Boryndo I. Structural-Parametric Synthesis of Convolutional Neural Networks in the Tasks of Production Process Automation

This dissertation explores and develops methods for the structural-parametric synthesis of convolutional neural networks (CNNs) for production process automation. The use of multi-criteria genetic algorithms (MCGA) for optimizing neural network architectures is substantiated to enhance their efficiency, accuracy, and computational performance. New approaches to parametric optimization are proposed, enabling the adaptation of neural network models to specific industrial tasks. Modern topological variants of CNNs, attention mechanisms, and spatial-channel reconstruction methods have been analyzed to improve real-time graphical data analysis. A comparative study of the proposed methods with existing automated neural network design approaches has demonstrated their advantages in classification accuracy, processing speed, and resource efficiency. The possibilities of applying the proposed methods in virtual and augmented reality technologies are considered, allowing for improved immersion and interactivity in visualization systems while reducing computational costs. The developed algorithmic solutions have been implemented as software using the TensorFlow and Keras platforms, confirming their practical significance and potential integration into real-world production processes. Chapter 1. A review of existing CNN architectures, their topological features, and primary optimization approaches is conducted. Key issues in structural synthesis and future directions for CNN improvement are outlined. Chapter 2. The integration of convolutional neural networks into virtual (VR) and augmented reality (AR) technologies is examined. Methods for enhancing immersion, interactivity, and VR/AR system performance through neural networks are analyzed. Major implementation challenges and possible solutions are identified. Chapter 3. The main structural components affecting CNN performance are identified. Mathematical models and algorithmic solutions for improving parametric adaptation are studied. An approach to automated design based on multi-criteria evolutionary algorithms is proposed, with justification for optimization criteria and algorithmic solutions that enhance real-time image recognition and processing efficiency. Chapter 4. The algorithmic and software implementation of the structural-parametric synthesis algorithm for CNNs is described and proposed. The synthesized network model was obtained, tested, and its qualitative characteristics were evaluated. The research results have been presented at international conferences and published in leading scientific journals. The proposed methods and algorithms can be used for the further development of automated neural network design systems and the expansion of their applications in industrial and information technologies.