Quantitative analysis of 3-monochloropropane-1,2-diol in fried oil using convolutional neural networks optimizing with a hybrid feature selection based on Fourier transform infrared spectroscopy.

To address challenges in quantifying 3-monochloropropane-1,2-diol (3-MCPD), a toxic compound formed during oil refining with nephrotoxic, reproductive toxic, and carcinogenic properties, a convolutional neural network (CNN) integrated with a hybrid feature selection strategy based on Fourier transform infrared spectroscopy was constructed. Before the construction of CNN, relevant spectral variables were selected by the SiPLSR-IRIV method, which achieved a 97 % variable reduction. Under the optimal conditions, the model attained a determination coefficient (R2C) of 0.9689 and root mean square error of calibration (RMSEC) of 0.0734. The limits of detection (0.41 μg g-1) and quantification (1.23 μg g-1) complied with EU regulatory standards for 3-MCPD detection in oils. This hybrid CNN approach demonstrated superior performance with smaller computational cost and narrower spectral data sampling range compared to the conventional methods, providing a robust solution for quality monitoring in edible oil processing industries.