Name	Chronicle discovery by mining data streams
Theme	Machine learning for model acquisition
Begin	2005
State	Ongoing
Description	We are interested in diagnosing and monitoring methods which aim at recognizing, on signal streams, temporal patterns that can be related to specific interesting events. We are investigating learning from examples or mining from data streams where data arrive continuously at a very high rate. The volume of data is such that it is impossible to store them for offline processing. In fact, data can only be seen once and consequently they have to be processed on line. This is a challenge for monitoring systems using model-based techniques for diagnosis, prediction or prognosis. When observing a dynamic system, the underlying concept of the data may change and so the model must be updated. We are particularly interested in models composed of temporal patterns in the form of chronicles. In this case adaptation means to add or to retract patterns from the model or to add or to retract events from chronicles as well as to enlarge or to restrict temporal constraints. Learning from data streams and model adaptation is achieved in the framework of the Calicot platform and within the SéSur project. For the first action, our aim is to take into account the patient evolution. For the latter project, our aim is to take into account new attack methods which should be retrieved when some data evolution is noticed. A third application field is also investigated: zootechnic data recorded on dairy cattle. After applying signal processing algorithms in order to improve the data quality, a methodology close to the one we have used on cardiac data for chronicle learning is being used. The results are currently evaluated and discussed by veterinary experts.