Feasibility study on Raman spectra-based deep learning models for monitoring the contamination degree and level of aflatoxin B1 in edible oil

Abstract

Edible oil is one of the three major nutritional sources of human body, and its quality is directly related to human health. This study proposes two novel deep learning architectures to achieve qualitative and quantitative monitoring of the degree and level of aflatoxin B1 (AFB1) contamination synthetically of edible oil in laboratory level. Based on Raman spectra acquired, the structures of a convolutional neural network (CNN) and a recurrent neural network (RNN) were designed, respectively; and deep learning models were established to achieve qualitative identification of the AFB1 contamination degree and quantitative detection of the AFB1 contamination level of edible oil samples. The results showed that both the CNN model and the RNN model showed perfect recognition performance when identifying the contamination degree of edible oil samples by the AFB1, and their recognition accuracy reached 100% when predicting the independent samples in the prediction set. The RNN model has better detection performance than the CNN model when detecting the AFB1 contamination level of edible oil samples. In the prediction set, the coefficient of determination (RP2) and ration of prediction to deviation (RPD) of the RNN model were 0.95 and 4.86, respectively. The feasibility study results demonstrate that deep learning combined with Raman spectroscopy can achieve high-precision monitoring of edible oil mycotoxins. In addition, deep learning has a good promising tool in the field of spectral chemometrics analysis.