Вирішення регресійної проблеми демодуляції сигналів з амплітудною модуляцією багатьох складових

Abstract

Розглянуто підхід до демодуляції сигналів з амплітудною модуляцією багатьох складових (АМБС), який ґрунтується на використанні регресійної нейронної мережі. Пояснено принцип дії попередньо розробленого алгоритму генерації сигналів із АМБС, висвітлено проблематику визначення вирішуваного завдання, запропоновано демодуляційну нейронну мережу на основі цього алгоритму, її протестовано і запроваджено метод валідації.

The article is devoted to the consideration of AMBS features, highlighting the methodology of AMBS signal calculation for presentation in the form of a signal constellation and time graphs, and the use of calculated signals as input data for training a neural network that performs the task of signal demodulation. To represent sets of random values of different symbols of AMBS signals, a method was proposed, the essence of which is the use of Voronoi cells as a way of dividing the space between the points of the signal constellation, which is more efficient from a geometric point of view, compared to how signals are represented in trivial information transmission systems. The theoretical increase in the efficiency of the proposed method was calculated in comparison with the trivial approach assuming a higher efficiency of Voronoi cells as a way to divide the space between points. The described methods and techniques were embodied in the algorithm of the software product, which performs the task of forming the AMBS constellation, creating noisy variations of the signal around the points, recording these variations in a file, which is later used in the training of the neural network. The principle of operation of the software product based on previously formed algorithms is described, the algorithms themselves are described, their effectiveness is evaluated, the design decisions of the software product structure are explained, in particular, attention is paid to flexibility and the possibility of adjustment for specific cases. It is described with what data and in what form the created system operates. The efficiency of the created system was evaluated using relatively high values of added noise in the analyzed signal. Conclusions are drawn regarding ways to maximize system efficiency, and the dependence of accuracy on various model parameters is depicted. The algorithm for assessing the accuracy of the prediction of the neural network was formed, implemented in the form of a subroutine of the software product, the accuracy of the proposed system was evaluated, and conclusions were drawn about the work done.