Dynamic clustering and neuro-fuzzy identification for the analysis of fusion plasma signals

Measurements in long pulse devices like ITER require the use of intelligent techniques to detect interesting events and anomalous behaviors within a continuous data flow. This detection will trigger the execution of some experimental procedures such as: increasing sampling rates, starting data sampling in additional channels or notifying the event to other diagnostics. In a first approach, an interesting event can be any non-average behavior in the expected temporal evolution of the waveforms. Therefore, a model of the signals is needed. In this work, a model that represents each type of plasma signal is obtained by means of fuzzy inference systems (FIS) which are generated by applying adaptive neuro-fuzzy techniques. The purpose of this neuro-fuzzy modeling is to identify patterns of these groups of data to produce a concise representation of a signal. Previously the signals have been preprocessed and a new dynamic clustering strategy based on a partitioning method has been applied to obtain the clusters. Off-line analyses have been applied to bolometric signals of the fusion device TJ-II Stellator with encouraging results.