Radial Basis Neural Network Construction and Training for Telegraph-Code Structure Reception

Abstract

The use of neural network classification algorithms for solving the problem of receiving telegram-code structures is considered. The article provides comparison of the neural network classifiers analyzing the normalized input signal as well as the signal after the binary conversion. Various measures of the code distance in the space of informative features are considered. Recognition comparative results for the selected pair of symbols are given. On the basis of these results the code distance is determined, which ensures the minimum recognition error probability. The results obtained in the developed neural network classifier are compared with those obtained in correlation receivers operating in the signal time and frequency domains. The advantage of neural network algorithm is shown. The structure implementing the developed neural network classifier is provided. It is shown that the procedure for the classifier developing, k \ including selection of information signs and their amount, as well as code distance, is not of general nature and is to be performed for each set of recognizable symbols. It is stated that to generalize the received alphanumeric blocks it is necessary to use the second decision contour where current information on the reception and information on the duration of the observed symbol is supplied, which is the subject of further research.