ИСТИНА

The outer radiation belt of the Earth (ERB) is a part of inner Earth’s magnetosphere. The intensity of the flux of relativistic electrons in the outer ERB is subject to strong and abrupt changes under the influence of factors that are difficult to take into account. This kind of electrons may cause dangerous malfunction of the electronics on board spacecraft and therefore they are often called killer electrons. Prediction of the relativistic electron flux in the outer ERB is a complicated problem. This is due to the fact that Earth’s magnetosphere is a complex dynamical system. The state of this system can be described by a multi-dimensional time series, including various physical features - parameters of interplanetary magnetic field, solar wind, geomagnetic indexes, relativistic electron flux at geostationary orbit etc. Here we consider the approach to investigation of time series characterizing the dynamics of the outer ERB with the help of machine learning algorithms. There is every reason to believe that there may exist several different basic states of the outer ERB – significantly diverse regions in the phase space of the states of this dynamical system, for which the most efficient prediction models may be different. To separate out such regions, in this study we used segmentation of multi-dimensional time series with the help of k-means clusterization algorithm and Kohonen neural networks. The initial data which are a multi-dimensional time series with delay embedding were split into three, four, and five clusters (segments) with Kohonen self-organizing map and k-means algorithm. The same operations were repeated for different sets of input physical features. The obtained variants of segmentation of the time series were compared to each other and correlated with various possible states of the outer ERB.