A neural network approach to the classification of electron and proton whistlers

Abstract

Fractional whistlers, whistlers, and proton whistlers are automatically identified and characterized by means of a neural network. A feed-forward neural network with Time Delay Neural Network (TDNN) architecture is used. It has the ability to represent structures in frequency time diagrams; a set of 50 spectrogram elements (5 Fourier components × 10 time intervals) serves as input to the network. Applications to date have used ELF data recorded on board the low-altitude AUREOL-3 satellite. A first neural network was designed to identify and characterize fractional whistlers and whistlers. A set of 997 vector data is used for the training phase and 1088 other vector data are used for evaluating performance. It is observed that fractional whistlers and whistlers can be distinguished from noise with an accuracy of 90%. A second neural network, with the same architecture, was used for studying proton whistlers. Although the training database contains less examples, the accuracy of the classification is 89%. Neural networks of this type could be used in satellites for real-time classification and characterization of electron and proton whistlers.