부산대학교 도서관

Milk weight records are an important source of information for daily decisions on dairy farms. Most milking parlor systems have technologies that record milk production at cow level on a daily basis. However, a great number of dairy farms simply do not have this crucial information because they lack the necessary technology in the parlor systems. The aim of this study is to implement a machine learning-based milk weighing scale exploring the dynamics of milking flow, looking for an alternative strategy for low-cost mechanical milking systems, and/or parlor milking systems without appropriated hardware/software integrated. We developed a milk flow sensor integrated to the milking parlor system to compute the milk flow during milking, and installed some cameras to label real milk weight. The regression model based on machine learning predicted the milk weight with high precision, achieving minimum values of 2.81, 1.29, and 10.26 in terms of mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE), respectively, along with a maximum value of 0.84 in terms of the R2 index. We found that the area under milk flow curve (AUC), the number of milk flushes (MF), and the thirty-seventh to forty-fourth milk flow during milking have the most important influence in the efficacy of our proposed model. Milk flow of 250 cows milked 3 times a day were recorded during 12 days, generating 8,601 milk flow curves. Our preliminary results showed that we can confidently estimate the weight milk produced by a cow using machine learning algorithms to empower dairy farms to make better decisions.