Abstract:
Quality of raw milk is necessary to assure food safety, human health and the survival of the
dairy industry in Sri Lanka, though the existing approaches to quality assessment are mostly
manual, time consuming, and are not viable in making immediate and accurate decisions by
the dairy suppliers in the rural collection centers. The traditional tests like the Alcohol test,
Resazurin test, Fat percentage, Solid Non Fat (SNF) percentage and pH value involve laboratory
methods and experienced operators and result in a significant disparity in effective field-level
quality monitoring. This paper suggests a method to identify the status of milk quality using
periodically measured parameters of physicochemical tests on the milk samples at Sri Lankan
milk collection centers. Following the consultation with domain experts, reading of industry
guidelines, and referring to existing work major features such as Alcohol stability, Resazurin
reduction time, pH, fat content, SNF value, temperature, etc. were chosen as predictive features.
1056 milk sample data were collected from milk collection centers in the Uva Province of Sri
Lanka, and was preprocessed. Then the preprocessed dataset was trained and evaluated using
various Machine Learning (ML) models. This research mainly focuses on six ML algorithms
to predict the milk quality. Here, above 80% of accuracy has given by Deep Neural Network
(DNN) and XGBoost models, above 75% accuracy from SVM and Random Forest models
and Logistic Regression and KNN models give an accuracy of below 75%. The results of the
experiment show that DNN had the best accuracy followed by Random Forest and SVM, which
proves that machine learning has the potential to improve the quick and reliable quality of milk
in Sri Lanka.
