Synthesis of Informative Features for Recognition of the Type of Pulse Repetition Interval Modulation of Signals from Radars

The modern development of radio and radar facilities leads to the need to create special radio monitoring tools to monitor the electromagnetic spectrum. In the process of radio monitoring there is a need to detect and identify (recognition) of radio sources, in particular radars. The peculiarity of pulse radars is that the emitted radar signals can have both intra-pulse and inter-pulse modulation. There are quite a number of methods for recognizing types of pulse modulation (based on artificial neural networks, logical rules base, wavelet analysis, histogram analysis). Each of these methods uses its own set of features for recognition, which are certain parameters of the registered signal or the results of their special transformations. The effectiveness of recognition methods directly depends on the set of selected recognition features. The aim of the study is to select a dictionary of informative features for recognition of the following types of pulse repetition interval modulation (PRI) signals of radar: constant, stagger, sliding, dwell&switch, jittered, sin. To achieve this goal, two approaches were used: the choice of informative features using decision trees and the synthesis of informative features using an auto-associative neural network with a narrow throat. These approaches work well at the decision of tasks of recognition of radio signals. As a basic set of recognition features, statistical features of the distribution of the values of the pulse intervals of the radar signal proposed in [13], [14] were used: J
₁
, J
₂
, J
₃
, f
₁
, f
₂
, f
₃
, f
₄
, f
₅
. The analysis of different approaches to the selection of informative features showed that the mathematical apparatus of decision trees provides a better result compared to auto-associative neural networks. It is advisable to use features J
₁
, J
₂
, J
₃
, f
₂
to recognize PRI modulation. At the same time, auto-associative neural networks allow to minimize the feature space without significant deterioration of its separating properties. Due to the reduction of the characteristic space at the stage of pre-processing of radio monitoring data, it is possible to reduce the computing power at the stages of identification of radio sources.