Establishment Of Adulteration Identification Models For Oil-tea Camellia Seed Oil Based On Machine Learning Algorithms

Due to the relatively high nutritional functional value and commodity price of oil-tea camellia seed oil(OCSO),the phenomenon of OCSO adulteration is becoming common in the market,which not only harms the interests of edible vegetable oil producers and consumers but also is not conducive to the healthy development of the edible oil industry in China.At present,the common adulteration methods of OCSO can be divided into two categories,one is to mix other kinds of low-quality and low-price edible oil into OCSO,the other is to mix the solvent extracted OCSO with poor processing technology into pressed OCSO.It is difficult to identify the adulterated edible vegetable oil accurately and quickly because of the wide range of adulterated oils and the complicated ways of adulteration.Aiming at the adulteration identification issue of the OCSO mixed with low-quality and low-priced edible oils,53 OCSO samples of different cultivars and origins were selected,whose fatty acid and triglyceride composition data were measured by gas chromatography and high-performance liquid chromatography.Aiming at the adulteration identification issue of the pressed OCSO mixed with solvent extracted OCSO,4 original pressed and 10 roasted pressed OCSO samples that were commercially supplied or obtained from the laboratory were selected,whose volatile component data were measured by headspace solid-phase microextraction(HS-SPME)and gas chromatography-mass spectrometry(GC-MS).Then,by different adulteration gradients,different kinds of edible oils were mixed into pure OCSO,and the solvent extracted OCSO was mixed into expressed OCSO,to build the high adulteration gradient model and the low adulteration gradient model.Then,based on the fatty acid,triglyceride,and volatile component data,machine learning adulteration identification models were established to identify the adulteration category and predict the adulteration concentration of OCSO by various machine learning algorithms,which achieved good results in both qualitative identification and quantitative prediction.The results are as follows:(1)Aiming at the adulteration category identification issue of OCSO adulterated with a single kind of low-quality and low-price edible oil,based on the fatty acid and triglyceride data of adulterated OCSO samples,a decision tree model and a multi-layer perceptron neural network model for identifying the adulteration category are built.The neural network model achieves a better identification effect.For the adulteration category identification of OCSO under high adulteration gradient and low adulteration gradient,its average identification precision reaches 0.977 and 0.981,and its average identification accuracy reaches 0.974 and 0.992,respectively.(2)Aiming at the adulteration concentration prediction issue of OCSO adulterated with a single kind of low-quality and low-price edible oil,a partial least squares regression model and a multiple linear regression model for adulteration concentration prediction are built based on the fatty acid and triglyceride data of adulterated OCSO samples.The multiple linear regression model has a more accurate prediction ability of adulteration concentration.Under high adulteration gradient and low adulteration gradient,the average R2 values of adulteration concentration prediction of the multiple linear regression model are 0.999 and 0.994,and the average RMSE values are 0.146 and 0.136,respectively.(3)To identify the adulteration of pressed OCSO with solvent extracted OCSO.a principal component analysis model and a logistic regression model are established based on the volatile component data of adulterated OCSO.The logistic regression model has a stronger identification ability.For the original and roasted pressed OCSO,the lowest detection limits of the logistic regression model are 10%under the high adulteration gradient and 4%under the low adulteration gradient.(4)To predict the adulteration concentration of pressed OCSO adulterated with extracted solvent OCSO,a partial least squares regression model is established based on the volatile component data of adulterated OCSO,which has a strong prediction ability.For original pressed OCSO,the average R2 values of adulteration concentration prediction of the partial least squares regression model for high adulteration gradient and low adulteration gradient samples are 0.998 and 0.956,and the average RMSE values are 1.127 and 0.592,respectively.For roasted pressed OCSO,the average R2 values of adulteration concentration prediction of the partial least squares regression model for high adulteration gradient and low adulteration gradient samples are 0.998 and 0.999 respectively,and the average RMSE values are 1.166 and 0.094 respectively.This paper covers two common types of OCSO adulteration phenomenon in the current market and is the first systematic study combined with machine learning methods in the research field of adulteration identification of OCSO.By comprehensively considering the characteristics of two kinds of common adulteration phenomenon,suitable machine learning methods are selected to establish the adulteration identification models,and good identification effects are achieved.Compared with the existing research on OCSO adulterated with low-quality and low-priced edible oil,this paper comprehensively considers a variety of common edible vegetable oils,and the established machine learning adulteration identification models have a large improvement in both identification accuracy and prediction accuracy.Aiming at the adulteration phenomenon of pressed OCSO adulterated with extracted solvent OCSO,current studies only focus on the qualitative identification of whether the samples are adulterated,while this paper is the first research to solve the adulteration concentration prediction problem and the minimum detection limit of qualitative identification is reduced to 4%.