Ruban O. Diagnostic models and methods of their reduction in continuous dynamic objects identification systems

The dissertation is devoted to solving the actual scientific and technical problem – to increase reliability and efficiency of diagnosis of nonlinear dynamic objects with continuous characteristics through the development of methods of constructing diagnostic models and creating diagnostic software based on correlation methods of information model reduction. The analysis of the current state of the problem of identification of nonlinear dynamic objects information models with continuous characteristics, systematized methods of their reduction, and identified the features of their use. The choice of direction for research in building diagnostic models on the basis of correlation methods of information model reduction and their information optimization procedure for creating automated technical diagnostics systems has been substantiated. The information connection between non-linear dynamic models in the form of Volterra series and neural networks with time delays has been established; the method of evaluation of higher-order Volterra kernels by means of neural networks with time delays using activation function in the form of rectified linear unit has been developed. This allowed to increase the learning speed of the time-delay neural network and improve the accuracy of identification of information models in the form of Volterra series. The method of diagnostics of nonlinear dynamic objects with continuous characteristics has been further developed by using diagnostic models of objects based on correlation estimations of information model counts for reduction of space of diagnostic attributes and increasing of efficiency of diagnostics and procedures of information optimization of diagnostic models based on statistical methods of machine learning for increasing of reliability of diagnostics. To increase reliability of diagnosing in case of nonnormal distribution of learning sampling data, method of statistical classification is improved on the base of Bayesian approach for construction of diagnostic models in the form of second order polynomials by using power transformation of educating sampling attributes. Computer simulation tools for building diagnostic models of nonlinear dynamic objects with continuous characteristics in identification systems to solve resource-intensive tasks of diagnosis have been developed. On the basis of the specified algorithms the effective means of computer modeling in structure of the automated system of technical diagnosing for the decision of resource-intensive problems of diagnosing, provide high computing capacity and speed of search of decisions: system of technical diagnosing of electric motors for the organization of resource-saving operation of the process equipment, information system of identification of models of cutting process for prediction of residual working resource. With the help of the developed computer simulation tools the actual applied problems of technical diagnostics were solved. The practical significance of these results lies in the development of computational algorithms for identification of diagnostic models of nonlinear dynamic objects, extends the class of important practical problems of technical diagnosis, for which effective solutions can be constructed to objects with continuous characteristics in terms of growing dimensionality of the feature space.