Near-infrared spectroscopic analysis of diesel properties based on back propagation neural network optimized by a hybrid genetic algorithm-particle swarm optimization algorithm

Rapid determination of the physico-chemical properties of diesel is of significant importance for both refinery production and environmental protection. To address the time-consuming nature of traditional chemical analysis and the tendency of back propagation neural networks (BPNN) to converge into local optima, this study proposes a near infrared (NIR) spectroscopic method for determining diesel properties by integrating chemometrics with a genetic algorithm-particle swarm optimization BPNN (GA-PSO-BP). The diesel spectral data were preprocessed using Savitzky-Golay (S-G) smoothing and differentiation, while outliers were identified through Monte Carlo cross-validation (MCCV). The boosting soft shrinkage (BOSS) algorithm was employed to extract key characteristic wavelengths from the high-dimensional spectra, significantly compressing the feature dimensionality and reducing model complexity. Finally, a hybrid GA-PSO algorithm was utilized to optimize the initial parameters of the BPNN, effectively overcoming its inherent limitations regarding local optima. The results demonstrate that the GA-PSO-BP model outperforms traditional machine learning models in predicting six diesel properties, with the coefficient of determination (R ² ) reaching maximum values across all parameters. Furthermore, SHAP (SHapley Additive exPlanations) analysis was introduced to reveal the nonlinear mapping relationship between spectral absorbance and physicochemical indices, thereby enhancing model interpretability. This research provides an efficient and robust solution for the rapid and non-destructive monitoring of diesel quality.