The Mapping Ability Of Artificial Neural Networks And Their Applications In The Study Of Space Physics

This dissertation may be roughly divided into two parts. The first part introduces the theory of artificial neural networks and the second part investigates the applications of artificial neural networks in the prediction of space environments.In part 1, this dissertation firstly discusses the theoretic knowledge of artificial neural networks and their mapping ability. It reviews the developing history of artificial neural networks, analyses the architecture of artificial neural networks and learning rules and learning algorithms of artificial neural networks. It deeply studies the back-propagation algorithm of feed-forward networks, discusses the network structure's influences on the mapping abilities of neural networks. Secondly, this dissertation minutely investigates the effects of neural networks' algorithms and the option order of training signals on the mapping abilities of neural networks. We compare of convergence speed and networks output errors of tangent planes to constraint surface algorithm and gradient descent algorithm in the training of neural networks and draw the conclusion that the tangent planes algorithm converges speedily under the same circumstance. We also study the impacts of the order of training signals used to train neural networks on the convergence speed of neural networks, obtain the important results that there don't exist any intrinsic relations of the order of teaching signals and the convergence speed and the convergence speed of neural networks is mainly controlled by the topology architecture, the neuron number of the network, the quality of teaching signals and many other factors.In part 2, we solve two nonlinear practical problems in the study of space physics by using the mapping ability of artificial neural networks. One is the ionosphere model building in the region of the South China Sea. There exists complicated ionosphere structure in the equatorial South China Sea. We precisely select the components of the independent variable phase space by the correlated dimension analysis of the monthly medians of the ionosphere critical frequency and use them as the input neurons of the neural network. The plentiful historical ionosphere observation data are used to train the neural network. The mapping relation of these components and the ionosphere critical frequency, which is also called the regional ionosphere model, is then founded. The mapping relation can be used to predict the critical frequency and the HF communication channel. Another is the prediction of monthly mean number of sunspot. Because of the deficient knowledge of solar physics and the lack of observation of solar activity parameters, we develop a method of combination of reconstruction of phase...