Internet of Animals: Unsupervised Foaling Detection Based on Accelerometer Data

The coming of a foal is often one of the most stressful moments in a horse owners life. Lots of effort is being put in the monitoring of periparturient mares since 10% of foalings require human intervention to assure the well-being of both mare and foal. To remedy this, an unsupervised, machine learning based, anomaly detection algorithm is proposed to detect any abnormalities in a mares behavior that could indicate the onset of parturition. An accelerometer was attached to the mares halter to capture head movements during both regular behavior and parturition. This data was then used to train and evaluate the proposed autoencoder based anomaly detection algorithm. The proposed algorithm correctly identifies 100% of the foalings in the dataset. However, this comes at the cost of a high number of false alarms, on average 4 per mare in the 24-hour period before partus. By changing the decision threshold, and thus reducing the sensitivity of the decision, the accuracy dropped to 80% of foalings detected but with only 3 false alarms over the entire dataset of 10 mares.