Research On Edible Oil Identification Based On Neighborhood Rough Set Attribute Reduction And Spectral Information

The safety detection of edible oil is a persistent hot topic in the food field,and its existing detection methods have problems such as limited range of detection indexes,and the integration and intelligence of equipment need to be improved,which lead to the inability to form an efficient detection system.Therefore,new theories and technologies for edible oil detection need to be explored.In this work,common four kinds of edible oils are used as the research objects,and the spectral information of the oils is obtained by Fourier transform spectrometer and fiber optic spectrometer to establish the data set.Based on the Neighborhood Rough Set Attribute Reduction(NRSAR)theory and technique,the dimensionality reduction processing and feature selection of oil product information are completed.Subsequently,optimized random forest and extreme learning machine algorithms are established to construct oil product identification models and compares them the results of principal component analysis.The classification performance of the models is finally evaluated using multiple indicators of confusion matrix to determine the optimal model suitable for oil product identification.The main contents and results have the following four parts.(1)The Fourier transform spectrometer was used to collect the visible spectra of colza oil,peanut oil and waste oil in the range of 400 to 800 nm.In addition,the oil types were increased and a fiber optic spectrometer with a wider wavelength band was used to collect oil spectral information from 178 to 871 nm for colza oil,peanut oil,soybean oil and waste oil.The primary comparison and analysis revealed a high similarity of the spectral curves of each of the oils,but there still existed differential spectral information,especially in the VIS region,which reflected minor differences in the composition of functional groups and trace elements contained in various edible oils.(2)For the oil information obtained by Fourier transform spectrometer,the NRSAR method was used to reduce 193 wavelengths to 10 characteristic wavelengths,and the wavelength number compression rate reached over 88.61%.The accuracy of the created NRS-RF and NRS-ELM models were 91.67% and 93.33%,respectively,while the PCA-RF and PCA-ELM models were 81.67% and 90.00%,respectively.When compared with the feature extraction methods,the NRSAR method has pointing advantages in terms of accuracy,sensitivity,specificity and misidentification distribution.Therefore,the NRSAR method and the NRS-ELM model have excellent results in the spectral identification of edible oils.(3)For the above four edible oils,the data measured by using fiber optic spectrometer,NRSAR can be used to approximate the wavelength number from 694 to 18,and the accuracy of models NRS-RF and NRS-ELM is 91.67% and 94.17%,respectively,which is better than PCA-RF and PCA-ELM models of principal component analysis method.When examining the accuracy rate,which is the core index of oil spectral identification,ELM algorithm is better than RF algorithm.Therefore,the feature selection method implemented with NRS technology in oil spectral recognition runs faster and has higher accuracy than the traditional feature extraction method.(4)Several evaluation indexes,such as accuracy,sensitivity,specificity and misidentification distribution,were used to evaluate the discrimination ability of each model for edible oil,and it was determined that the Neighborhood Rough Set Attribute Reduction Algorithm combined with Extreme Learning Machine Model(NRS-ELM)could be the best model for the detection of spectral information of edible oil.The three innovative points of this thesis are as follows.(1)Successfully extended the rough set theory to the field of edible oil identification,using the NRS method instead of the traditional method for feature selection of oil spectral information to achieve oil identification,which was highly evaluated by the editors of the well-known journal Applied Optics.(2)In the field of edible oil detection,the differential spectral information of each oil in the UV and IR regions has received attention from domestic and foreign researchers,while the VIS spectral information of oil is less studied.Therefore,this work refines the oil identification scheme of the spectroscopic method,which provides the basis for the full waveband study of oil information.(3)The model optimization was achieved by k-fold cross-validation technique,multiple indicators were used to evaluate the performance of the model for classification,and a preliminary oil product identification scheme based on Rough Set Theory was established.In conclusion,the NRSAR method can effectively handle the spectral information of edible oil with high redundant features,and the characteristic wavelengths obtained from it are combined with K-fold cross-validation to optimize the classification model,and multiple indicators of confusion matrix are used to evaluate the model,and the NRS-ELM model is finally established as the optimal model for oil detection.This work not only tapped the wider spectral information of oil products,but also successfully applied the Rough Set theory and technique to the field of oil spectral information,and innovated the method for the subsequent identification of edible oils.