Comparison of Machine Learning Methods to Detect Anomalies in the Activity of Dairy Cows

Farmers need to detect any anomaly in animals as soon as possible for production efficiency (e.g. detection of estrus) and animal welfare (e.g. detection of diseases). The number of animals per farm is however increasing, making it difficult to detect anomalies. To help solving this problem, we undertook a study on dairy cows, in which their activity was captured by an indoor tracking system and considered as time series. The state of cows (diseases, estrus, no problem) was manually labelled by animal caretakers or by a sensor for ruminal pH (acidosis). In the present study, we propose a new Fourier based method (FBAT) to detect anomalies in time series. We compare FBAT with the best machine learning methods for time series classification in the current literature (BOSS, Hive-Cote, DTW, FCN and ResNet). It follows that BOSS, FBAT and deep learning methods yield the best performance but with different characteristics.