Physics of auroral phenomena : proceedings of the 37th Annual seminar, Apatity, 25 - 28 February, 2014 / [ed. board: A. G. Yahnin, N. V. Semenova]. - Апатиты : Изд-во Кольского научного центра РАН, 2014. - 125 с. : ил., табл.

о 50' N.A. Barkhatov el al. 30' "8 i a 2 . 30' 2 Figure 1. Example of input data, wavelet spectrum and the corresponding of local maxima or "wavelet skeleton" of the spectrum (the result of post-processing) is demonstrated The resulting images of the wavelet skeleton for one minute data of density, velocity, temperature, pressure, module and component of the interplanetary magnetic field to quantify the consistency of the obtained spectrums was used. Obtained for this work skeleton images according to a previously known types of plasma flows (MC, CIR, Shocks, HSS) in order to obtain material for training and testing the classification neural network are grouped. Within each group looks for moments of coincidence skeletons (synchronizing the change of oscillation modes) for parameter sets flows is proposed. The presence and/or absence o f sync points within selected flow stream as a unique feature of this type is used. Subtraction of skeleton spectrums matrices from each other with record values of the difference in absolute value a labeled matrix are provides, where at times 100% synchronization corresponds to zero difference. The resulting picture with markers by a Gaussian filter to blur is processed. In this step unit, not grouped markers will be screened (blown out), and the existing group of markers will be visible (see Fig. 2). Then the histogram of the distribution likely to appear moments oscillation synchronization parameters in the stream on a scale from 0 (white, less likely) to 1 (black, most likely) we calculate. To obtain classification features of each type of individual flow histogram likely to appear moments oscillation synchronization parameters in this flow at given periods are summarized. Thus, the neural network are classified cumulative histogram comprising a compressed information of the stream. Л \ ( A if j & СЙ•d'V'n fX W i(ЫA -------------------------------------------------------------------------------- ! time a) в) jar— At w «> time, minutes Figure 2. Example of histogram for moments of oscillations synchronization for the flow parameters: (a) on the markers; (b) processed by a Gaussian filter; (c) within scale from 0 (less likely) to 1 (most likely) 3. Classification artificial neural network and the classification results Classification of obtained summary histograms with a self-learning neural network like Kohonen layer (see Fig. 3) is performed. This self-learning system, adjustable by self-imposed on the input data. In this case, the neural network generalization and compression imposes information performs. The number o f nodes (k) o f the neural network (the number of candidate classes or neurons in the competitive layer) in advance is given. In this case, the number of classes to four (MC, CIR, Shocks, HSS) is equal. Number of inputs ( n ) of each neuron in’ the classification layer by the size of the input image is determined. In our case it is 1400 (the total size of the histogram of each event). 76