Disruption precursor detection: Combining the time and frequency domains

This work analyses the recent evolution of statistical learning methods in JET for the prediction of disruptions. Disruption predictors are implemented as binary classification systems (two labels are possible: ‘disruptive’ and ‘non-disruptive’) whose training process is strongly related to the existence of disruption precursors in the signals. So far, the best predictors (in terms of success rate, false alarm rate and enough anticipation time) have used a combination of the time and frequency domains from the plasma signals to distinguish between disruptive and non-disruptive behaviors. Three different types of predictors are reviewed. Their difference is the amount of information that is needed to carry out the respective training processes: thousands of past discharges (for instance, the APODIS predictor), very limited data from previous discharges (for example, a particular APODIS version and Venn predictors) and no requirement at all from earlier shots.