Rapid recognition of processed milk type using electrical impedance spectroscopy and machine learning

Abstract

SummaryUnscrupulous merchants would sell cheap, low‐nutrition formula milk powder as pasteurised milk, or use it as a raw material in dairy products, for the purpose of making a profit. Currently, biochemical methods are utilised to identify the type of processed milk, which could involve chemical reagents, sample preparation and costly instruments. This paper investigates the utility of electrical impedance spectroscopy (EIS) in distinguishing different types of processed milk. Ultra‐high temperature sterilised milk, pasteurised milk and formula milk powder were studied. The one‐factor analysis of variance revealed that milk type had a significant influence on electrical properties, whereas sensory analysis, pH and total soluble solids showed poor discrimination of milk type. Recognition models were developed using Support Vector Classification, Partial Least Squares‐Discriminant Analysis (PLS‐DA), Gaussian Naive Bayes and Random Forest (RF). The nested cross‐validation and the external validation set (N = 45) indicated the RF model had the highest predicting accuracy of 97.8%; whereas PLS‐DA had a misclassification rate of 33.3% on positive samples and was, therefore, unsuitable for the authentication. The results demonstrated that the EIS data was featured enough for a small sample modelling (N = 75), and the established method was simple, fast and effective for liquid food classification.