Abstract

Agriculture is currently undergoing a rapid transformation, driven by digitization. One aspect is heat detection in dairy cows. Manual detection is prohibitively time consuming, especially in larger farms, but since the advent of the Internet of Things, it is has become possible to continuously collect time series of cow health parameters that enable automation. Livestock farming features diverse processes and conventions all over the world, owed to different climate zones, farm sizes and local conditions, which in term is reflected in the cattle monitoring data. Conventional algorithms struggle to robustly detect heats from dairy biosignals out-of-the-box, instead requiring manual parameter adaptation on a case by case basis. This thesis presents two machine learning models based on feed forward and recurrent neural networks respectively, in an attempt to improve generalization of detection. While ultimately their performance falls short of what is required by productive use, experiments performed reveiled shortcomings in data labels, that, if adressed correctly, hold the potential for great improvements. Despite the lack of performance, the recurrent neural network model demonstrated that it was able to learn the underlying problem and could, with further improvements, achieve the desired outcome.