Abstract
High-quality gasoline combustion reduces atmospheric emission. The evaluation of gasoline quality is needed reduce air pollution. The purity of gasoline is commonly quantitatively evaluated by octane number. Near-infrared (NIR) spectral determination of the octane number is practical for the rapid determination of gasoline quality. In NIR, the calibration model must be optimized for the improvement of the accuracy using deep learning methodologies. In this work, an algorithmic architecture was constructed for the deep learning of hybrid optimization by the combination fractional derivative (FD) and partial least squares (PLS). The design of FD-PLS methodology refers to the joint optimization for the data pretreatment and the calibration modeling. The FD order was tunable from 0.1 to 2.0 to search for the optimal pretreatment. Calibration models were established and trained by screening the optimal parameter combination for the model. The experimental results show that the calibration model was most improved by the FD-PLS deep learning with a 1.3 derivative order and 7 informative latent variables. The optimal model performed better than using integer derivative orders. In conclusion, the hybrid parameter optimization for the FD-PLS deep learning architecture improved the NIR assessment to characterize gasoline quality with quantitative prediction of the octane number. These results allow the production of high-quality gasoline to minimize air pollution.