An integrated framework combining CenFormer and PLS regression for rapid distillate oil classification and property prediction

Abstract

Rapid and accurate classification and property prediction of distillate oil are essential for intelligent quality control and process optimization in modern refineries. Traditional methods, such as spectral analysis with chemometrics, are widely applied, but heavily depend on manual feature engineering and offer limited representation capacities. Recent advances in deep learning have shown promise for oil analysis, yet existing models often struggle to jointly capture fine-grained local patterns and long-range spectral dependencies, and rarely optimize feature space geometry. To address these challenges, an integrated framework is proposed, integrating spectral preprocessing, a dual-branch CenFormer model, a joint loss function, and dynamic property prediction. Spectral preprocessing is employed to sharpen spectral features by applying baseline correction, spectral truncation, and vector normalization. The CenFormer model leverages a CNN-Transformer dual-branch architecture, enabling the simultaneous capture of fine-grained local patterns and long-range spectral dependencies. A joint loss function, combining softmax and center loss, enforces intra-class compactness and inter-class separability, thereby improving feature discriminability. For property prediction, a similarity-based sample selection strategy is performed, followed by PLS regression, to enable adaptive modeling of physicochemical attributes. Experimental results demonstrate the effectiveness of the framework, achieving a classification accuracy of 99.51 %, low RMSEs and rRMSEs, and high $R^2$ in property prediction, highlighting its potential for rapid and reliable spectral analysis in industrial applications.