In the dissertation work, an actual scientific and applied problem in the field of computer science is solved: improving the accuracy and reducing the time of constructing diagnostic models, increasing their interpretability and generalizing abilities by developing new and improving existing methods for synthesizing recurrent neural network diagnostic models that combine the principles of intelligent and parallel computing.
The dissertation consists of an introduction, four chapters, conclusions, a list of references, and appendices.
In the introduction, the relevance of the topic of dissertation research is justified, the purpose, object and subject of research are given. Thus, the aim of the work is to develop and research methods and tools for synthesizing diagnostic models based on recurrent neural networks that combine intelligent and parallel computing to improve the accuracy of neuromodels, their level of data generalization and interpretability. The object of research is the process of synthesizing diagnostic neuromodels based on historical data. The subject of the research are methods for constructing diagnostic models based on recurrent neural networks.
In the first chapter, "Overview of the problem area and formulation of the research problem", the state of the diagnostic problem based on historical data about the object is considered. Various types of artificial neural networks that can be used as the basis for diagnostic models in solving diagnostic problems are analyzed. The process and methods of synthesis of such diagnostic neuromodels are analyzed. Significant disadvantages of existing neuroevolution methods for the synthesis of neuromodels are investigated. The necessity of developing new methods for synthesizing diagnostic models based on recurrent neural networks is justified.
In the second chapter, "Synthesis of recurrent artificial neural networks based on the neuroevolution approach", the actual problem of synthesis of diagnostic models based on recurrent neuromerges is solved. Methods for the synthesis of neuromodels based on the neuroevolution approach are proposed, which allow performing the process of synthesis of neuromodels with a high level of accuracy, which can be used as a basis for the synthesis of models for non-destructive diagnostics based on historical data about the system and object. A modified genetic method for the synthesis of recurrent neural networks has been developed, which, unlike existing methods, uses a method of encoding information about a neuromodel based on sequencing at the coding stage, for more compact data storage. A parallel genetic method with the implementation of selective pressure mechanisms is proposed, in which the main stages of the process of evolutionary synthesis are performed on parallel computing nodes, selective pressure mechanisms and uniform crossing are used, which allows reducing the size of the population, without taking into account and without processing those individuals from the population that differ in a small value of the fitness function.
In the third chapter, "Structural optimization of synthesized neural network models", the actual problem of choosing mechanisms for additional adjustment of neural network based on an indicator system for assessing the level of complexity of the problem is solved. A method of structural adjustment of synthesized neuromodels based on the neuroevolution approach using a system of indicators and criteria for adaptive determination of mutational changes is proposed. A system of indicators for assessing the level of complexity of the problem for further modeling, diagnostics or forecasting is proposed. Thus, based on taking into account the characteristics of the input data set, the level of possible simplification of the structure, the total number of significant and insignificant factors, the level of measurement accuracy and the level of possible control and management, it is possible to choose the option of additional adjustment of the mathematical model in the future. A method of structural adjustment is proposed to improve the characteristics of neuroplasticity of networks. For example, when working with big data, the use of complex neural network topologies is often limited to computing resources. This is why structural optimization allows us to obtain a simplified and thinned structure of a pre-synthesized neuromodel.
In the fourth chapter, "Experimental study of methods for synthesizing recurrent neural network models for diagnosing complex technical objects and systems", an experimental study of the developed methods for data analysis and synthesis of diagnostic neuromodels is performed.
