Study On Identification Technology Of Honey Quality Based On Raman Spectroscopy And Deep Learning

The consumption of honey is increasing due to its own unique flavor and powerful health properties.The imbalance between increasing consumer demand and the limited availability of natural honey has contributed to the phenomenon of honey fraud,and honey has become the third most fraudulent food in the world.And long-term consumption of fake honey has been shown to have a negative impact on consumers’ health.There is therefore an urgent need to develop a rapid and robust detection method.Compared with the traditional chemical methods,the Raman spectroscopy does not require any sample pretreatment process and has the characteristics of green,fast,non-invasive and cost-effective,and it is widely used in food fingerprint analysis.Raman spectroscopy is also suitable for the detection of honey samples because it is less susceptible to the interference of moisture in the samples during the measurement.In addition,as a complex food matrix,honey often has the complex and abstract spectra with serious overlapping peaks,and the traditional chemometric algorithms are limited in their ability to resolve the spectral datasets with nonlinear and collinear characteristics.Therefore,the study employs two-dimensional correlation spectroscopy and deep learning techniques to better amplify,explore,and excavate the spectral features most relevant to the specific tasks.The main research of this article is as follows:(1)Different honey samples were prepared,and the corresponding Raman spectra were collected for the common fraudulent methods of honey,i.e.,adulteration of exogenous sugars,mislabeling of floral or geographical origin and adulteration of cheap honey in high-value honey.The same preprocessing operations were performed on the spectra and the overall visualization of the datasets was achieved by the dimensionality reduction algorithms.(2)For the sugar adulteration and botanical origin fraud in honey,the qualitative and quantitative models of convolutional neural network with similar structures were constructed to achieve the identification of floral origins and the direct prediction of adulteration concentrations for honey samples,based on Raman spectroscopy and without considering both honey and syrup types.The applied models achieved the best performance in both identification and quantification tasks compared to the traditional chemometric and machine learning algorithms.(3)Due to the high similarity of components between the high-value and cheap honey,there were no spectral changes related to the adulteration concentrations observed in the conventional Raman spectra of honey samples.The visualization and quantification results based on the Raman spectra and two-dimensional correlation spectra demonstrate that two-dimensional correlation spectroscopy provides a positive effect on the resolution and extraction of spectral features.The deep residual shrinkage network combined with synchronous correlation spectra achieved the concentration prediction of inexpensive honey,obtaining the best performance compared to the chemometric algorithms and several deep learning models.The study demonstrates that Raman and two-dimensional correlation spectroscopy combined with deep learning techniques can achieve the detection of honey fraud,and the proposed spectroscopic approach provides a non-destructive and efficient alternative for the honey quality control.