Optimization of identification of non-stationary objects due to information properties and features of models

Abstract

Constructive approaches, principles, methods of optimal identification of random time series of non-stationary objects with mechanisms of using statistical parameters, dynamic characteristics, specific features, and data patterns are proposed. Models and algorithms for hybrid identification have been developed that combine the capabilities of statistical, dynamic models, and a three-layer neural network. Mechanisms for adjusting model variables are implemented, taking into account the structural complexity, stochasticity of links between the elements of the random time series, ambiguity of dynamics, a large number of variables, the impact of the external environment. In the solutions of optimization problems, probabilistic and soft computing, self-organizing, predictive, adaptive properties of neural networks and dynamic modeling are used. A generalized algorithm for the identification of non-stationary objects is built on the basis of the mechanisms for identifying segments, boundaries, the general interval of element values, the selection of informative elements, and the formation of a training set. A simplified computational scheme for identifying non-stationary objects is implemented on the basis of 4-order orthogonal polynomials, a cubic extrapolation spline function, a linear Kalman filter, and a three-layer neural network. A software complex for identification in the C ++ language in the parallel computing environment “CUDA” has been developed and implemented.