Exploration of total synchronous fluorescence spectroscopy combined with pre-trained convolutional neural network in the identification and quantification of vegetable oil

• Published in: Food chemistry 2021-01, Vol.335, p.127640-127640, Article 127640
• Main Authors: Wu, Xijun, Zhao, Zhilei, Tian, Ruiling, Gao, Shibo, Niu, Yudong, Liu, Hailong
• Format: Article
• Language: English
• Subjects: Chemometrics; Convolutional neural network; Identification of vegetable oil; Total synchronous fluorescence spectroscopy
• ISSN: 0308-8146; 1873-7072
• DOI: 10.1016/j.foodchem.2020.127640

•Vegetable oil samples were characterized by total synchronous fluorescence spectra.•The convolutional neural network was used for spectral analysis by transfer learning.•The different ways of using convolutional neural networks were compared.•A deep learning model for a small amount of spectral data was established.•The deficiency of traditional algorithms in spectral analysis was made up. In order to distinguish different vegetable oils, adulterated vegetable oils, and to identify and quantify counterfeit vegetable oils, a method based on a small sample size of total synchronous fluorescence (TSyF) spectra combined with convolutional neural network (CNN) was proposed. Four typical vegetable oils were classified by three ways of fine-tuning the pre-trained CNN, the pre-trained CNN as a feature extractor, and traditional chemometrics. The pre-trained CNN was combined with support vector machines to distinguish adulterated sesame oil and counterfeit sesame oil separately with 100% correct classification rates. The pre-trained CNN combined with partial least square regression was used to predict the level of counterfeit sesame oil. The coefficient of determination for calibration (Rc2) values were all greater than 0.99, and the root mean square errors of validation were 0.81% and 1.72%, respectively. These results show that it is feasible to combine TSyF spectra with CNN for vegetable oil identification.