Recently, overexploitation of oil resources has become a major problem, attracting extensive attention. Thus, alternative new fuels are needed to alleviate the problem. Biodiesel is usually mixed with diesel as an alternative to diesel because of its similar properties and excellent environmental benefits, thereby reducing the actual consumption of diesel. However, the proportion of diesel in mixed oil containing diesel and biodiesel greatly affects the power transformed. A low diesel proportion will cause problems, such as poor oil atomization, insufficient oil-gas mixing, insufficient engine pressure, and increased nitrogen dioxide emission. Although several countries have different specifications for the proportion of diesel in mixed fuel, the proportion of diesel should be more than 50%. Therefore, the rapid and effective identification of the proportion of diesel in the mixed oil of diesel and biodiesel is greatly significant for the qualitative detection of imported diesel.

In this paper, the peak pick-up method was used to detect the proportion of diesel in the mixed oil. Since there were many hydrocarbons in the mixed oil, the fluorescence spectrum of the sample was complex. Thus, selecting the appropriate fluorescence peak was difficult. The sample was projected directly on the two-dimensional plane using two-dimensional principal component analysis (2DPCA), and the first three principal components were selected to reconstruct the fluorescence spectrum. Observe that the reconstructed fluorescence spectrum retained the fluorescence peak with more principal component information. Using this peak intensity to predict the proportion of the diesel oil will greatly improve the prediction accuracy. However, no linear relationship between the retained fluorescence peak intensity and the proportion of diesel oil exists, thus, introducing a neural network for further analysis is necessary. Furthermore, we selected a generalized regression neural network (GRNN) to predict the proportion of diesel oil because of its advantages in dealing with nonlinear relationships. Additionally, three sparrows of the small slope approximation (SSA) method were used to search the parameters of the GRNN-spread, and the GRNN was established using the best spread to accurately detect the proportion of diesel in the mixed oil.

For the oil mixture sample of diesel and biodiesel (Table 1), we divided the sample into test and training sets, respectively. It can be found from Fig. 1 that the fluorescence spectrum of the sample was complex, with more fluorescence peaks and less difference in fluorescence intensity; thus, accurately identifying fluorescence peaks containing more principal components was difficult. In the Fig. 1, the best fluorescence peak of diesel oil was at 450/480 nm and that of biodiesel was at 370/410 nm. With the decrease in diesel proportion, the fluorescence intensity at the best fluorescence peak position of diesel failed to satisfy the linear relationship with the proportion of diesel, as shown in Fig. 3. Furthermore, the fluorescence spectrum of biodiesel greatly affected that of diesel. It was impossible to simply extract the strongest fluorescence intensity and establish a neural network for prediction. The three-dimensional fluorescence spectrum of the sample was reconstructed using 2DPCA, as shown in Fig. 5. Since the remaining fluorescence peaks of the reconstructed fluorescence spectrum contain more principal components, selecting the spectral information with strong fluorescence intensity for prediction reduced the error. According to Fig. 4, the reconstructed emission spectrum of the sample matched the original diesel sample. Therefore, the strong part of the reconstructed fluorescence spectrum (9×6) was selected as the training set. The GRNN optimized using SSA was established to predict the test set and the prediction results are shown in Table 3. The results revealed that the prediction effect of the 2DPCA-SSA-GRNN model is good, with an average recovery of 98.39% and a mean square error of 0.90%, which is about 1/2 the mean square error predicted by SSA-GRNN alone.

The intelligent algorithm was used to optimize the neural network, which realized the quantitative analysis of substances. However, its accuracy was low considering the complex situation of the fluorescence spectrum. In this paper, we proposed 2DPCA to reconstruct the fluorescence spectrum of the sample. Our method retains the original fluorescence characteristics of the spectrum, simplifies the redundant information in the spectrum, improves the quantitative analysis results of diesel in mixed oil, and provides a new idea for the quantitative analysis of substances. With the progress in science and technology, more additives can be used to optimize the performance of substances, even with more complex spectra. 2DPCA can screen out the main fluorescence information in the spectrum and improve the peak picking method to analyze substances. In future work, 2DPCA can be combined with other spectral analysis methods to further improve the accuracy of fluorescence information extraction for serious spectral overlap and realize the quantitative and qualitative analysis of complex substances.