Tyshchenko O. Multivariate nonlinear time series intelligent analysis based on heterogeneous neural networks

The thesis is devoted to architectures development for multivariate nonlinear time series processing with the help of heterogeneous neural networks. A predictive neuro-fuzzy architecture based on reservoir computing techniques is proposed. It demonstrates high speed parameter setting and has linear outputs and high processing speed. A multidimensional neo-fuzzy-neuron architecture is proposed which possesses improved approximation properties and its learning algorithm which is characterized with simple realization. It provides smaller computing complexity by means of membership functions amount lessening. A predictive heterogeneous neural network architecture is proposed that consists of a fuzzy clusterization layer and a neuro-neo-fuzzy counter propagation network which is characterized with high processing speed and improved approximation properties. It provides nonlinear time series processing of arbitrary nature under uncertainty conditions. A predictive heterogeneous neural network architecture of compressed data based on multilayer perceptron "bottle-neck" is improved due to joining a neuro-neo-fuzzy counter propagation network to the neuro-compressor. This computational system can predict reduced dimension time series without quality loss of connections inside data. A multivariate time series segmentation method based on a fuzzy clusterization method by means of a special center setting is improved. Now it can segment data in an online mode and also detect time series changes of characteristics. The effectiveness of the proposed heterogeneous neural network architectures has been proved while solving practical problems. The results of the dissertation work can be used for solving intelligent analysis tasks of multivariate nonlinear time series of different nature processing characterized with a short learning sample in an online mode under a priori and current uncertainty.