RECURRENT NEURAL NETWORK FOR CONTROLLING THE SPECTRUM WIDTH OF A NON-STATIONARY RANDOM SIGNAL

Abstract

The purpose of the study is to develop a recurrent neural network for detecting the moment of the beginning of the transient process of a random non-stationary signal in a sliding window. The possibility of using the apparatus of artificial neural networks (ANN) in analyzing non-stationary random processes is investigated. Rapid detection of the moment at which a non-stationary random process changes its behavior is an urgent task of electrical engineering. Materials and methods. A comparison is made between the use of the autocorrelation function and a neural network algorithm based on a recurrent neural network to control non-stationary noise. Results. The novelty of the study consists in developing a new algorithm for estimating the spectrum width of a non-stationary random signal based on the use of the ANN apparatus. It is shown that recurrent neural networks are capable of processing the original signal. They do not require special pre-processing and data preparation. A study of the quality of ANN operation depending on the parameters of the signals and the size of the sliding window was carried out. The ways to improve the architecture of the neural network and to enrich the data to improve the quality of the classifier are proposed. Conclusions. It was found that there is an optimal ratio between the time of detecting a change in signal parameters and the size of the sliding window, which imposes restrictions on the choice of the latter and the method of applying the trained neural network model.