Neural network identification of ECR events - preliminary results based on synthetic data

Vlad Constantinescu, Octav Marghitu

Institute for Space Sciences, Bucharest, Romania


Neural networks (NN) provide a powerful pattern recognition tool, that can be used to search large 
amounts of data for certain types of 'events'. Our specific goal is to make use of NN in order to identify 
energy conversion region (ECR) events in the Cluster data, that is regions where E.J<>0 is rather well 
defined and observed on time scales from a few minutes to a few tens of minutes (E is the electric field 
and J the current density). The manual examination of the Cluster plasma sheet data from the summer of 
2001 provided a start-up set of several concentrated generator regions (CGRs, E.J<0) and concentrated 
load regions (CLRs, E.J>0), that we tried to use initially for training a feed-forward back-propagation 
NN. The input data was divided in intervals of fixed size (e.g. 100 points) and for each interval the target 
output of the network was an equally long vector, filled with 1 for CLRs, -1 for CGRs, and 0 in rest. 
However, using Cluster data for training the NN proved to have two disadvantages: first, the training set 
was limited, and second, the regions marked with 0 in the training set could not be explored later - in a 
consistent manner - for the presence of CLRs / CGRs. In oder to overcome these problems we considered 
using synthetic data for training and testing the NN. For the time being each input interval contains one 
ECR event, with random amplitude, duration, and sign, to which a certain noise level is added. We 
investigate the results provided by the NN, as used with synthetic data, depending on the size of the 
training set, the noise level, and the NN configuration.