Lobachev I. Models and methods for improving the efficiency of distributed transducer networks based on machine learning and peripheral computing.

Українська версія

Thesis for the degree of Doctor of Philosophy (PhD)

State registration number

0822U100271

Applicant for

Specialization

  • 122 - Комп’ютерні науки

29-12-2021

Specialized Academic Board

ДФ 41.052.025

Odessа Polytechnic State University

Essay

The qualification scientific work is devoted to the development of models and methods for improving the efficiency of distributed transducer networks based on machine learning and peripheral computing. The analysis of the peculiarities of the use of distributed transducer networks in the construction of various Internet of Things systems, as well as methods of data processing and analysis in these networks was carried out. The paper also analyzes the advantages and disadvantages of typical architectures of distributed transducer networks in the use of deep neural networks as a tool for data analysis. It is shown that the functioning of modern IoT systems is based on the analysis of the readings of many sensors of different types, with different data dimensions and frequency of obtaining new data. To solve this problem, a method of pre-processing and formatting of sensor readings for their further use as input data in the proposed hierarchical neural network model is proposed. It is shown that the structures of typical models of neural networks can be redundant to solve various machine learning problems. To solve this problem, a method of neural network compression is proposed to reduce the requirements for computing and energy resources by reducing the number of connections between network neurons. A hierarchical neural network model of peripheral computing for multisensory data analysis has been developed, which allows to organize neural network computations on a hierarchical principle without the use of a centralized cloud server. A method of forming an input tensor from a set of indicators collected by sensors of different types over a period of time has been proposed. A method of compressing neural networks was developed to reduce the requirements for computing and energy resources by reducing the connections between network neurons. Programming tools have been developed to implement the proposed solutions. On the basis of the developed tools the software component for modeling of work of the distributed transducer networks in various areas of application of systems of the Internet of Things is created.

Files

Similar theses