Abstract
The purpose of this research is to construct the system which can extract and recognize action and a phenomenon dynamically from the sensor information. The system to propose divides arbitrary signal patterns with the short-term pattern based on time constant for searching events, and by making the shortest unit of a phenomenon as the "event primitive", a phenomenon can be recognized as an event primitive sequence. Moreover, on a higher order layer, system updates the time constant for searching and make a prediction based on known event primitive sequence, and has the composition that can perceive the information-unexpected according to the differential from putative event sequence by controlling background information and making active attention. On the other hand, when the information observed frequently, a gaze to the information of phenomenon falls and the system which aims at improvement in the recognition accuracy of an anomalous behavior by being buried in background information by proposed architecture. Proposed system extracts the features based on primitive pattern in the power-spectrum pattern in time series and classifying the observed phenomenon according to the event primitive, which is the shortest unit of the event.
