Quantitative Analysis of 3-Monochloropropane-1,2-diol in Fried Oil Using Convolutional Neural Networks Optimizing with a Stepwise Hybrid Preprocessing Strategy Based on Fourier Transform Infrared Spectroscopy

As one kind of 'probable human carcinogen' (Group 2B) compound classified by the International Agency for Research on Cancer, 3-MCPD is mainly formed during the thermal processing of food. Tedious pretreatment techniques are needed for the existing analytical methods to quantify 3-MCPD. Hence, a nondestructive sensing technique that offers low noise interference and high quantitative precision must be developed to address this problem. Following this, Fourier transform infrared spectroscopy association with an convolutional neural network (CNN) model was employed in this investigation for the nondestructive quantitative measurement of 3-MCPD in oil samples. Before building the CNN model, NL-SGS-D₂ was utilized to enhance the feature extraction capability of model by eliminating the background noise. Under the optimal hyperparameter settings, calibration model achieved a determination coefficient (R²_C) of 0.9982 and root mean square error (RMSEC) of 0.0181 during validation, along with a 16% performance enhancement enabled by the stepwise hybrid preprocessing strategy. The LODs (0.36 μg/g) and LOQs (1.10 μg/g) of the proposed method met the requirements for 3-MCPD detection in oil samples by the Commission Regulation issued of EU. The method proposed by CNN model with hybrid preprocessing was superior to the traditional model, and contributed to the quality monitoring of edible oil processing industry.