Research On Identification Of Vegetable Oil Quality Based On Fluorescence Spectroscopy And Deep Learning

Edible vegetable oil is an essential part of human diet.The health of consumers will be seriously endangered if there are quality problems in vegetable oil.Therefore,the identification of vegetable oil quality has always been an important topic in the field of food safety.As an important optical detection technique,fluorescence spectrum scanning is widely used in the detection of complex systems such as vegetable oils.However,the current fluorescence spectral modeling still relies on the preprocessing based on prior knowledge for feature selection,which makes their universality meet challenges in spectroscopic analysis.Therefore,this paper studies the deep learning combined with fluorescence spectroscopy to complete the few samples learning modeling of vegetable oil quality identification.The main research is as follows:(1)Total synchronous fluorescence spectrum sequence data combined with different recurrent neural networks to distinguish the types of typical vegetable oils.Four typical vegetable oils were scanned by total synchronous fluorescence spectroscopy,and the fluorophores corresponding to characteristic peaks were analyzed.Based on the influencing factors of fluorescence spectroscopy,two data augmentation methods were studied to increase the amount of spectral data,and five deep network structures with different principles were designed.Four typical vegetable oils were identified through training these network architectures.(2)Total synchronous fluorescence spectrum image was combined with Vision Transformer network to identify adulterated sesame oil.The pure sesame oil and adulterated sesame oil samples were scanned by total synchronous fluorescence spectroscopy.Data augmentation methods were adopted to expand the data set,and the spectral data were represented as images.A Vision Transformer network architecture for fluorescent image characteristics of adulterated vegetable oil was established.The enhanced data set was used to train the Vision Transformer network to identify adulterated sesame oil.(3)The pre-trained Efficient Net-B0 was transferred and combined with the excitation emission matrix fluorescence spectrum image to realize the traceability and quantitative analysis of counterfeit sesame oil.Pure and counterfeit sesame oils were scanned by excitation emission matrix fluorescence spectra and characterized as images.In the discrimination of counterfeit sesame oil,Efficient Net-B0 was targeted to fine-tuned to improve the performance of the discrimination model.Efficient Net-B0 innovatively combines partial least squares regression algorithm to quantitatively analyze the degree of counterfeit sesame oil.The research results show that the study of fluorescence spectroscopy combined with deep learning to complete few sample learning has certain significance for the identification of vegetable oil quality.