Research On Correlation Model Between Fermented Milk Detection Data And Artificial Senses Based On Data Mining

Fermented milk is a kind of dairy product which is fermented by adding Lactobacillus after pasteurization of milk and other animal milk.Because of its unique flavor and health care function,the consumption has continued to increase in recent years.At present,in the production process of fermented milk,the main method of sensory quality control is artificial sensory evaluation.Because of its subjectivity,the artificial sensory evaluation method is difficult to reproduce and quantify,which limits the fermentation milk production enterprises to further improve the yield.In view of the shortcomings of artificial sensory evaluation methods,this paper proposes three solutions:1)In view of the shortage of artificial sensory evaluation to meet the current needs of fermented milk production,the electronic nose was used to detect fermented milk samples,and the detection data of fermented milk was obtained,and the artificial sensory evaluation was used to mark it.Using bagging integrated learning method to model it,reduce the variance of single model,and generate the mapping model of fermentation milk electronic nose detection data and artificial sensory index.The experimental results show that the bagging integrated learning method based on BP neural network achieves 96.67% accuracy in the testing data set of fermented milk electronic nose,and the classification performance is better than the traditional single model,which can accurately classify fermented milk samples into three classes,and preliminarily realize the auxiliary and replacement of artificial sensory evaluation of fermented milk.2)In order to further improve the classification granularity of sensory evaluation model of fermented milk,so that it can more accurately describe the sensory characteristics of fermented milk samples,on the basis of single electronic nose detection data,combined with a variety of additional detection methods,using colorimeter and acidity titration to detect fermented milk samples respectively,obtain additional sensory data dimensions,and build data set.In order to deal with more complex data sets and to deal with more precise classification objectives,the stacking integration model with stronger integration performance is used,and the stacking integration model is constructed based on four heterogeneous classifiers: k NN,support vector machine,random forest and XGBoost.The experimental results showed that when the feature dimension of dataset is increased and the classificationtarget is promoted to four classifications,the stacking integrated classifier can still achieve96.55% classification accuracy,which proves that it is feasible to integrate multiple detection methods and use the model with stronger learning ability to improve the performance of sensory evaluation model.3)In view of the problem that the sensory evaluation model of fermented milk based on data mining has poor interpretability and can not describe the correlation between sensory and test data in detail,Pearson correlation coefficient,random forest feature importance measurement,XGBoost feature importance measurement were integrated,and three different correlation analysis angles were combined.and seven key sensors with strong correlation with six kinds of fermented milk senses which are milk flavor,typical yoghurt flavor,cream flavor,sour milk flavor,sharp sour milk flavor and peculiar smell were found out,and the correlation between specific sensory indexes and electronic nose detection data was established,which provides ideas for the future fermented milk manufacturers to develop portable and simple electronic nose system.