Heat event detection in dairy cows with collar sensors: An unsupervised machine learning approach

The detection of heat (estrus) events in pasture-based dairy cows fitted with on-animal sensors was investigated using an unsupervised learning. Accelerometer data from the cow collar sensors were used in this approach where the aim was to identify increased activity level (restlessness, increased walking for mating) and to find association with recorded heat events. High dimensional time series data from accelerometers were first segmented in windows followed by feature extractions. The extracted features are standard deviation, amplitude, energy and Fast Fourier Transform (FFT). K-means clustering algorithm was then applied across the windows for grouping. The groups were labeled in terms of activity intensities: high, medium and low. An activity index level (AIxL) was derived from the activity intensity labels. We compared the AIxL with recorded heat events and observed significant associations between the increased activities through high AIxL values and the observed heat events.