Synthesis Method of Robust Neural Network Models of Systems and Processes

The work deals with the study of some problems of reconstruction multidimensional statistical dependences on the basis of empirical data by means of artificial neural networks. To provide stability (robustness) of systems and processes statistical model parameters on the basis of trained artificial neural networks (ANN) at the a priori input data uncertainty as well as practically sufficient accuracy of data approximation, it is appropriate to use stable (robust) methods of deep ANN training methods. The work uses the cricking model for statistical data since we couldn’t get precise parameter values; therefore, to achieve the required accuracy, some probability was introduced. The synthesis method of scalar convolution of selection functions for mathematical model identification, based on the law of requisite variety (Ashby law), Kolmogorov power overage concentration and the maximum likelihood principle, where Student and Romanovski statistics, are used as the proximity measure of true multidimensional samples. It makes it possible to structure preference systems of a person who makes decisions for multi-criterial problems to identify mathematical models in determinate and stochastic formulations (MV-, MH- problem). Neural network identification was made by the stochastic approximation method on the basis of ravine conjugate gradient method. The method of effective robust estimation of statistical model systems parameters was worked out by employing a regularizing sequential (adaption) algorithm for synthesis of solutions with deferred correction. Samples of Rosenbrock function data and corresponding parameters of aerodynamic characteristics of the jet engine multistage axial compressor were taken as examples.