Model-driven multivariate control chart and support vector machine as tools to detect variation in the milking process and monitor parlor performance

The efficiency of the milking process is the key to dairy farm management. However, due to the high variability of data from single or multiple milk meters, it is difficult to know whether the milking process is under control or not. The main objectives of this study were to develop a model to assess whether the milking process in herringbone and parallel parlors is out of control (i.e., deviates significantly from normal milking patterns) and to filter out the out of control points from the historical milking parlor data. To this end, 16 milking parlors equipped with herd management software and electronic milk meters were included in a 1-yr longitudinal field study and model-driven multivariate control charts (MDMVCC) were used in conjunction with a support vector machine (SVM). The milking variables that contributed most to uncontrolled milking were, in this order: percentage of kick-off (31.9%), percentage of multiple attachments (28.5%), percentage of irregular take-off (16.5%), percentage of milk produced in the first 2 min (10.8%), percentage of milking time spent in low milk flow (<1 kg/min; 7.9%), and average milk flow (kg/min; 4.4%). Total parlor performance, expressed as milk yield/stall per hour, was on average 7.7% higher when the milking process was under control than when it was out of control. The results indicate that MDMVCC in conjunction with SVM appear to be a potentially useful tool to detect variations in the milking process and monitor performance in herringbone and parallel milking parlors.