ultrasonic cleaning machine for oscillation beforehand.

NIR spectra of the 87 pure petrol samples were recorded on the NIR-Quest 256 spectrometer (Ocean Optics, USA) fitted with In GaAs detector and quartz-halogen light source. The scanning spectral range was set 800 - 2498 nm with 4 nm resolution, so that we have 850 wavelengths per spectrum. The experiment temperature was controlled at 25˚C ± 1 ˚C and the relative humidity was at 47% ± 1% RH throughout the scanning process. Each sample was measured thrice and the mean value was calculated for model establishment.

NIR analytical process requires a modeling-testing division for samples. And modeling optimization process should have the samples partitioned for calibration and validation. Firstly, 22 samples were randomly selected for testing, which were not subjected to the modeling process. The remaining 65 samples were used for modeling. We are planning to have 45 samples for calibration and 20 for validation. To obtain representative sample sets, the calibration-validation partition should be carried out based on both the absorbance values and the measured values from SKY2102-VII, so the SPXY method [16] was utilized.

The NIR spectra of 87 pure petrol samples were showed in Figure 1. As can be seen in Figure 1, a sharp absorption peak appeared around 1400 nm and another around 1900 nm, which respectively corresponds to the first overtone of the chemical bonds of H-O and C=O. Because of these two peaks and the reasons of high frequency random noise, baseline drift, light scattering in raw spectra, the multivariate calibration models cannot smoothly give out prospective prediction results, thus pretreatment methods should be employed to reduce the interferences in raw data.

According to the physical properties of the NIR absorption band of H-O and C=O, we utilized the 1st-derivative Savitzky-Golay smoother (1st-d’SG) for data pretreatment, and the pretreated spectra were showed in Figure 2. We can see from Figure 2 that the 1st-derivative + SG pretreatment well reserve the spectral features of the peak signals at 1400 nm and 1900 nm, and all spectral curves obviously became smoother.

The NIR spectral data after 1st-d’SG pretreatment were used for quantitative determination for the octane number in petrol samples. Based on the spectral in full scanning range, wavelength selection models were established for the calibration and validation samples by SPA method, so that the informative wavelengths can be selected. The selected wavelengths will be taken as the input variables of LSSVM regression and the NIR analytical models can be further established for optimization.

SPA was used to select wavelengths for NIR quantitative analysis of fuel petrol. The NIR data were beforehand pretreated by first derivative SG smoother (1st-d’SG). According to Balabin’s reports [18] [19] , a good wavelength selection model provided by SPA always includes 2 - 20 variables (i.e. wavelengths). We tried to expand the number to 25, and respectively established models for the number changed from 1 to 25. The RMSE corresponding to each number of variables by SPA was showed in Figure 3. As Figure 3 shows, the RMSE curve sharply declines when the number changes from 1 to 5, which indicated that the proper number should be larger than 5 in order to avoid over-fitting. And also, the curve seems to be stable when the number goes from 4 to 9, and rises when larger than 9. Thus we concluded that the number of informative wavelengths for the best NIR calibration model of fuel petrol should be no less than 4 and no more than 9.

Next, we reset the number of variables changed from 4 to 9 and repeat wavelength selection by SPA. Then we have the best 8 variables (the most informative wavelengths) for the NIR analysis of the octane in fuel petrol. The specific 8 wavelengths were 1040, 1316, 1392, 1476, 1800, 1856, 1904, 2298 (the 8 circles in Figure 4). These wavelengths distributed at the peak or trough locations of the 1st-d’SG spectrum. This result reflected the merit of SPA that the selected informative variables do not appear at the flat region but gather at the special locations of the 1st-derivative curve, which point to some dedicated functional groups. The 8 wavelengths selected by SPA based on the 1st-d’SG pretreated data reserved the main features of spectra. Thus we use these 8 SPA-selected wavelengths to establish calibration models for NIR determination of the octane in fuel petrol.

NIR calibration models were established by using LSSVM regression with kernel function. Theradial basis function (RBF) kernel is common used in chemometrics [14] . RBF kernel includes two tunable parameters γ and σ. We utilized genetic algorithm (GA) to optimize the parameters. We have the GA iteration error of 0.02 at the 30th iteration, and the optimal value of γ and σ can be found. According to the abovementioned SPA wavelength-selected results, we established NIR calibration models by LSSVM regression, and further to predict the octane number for the prediction samples. The prediction results were listed in Table 1. Here we used correlation coefficient (R) and root mean square error (RMSE) as the model indicators.

The results in Table 1 indicated that, on one hand, the raw spectra and the 1st-d’SG pretreated data let out the calibration models with relatively low prediction bias, while the 1st-d’SG pretreatment provided a noise-reduced data to give out an optimal model with higher R and minor RMSE. On the other hand, SPA wavelength selection method performs well for LSSVM models. The SPA-selected 8 informative wavelengths did improve the LSSVM modeling results, with the GA-optimized LSSVM modeling parameters (γ, σ) were (89.34, 32.45), and calibrating R and RMSE were 0.951 and 3.282, respectively, which was quite satisfactory for the NIR quantitative analysis of the octane number in fuel petrol.

For model evaluation, the 22 independent testing samples, not subjected to the modeling process, were used to examine the optimal SPA-LSSVM models with or without 1st-d’SG pretreatment. The predictive values of the 22 testing samples can be calculated respectively for raw data and the pretreated data, and the correlation charts were showed in Figure 5. The correlation coefficient for the raw data modeling was 0.876 and RMSE was 4.416 (Figure 5(a)). The correlation coefficient for the 1st-d’SG data modeling was 0.903 and RMSE was 4.128 (Figure 5(b)), which was a little better than the results on raw data. We can see from Figure 5(b) that the testing samples evenly distributed on both sides of the regression line and the distance to the regression line is relatively closer. It showed that the SPA-LSSVM model based on 1st-d’SG pretreated data gave out the best model testing results.

Combined considering the modeling results showed in Figure 5 and Table 1, we conclude that SPA method can effectively extract informative wavelengths from the full NIR scanning range. The 1st-d’SG pretreatment has the advantage of selecting wavelengths containing the implicit information. LSSVM regression can establish calibration models by linear and non-linear training mechanism. The combination of 1st-d’SG pretreatment and SPA-LSSVM show the potential to predict the octane number for fuel petrol samples.