Research On Near Infrared Detection Method Of Frying Oil Quality Based On Machine Learning

After several times of frying,the frying oil will not only affect the taste of the fried food,but also reduce the quality of the frying oil.As frying times increases,the amount of harmful substances in the oil increases,which in turn leads to a serious decline in the quality of the frying oil.Therefore,this study proposes the use of near-infrared spectroscopy(NIRS)technology combined with machine learning methods to detect the quality of frying oil under standard experimental operation.The frying oil quality was converted to the frying times under standard experimental operation,and the random forest and partial least squares method models were established from two different perspectives of machine learning classification and regression.Firstly,this paper designed the standard frying experiment operation.The process was based on soybean oil as frying oil and frozen fries as frying material,and ten frying operations were carried out.Each frying operation was repeated frying 15 times without adding new oil,and spectral data of 15 different frying times were obtained.Secondly,the first derivative(D1),second derivative(D2),standard normal variable transformation(SNV)and multivariate scatter correction(MSC)are used to preprocess such spectral data,and the optimal pretreatment method is selected based on the results of classification and regression.At the same time,according to the different angles of analysis,different feature wavelength selection methods are adopted.The random forest model uses the correlation coefficient method to extract the characteristic wavelengths,and the partial least squares model uses the forward interval partial least squares(FiPLS),the backward interval partial least squares(BiPLS)and the genetic algorithm(GA)to extract the characteristic wavelengths.Finally,it is found that the random forest model uses the D1 and correlation coefficient method to process the spectrum,and the obtained model has the highest accuracy.The train accuracy(TRA)is 100%and the test accuracy(TEA)is 93.33%,and the out-of-bag data error rate is 0.04148.For the partial least squares model,when the spectral data is processed by the combination of SNV,FiPLS and GA,the model has the best effect.The determination coefficient(R~2)is 0.9994,the root mean square error of prediction(RMSEP)is 0.1060 and the ratio of performance to standard deviate(RPD)is 40.7021.The results show that the two different types of modeling methods used in this paper can quickly and accurately identify the quality of frying oil,and also provide a convenient and efficient method for the detection of frying oil quality.