Page 7 - Detection and prediction of a beam-driven mode in field-reversed configuration plasma with recurrent neural networks

Nucl. Fusion 60 (2020) 126025 C. Scott et al
Figure 6. Schematic of C-2U confinement chamber, marked with position of magnetic probes. Reprinted from [28], with the permission of AIP Publishing.
Figure7. Resultsfor40runsofvariableimportanceexperiment.Eachrunconsistsoftherandomuniformselectionofasubsetofvariables upon which to train. Some variables are more important than others to accurate prediction of the staircase phenomenon, indicating that the behavior measured by the corresponding probes may play a role in the staircase instability. In particular, the successful training runs (those in the upper branch of the plot, in blue) all used prior excluded flux radius information, indicating that there may be trends in this signal which could be used to predict staircase bursts with some lead time. Orange runs did not use this signal.
(a) create a binary ‘mask’ of length n, where n is the total num- ber of input variables to the network;
(b) everywhere where the mask is a zero, replace that variable with its mean over the whole training dataset. Train and evaluate the network, as before.
The mean accuracy over all runs where a diagnostic was
present provides us with a rough measure of how important that diagnostic is to the the network’s ability to classify
windows as staircase or non-staircase. Furthermore, we restrict our random selection of masks so that all of a given type of probe are included or excluded (so, for example, the mask indices for Bdot probes are either all 0 or all 1 for a given mask). Forty runs of this experiment, with different randomly chosen masks, are plotted in figure 7. We clearly see that some variables are more important for the network to make this dis- crimination than others. Some collections of variables allow the model to reach accuracy values comparable to original,
7