Exploring the Potential of Machine Learning Algorithms Associated with the Use of Inertial Sensors for Goat Kidding Detection.

The autonomous identification of animal births has a significant added value, since it enables for a prompt timely human intervention in the process, protecting the young and the mothers' health, without requiring continuous human surveillance. Wearable inertial sensors have been employed for a variety of animal monitoring applications, thanks to their low cost and the fact that they allow less invasive monitoring process. Alarms triggered by the occurrence of events must be generated close to the events to avoid delays caused by communication latency, which is why this type of mechanism is typically implemented at the network's edge and integrated with existing auxiliary mechanisms on the Internet. Although the detection of births in cattle has been carried out commercially for some years, there is no solution for small ruminants, especially goats, where the literature does not even report any attempts. The current work consisted of a first attempt at developing an automatic birth monitor using inertial sensing, as well as detection techniques based on Machine Learning, implemented in a network edge device to assure real-time alarm triggering. Thus, two concept drift detection techniques and seven kidding detection mechanisms were developed using data classification models. The work also includes the testing and comparison of learning results, both in terms of accuracy and of computational costs of the detection module, for algorithms implemented. The results revealed that, despite their simplicity, concept drift algorithms do not allow kidding detection, whereas classification-algorithm-based static learning models do, despite the unbalanced character of the dataset and its reduced size. The learning findings are quite promising in terms of computational cost and its suitability for deployment on edge devices. The algorithm demonstrates behavior changes four hours before kidding and allows for the identification of the kidding hour with an accuracy of 61%, as well as the capacity to improve the overall learning process with a larger dataset.