Algorithms for the identification of yield anomalies in cattle dataset collected by automatic milking systems

Despite the growing interest in new animal housing, management strategies for reducing impacts, and collecting daily data, there is a lack of studies investigating factors that can lead to productive anomalies. On the other hand, the use of automatic milking robots, milking parlours, collars and pedometers allows for a precise monitoring of dairy cows, providing farmers with real time information. In this context, the early detection of production anomalies is fundamental for animal health and safety. In this work, two numerical methods for detecting daily milk production anomalies are presented and applied to three different farms selected as case studies. The methods described in this paper provide a numerical procedure having the scope of detecting milk yield anomalies. Both the algorithms presented hereinafter are based on statistical calculations and take as input daily resting time and daily milk yield recorded respectively by the pedometers worn by the cows and by the automatic milking system of the barns. The first method take into consideration two indicators, namely the Difference in Relative Production (DRP) and the Daily Rest time (DR). DRP is defined as the relative difference in daily milk yield between real-time data of a single animal and a baseline curve considered as an ideal trend. An anomaly (i.e. a deviation from the normal value) is determined, for a single cow, for a specific day, if two conditions on DRP and DR are contemporary verified. In the second method, starting from the Wood function, maybe the most famous model to fit the production of the cow in dependence of day in milk, the concept of reliability of robust statistics has been introduced in order to obtain, for each animal, a more solid and realistic lactation curve since not affected by outlier values.