Detection Of Saturated Fatty Acids In Edible Oil Based On Reflectance Spectroscopy And Deep Learning Network

When the intake of saturated fatty acids in edible vegetable oils is excessive,the risk of cholesterol and cardiovascular disease is increased.The reference values for different saturated fatty acids in different edible vegetable oils can be used not only to classify the kind of edible vegetable oils,but also to determine the degree of spoilage and to assess the degree of oxidation.Therefore,to establish a reasonable effective and convenient detection method of saturated fatty acid has important significance.In this study,a detection method of saturated fatty acids in edible vegetable oil was established by combining reflectivity spectroscopy and deep learning network,and the polynomial correction method was explored to solve the failure problem of spectral analysis model caused by operator differences.The definite research contents are as follows:(1)A method for the determination of saturated fatty acids in edible vegetable oil was developed by combining reflectivity spectroscopy and deep learning network.First of all,the reflectivity spectroscopy of seven edible vegetable oils,including rapeseed oil,soybean oil,sunflower seed oil,corn oil,olive oil,sesame oil and peanut oil,were measured at 350-2500 nm.The reference values of saturated fatty acids,such as palmitic acid,arachidic acid and behenic acid,in the oils were measured by GC-MS.Centralization(CEN),multiple scattering correction(MSC),standard normal variable transformation(SNV)and standardization(STA)were used to eliminate the interference such as spectral irrelevant information,noise and spectral scattering.Then,a novel two-dimensional spectral convolution regression network(S2DCRN)was constructed for the regression analysis of saturated fatty acids.At the same time,S2 DCRN and three kinds of machine learning were compared.Finally,sequence forward selection(SFS),uninformative variable elimination algorithm(UVE),Relief algorithm,random frog(RFrog),stepwise regression and genetic algorithm were used to select the spectra of important wavelengths,and then a more simple and robust analysis model was constructed.The experimental results showed that the S2 DCRN model had the best performance after the spectral pretreatment of edible vegetable oil.Although the analysis result obtained after the selection of variables is slightly poor,the size of the characteristic variable is only 1% of the original spectral data,which saves the spectral data collection and greatly reduces the complexity of the model,which is helpful for the subsequent research and development of portable simplified detection devices.In addition,the S2 DCRN model was applied to predict the content of other saturated fatty acids in edible vegetable oil,and a better prediction result was also obtained.(2)A polynomial correction method is used to solve the operator differences in the application of reflectivity spectroscopy.Firstly,the reflectivity spectroscopy of seven edible vegetable oils,including rapeseed oil,soybean oil,sunflower seed oil,corn oil,olive oil,sesame oil and peanut oil,in the range of 350-2500 nm were measured by different experimenters,and the reference value of the content of behenic acid in the oil was measured by GC-MS.In order to eliminate the spectral scattering interference in the oil sample before modeling,a multi-scattering correction algorithm is used to process the Spectral data.Then,in order to solve the problem of spectral analysis model failure caused by operator differences in the application of reflectivity spectroscopy,a new polynomial correction method,namely order adaptive polynomial correction(OAPC),is proposed.S2 DCRN and three kinds of machine learning were used to analyze the content of behenic acid.At the same time,use the bootstrapping soft shrinkage(BOSS)and random frog(RFrog)to select the important wavelengths of the spectrum,Pearson's correlation was used to analyze their correlation.Finally,in order to further discuss the influence of the selection method and number of correction samples on OAPC performance,random selection and Kennard-Stone(KS)algorithm were adopted to select data with a proportion of 1% to 10% at 1% step size respectively.The experimental results show that the CNN was the best model after the application of OAPC.The proposed method can effectively solve the problem of spectroscopy model failure caused by operator differences in the application of reflectance spectroscopy,and also provide a potential solution to the problem of spectral difference caused by equipment or environmental factors.In conclusion,the method combining reflectivity spectroscopy and deep learning network can realize the rapid detection of saturated fatty acid content in edible vegetable oil,and the combination of polynomial correction can effectively solve the failure problem of model caused by operator differences in the application of reflectivity spectroscopy.In addition,the simplified model based on the reflectivity spectroscopy with selected characteristic wavelengths can provide a reference for the future development of portable instruments.