Study On Honey Species And Adulteration Based On Multi-spectral Dimension And Machine Learning

Traditional honey variety identification and adulteration detection have some problems,such as complex operation,low precision,and few kinds of single identification,which can not meet the increasingly refined and rapid detection requirements of honey products.Hyperspectral imaging technology and transmission spectrum technology are the hot research directions of nondestructive testing in recent years,and their advantages are the high number of bands,complete information,and high resolution.In this study,hyperspectral imaging technology and transmission spectrum technology are combined with spectral analysis and deep learning to design a new method to identify honey quality quickly,accurately,and multi-dimensionally.In this study,the variety identification of Baihua honey,coffee honey,Vicia japonica honey,Sophora japonica honey,and jujube honey was started.At first,the hyperspectral data of five kinds of honey were collected,and the average spectral reflectance curves of five kinds of honey samples were obtained through the hyperspectral data.The spectral characteristics were extracted from the characteristic absorption peaks of the reflectance curves,the training set and the test set were divided,and a BP neural network model was established for honey variety identification.The results show that the original data collected by hyperspectral imaging technology carries complete feature information,and the accuracy of classification and identification is high.However,by combining the variety identification result graph and root mean square error(RMSE)analysis,it is found that the identification result of the seed honey is slightly less robust.Then,to identify the five kinds of honey more accurately and improve the generalization of the whole,the transmission spectrum data were collected by comparing the seed honey with two kinds of transparent honey at room temperature,acacia honey and jujube honey,which were used for multi-angle honey variety identification.Due to the complexity of ambient light,it is necessary to preprocess the transmission spectrum.In this study,three preprocessing methods are selected: standard normal variable transformation(SNV),multivariate scattering correction(MSC),and polynomial smoothing algorithm(SG).Because the dimension of transmission spectrum data is much higher than that of hyperspectral data,principal component analysis(PCA)is used to extract features from transmission spectrum data to build a classification model more quickly and eliminate redundant data.Finally,through the identification results of the BP neural network,it is established that the polynomial smoothing-principal component analysis-BP neural network model has the best identification effect,and the identification accuracy rate reaches 100%.From the research of honey variety identification,it is found that the data collected by hyperspectral imaging technology carries high information integrity and many dimensions,which is very suitable for the situation of fine identification.In view of this feature,hyperspectral imaging technology is used to further explore the feasibility of adulteration detection methods for multi-variety honey.The accuracy of the test set obtained by substituting RAW data into CNN model reaches 72.5%.Because of the noise and other errors in the collected honey spectral data,the detection effect of the model is directly affected,and the accuracy of the detection results is reduced.Therefore,six spectral preprocessing methods,namely centralization(MS),standardization(SS),first-order difference(D1),second-order difference(D2),standard normal variable transformation and multivariate scattering correction,are adopted to optimize the original spectral data,so as to improve the accuracy of the adulteration detection model.Then one-dimensional convolutional neural network(1DCNN)is selected for feature extraction,which simplifies the calculation of data and the running time of the model to achieve the purpose of rapid detection.Finally,the support vector machine(SVM)with good generalization ability is used to replace the Softmax classifier,which expands the application scenarios of identifying honey varieties and concentrations.After model identification,it is found that the first-order difference(D1)is the best pretreatment method,and the detection accuracy can reach 100%.To sum up,the model and pretreatment method adopted in this study can well realize honey variety identification and adulteration detection,and the detection method proposed at present can meet the basic conditions from experiments to market application and promotion.