Quantitative analysis of multi-component adulteration in camellia oil by near-infrared spectroscopy combined with long short-term memory neural networks algorithm

Abstract

Deep learning algorithms have provided new alternatives for the analysis of multi-component adulteration in foods with considerable attention recently. Near-infrared spectroscopy (NIRS) was employed to combine with various neural network algorithms such as convolutional neural networks (CNNs) and long short-term memory (LSTM) neural networks for quantitative identification of multi-component adulteration in camellia oil. The findings showed that the developed LSTM regression model for predicting the adulteration level in camellia oil exhibited satisfactory accuracy and excellent generalization ability. The sample datasets partitioning based on joint x-y distance (SPXY)-savitzky-golay smoothing (SG)-LSTM model corresponds to camellia oil adulterated with maize oil and soybean oil (CMS) (determination coefficient of the prediction datasets (R²p)= 0.9920, root mean square error of the prediction datasets (RMSEP)= 0.0264, residual predictive deviation of the validation set (RPDv)= 6.67); the SPXY-SG second derivative (SG-SD)-LSTM model corresponds to camellia oil adulterated with rapeseed oil and maize oil (CRM) (R²p = 0.9377, RMSEP=0.0716, RPDv=4.02); and the SPXY-standard normal variate (SNV)-LSTM model corresponds to camellia oil adulterated with soybean oil and rapeseed oil (CSR) (R²p = 0.9504, RMSEP=0.0547, RPDv=4.39). As above, the findings provide new support for the development of modeling methods for multi-component adulteration in vegetable oils, which contributes to promoting the quality control of vegetable oils.