Application of multivariate analysis of sensor data for the detection of metabolic disorders in dairy cows

Abstract

Sensor measurements, such as milk yield, concentrate intake, activity, rumination and body weight, can be used to detect metabolic disorders in dairy cows. In an on-farm study, different types of sensor were installed and evaluated on three dairy farms during a period of up to two years. On two farms, cows were milked in an automatic milking system, and the third farm used conventional milking. Each farm was equipped with a different set of sensors. Sensor data were pre-processed to convert values per measurement into one aggregated value per cow per day. Alerts for a single variable were based on a change in level (e.g. decrease in activity), a change in trend (e.g. unexpected decrease in milk yield) or a difference in index (e.g. excessive body weight before calving compared with the average weight). To improve test characteristics, a combined alert was created by using the number of individual alerts per cow per day. A combined alert was triggered when the number of alerts exceeded a threshold. Farmers' recordings of diseases were used as reference data. This resulted in a sensitivity (percentage of detected cases) of up to 83% for metabolic disorders. The specificity (percentage of healthy cows classified as healthy) was too low for practical application. Alerts were combined in Standard Operating Procedures to improve the test characteristics. Other multivariate analysis methods might lead to further improvements in sensitivity and specificity.