Rapid classification of whole milk powder and skimmed milk powder by laser-induced breakdown spectroscopy combined with feature processing method and logistic regression

Abstract

Whole milk powder and skimmed milk powder are suitable for different groups of people due to their differences in composition. Therefore, a rapid classification method for whole milk powder and skimmed milk powder is urgently needed. In this paper, a novel strategy based on laser-induced breakdown spectroscopy (LIBS) and feature processing methods combined with logistic regression (LR) was constructed for the classification of milk powder. A LR classification model based on mini-batch gradient descent (MGD) was employed first. As indicated by the research results, the accuracy of the MGD-LR model for the milk powder samples in the test set is 96.33% and the modeling time is 33.07 s. The modeling efficiency is low and needs to be improved. Principal components analysis (PCA) and mutual information (MI) were used as feature processing methods to reduce the high dimensional LIBS data into fewer features for improving the modeling efficiency of the classification model. The research results indicate that the accuracy of the PCA-MGD-LR model and the MI-MGD-LR model for the test set of milk powder samples was 99.33% and 99.67%, respectively. Compared with MGD-LR model, the modeling efficiency of PCA-MGD-LR and MI-MGD-LR models has increased by 89.7% and 74.8%, respectively. The results fully demonstrate the feasibility of rapid milk powder classification based on LIBS and feature processing methods combined with LR, and it will provide a new technology for the identification and classification of milk powder.

Keywords: Laser-induced breakdown spectroscopy; Logistic regression; Milk powder; Mutual information; Principal component analysis.