Contrastive Learning Of Near Infrared Spectroscopy And Its Application In Milk Adulteration Detection

Near-infrared spectroscopy is a mature material component detection method,which has a wide range of applications in agriculture,food,pharmaceutical,chemical and other fields.Due to the serious overlap of near-infrared spectral peaks,data dimensionality reduction and feature extraction algorithms are needed.Among them,principal component analysis and variational autoencoder are commonly used algorithms for extracting low-dimensional features,and they are very useful when there is a single influencing factor in the data set.However,when other background factors in the data set interfere seriously,conventional dimensionality reduction algorithms and feature extraction algorithms will only obtain background information features,such as changes in instruments(model,usage environment,etc.)and changes in samples(type,batch,etc.),causing model delivery problems.Contrastive learning is a new type of machine learning algorithm that has been proposed recently.The main idea is to obtain the unique target features of one dataset relative to the other through the comparison between two datasets,which is expected to be used to solve the calibration transfer.This algorithm has been popularized and applied from the field of biomedicine to the field of agriculture and food testing.It is used for the analysis of near-infrared pesticide residues affected by fruit types and the quantitative analysis of nitrogen content in fertilizers produced by different manufacturers,and has achieved good results.However,the current main algorithms for contrastive learning applied to NIR calibration transfer are limited to the use of linear models(contrastive principal component analysis)and the analysis of solid samples,and the application of nonlinear models and the analysis of liquid samples is still lacking.At the same time,there is a lack of a more systematic analysis of the qualitative and quantitative capabilities of related algorithms.Therefore,this paper takes the hot issue of melamine adulteration in milk as the breakthrough point,discusses the calibration transfer problems encountered in related research work,and hopes to solve this problem by applying linear and nonlinear algorithms to extract target features at the same time.The study found that melamine adulteration information was masked by irrelevant contextual information about the brand and batch when the instrument was unchanged.This paper will eliminate these irrelevant information through contrastive learning,realize the qualitative detection of melamine in milk of different brands and batches,and analyze the quantitative effect of related target features.The work of this paper mainly includes the following points:(1)The linear contrastive learning algorithm,contrastive principal component analysis(c PCA,contrastive Principal Component analysis),was used for the detection of melamine in pure milk affected by brand and batch.The experimental results verify that the linear contrastive learning algorithm can extract the key information of melamine adulteration in pure milk under the influence of brand and batch.(2)The nonlinear contrastive learning algorithm,contrastive variational autoencoder(c VAE,contrastive Variational Autoencoder),was used for the detection of melamine in pure milk affected by brand and batch.The experimental results also verified that the nonlinear contrastive learning algorithm can extract the key information of melamine adulteration in pure milk under the influence of the brand and batch.(3)Combining two contrastive learning algorithms—contrastive principal component analysis and contrastive variational autoencoders with quantitative algorithms(partial least squares regression),respectively,to determine the specificity of melamine adulteration in brand-influenced pure milk.Quantitative analysis of the content was carried out,and the results were compared with the ordinary partial least squares regression algorithm.The relevant experimental results show that the two contrastive learning algorithms cannot effectively improve the quantitative effect of melamine,but will make the model worse.Some of these influencing factors are analyzed,and some specific limitations of contrastive learning for quantitative aspects are explained.This paper mainly uses two kinds of contrastive learning algorithms to analyze the related calibration transfer problems they face when milk samples change,and through experiments,analyzes the effects of melamine doping in both hypothetical and quantitative aspects.Contrastive learning can effectively improve the qualitative ability,but cannot improve the quantitative effect,in the adulteration detection calibration transfer problem of milk changes.This paper also discusses this issue.