The project aims to develop methods for the early detection of health problems or stress through the behavior of an animal recorded by sensors. We will study data sets of dairy cow activity level available from the French partners. The final goal is to implement these methods in precision livestock farming tools. A first machine learning method, to detect anomalies in activity rhythm, was designed using a Fourier transform with promising results. These methods will be now further developed aiming to improve the reliability and accuracy of detection. In particular, we will study methodologies to refine the registration of time series in order to better compare them, relate the magnitude and direction of changes in the Fourier harmonics causing the anomaly (e.g., overactivity in case of mastitis, less marked circadian rhythm in case of lameness) using higher order harmonics and wavelet transform that are better to capture ultradians (i.e. peaks of activity during the day). Fuzzy logic will also be studied to take into account the progressive onset and disappearance of disorders.