Recognition of Images of Continuous Wavelet Spectra of Noised Radio Location Signals Using a Convolutional Neural Network

Abstract

The paper considers the existing methods of image recognition of continuous wavelet spectra of noisy signals with linear and nonlinear frequency modulation using convolutional neural networks. A procedure for preparing spectral images for processing in a neural network is proposed, which ensures a sufficient probability of recognizing a given type of signal out of twenty possible ones. The methodology for solving the problem consists of finding an image preparation algorithm that provides image augmentation by the method of changing continuous wavelets, which ensures the identification of signals under conditions of limited resonance frequency and bandwidth. The algorithm involves changing the frequency of the continuous spectrum by processing the phase grating signal with different continuous wavelets after the additive addition of non-stationary noise. Signals with linear and nonlinear modulation prepared in this way, as well as signal spectra of other regular forms, are used as input data of the convolutional neural network. The procedure of dividing wavelet spectrum images into classes is performed by checking the homogeneity of the class based on the Shannon entropy value. The minimum entropy value indicates the homogeneity of the subset and the absence of “impurities” from images of other classes. The developed model of a neural network with augmentation by continuous wavelet spectra in the conditions of a limited data set has an accuracy of up to 97.95%. Keywords: augmentation, wavelet spectrum, convolutional neural networks, continuous wavelets, unmanned aerial vehicles.