Data mining technique for fast retrieval of similar waveforms in Fusion massive databases

Fusion measurement systems generate similar waveforms for reproducible behavior. A major difficulty related to data analysis is the identification, in a rapid and automated way, of a set of discharges with comparable behaviour, i.e. discharges with "similar" waveforms. Here we introduce a new technique for rapid searching and retrieval of "similar" signals. The approach consists of building a classification system that avoids traversing the whole database looking for similarities. The classification system diminishes the problem dimensionality (by means of waveform feature extraction) and reduces the searching space to just the most probable "similar" waveforms (clustering techniques). In the searching procedure, the input waveform is classified in any of the existing clusters. Then, a similarity measure is computed between the input signal and all cluster elements in order to identify the most similar waveforms. The inner product of normalized vectors is used as the similarity measure as it allows the searching process to be independent of signal gain and polarity. This development has been applied recently to TJ-II stellarator databases and has been integrated into its remote participation system.