This thesis describes studies conducted on the subject of detecting oestrus and lameness in dairy cows. The studies comprise methods of statistical change detection and model based diagnosis, respectively. In the case of statistical change detection the development of algorithms for a decision support system is based on identifying behaviour from patterns of normal and deviant behaviour. Signal processing combined with statistical methods, e.g. likelihood ratio tests, are utilized to correlate observed behaviours with normal and detect changes. Diagnosis includes data from the available population of animals in order to isolate patterns of behaviours outside the norm for individuals, while being robust to common disturbance factors. The research is based on methods from change detection and fault diagnosis. Fault diagnosis techniques are employed to reduce the false alarm ratio, and attempts are made to isolate events and artefacts in signals that otherwise can give rise to false alarms. For the model based diagnosis the diagnosis is generally done evaluating an estimated probability distribution against hypotheses about causes of change behaviour, e.g. oestrus or lameness. The models used for diagnosis are chosen to represent the behaviours. A quantized system description is used as a diagnostic model. This technique is based on automata theory. The methods are in most cases specified to take into account parameters specific to the differences between production systems. The development of these methods and algorithms is an interdisciplinary activity including methods from fault diagnosis, information technology and statistics.