Research Of Low-frequency Radiation Recognition Based On Deep Learning

Identification of radiation sources is one of the important research topics in the field of electronic confrontation.Even radiation source individuals under the same type are from the same batch,there will be differences in the composition of components.These differences caused by hardware will be reflected in the transmitted signal,so the signals produced by them will not be exactly the same.The key of individual identification is to extract all kinds of subtle features which can distinguish different radiation sources and select the appropriate classifier to give the recognition result.The main contents of the study are as follows:For the behavior modeling of radiation source,this thesis mainly studies the behavior modeling of class-D power amplifier circuit.Recurrent neural network and its improved models are introduced to behavior modeling of power amplifier.This thesis also introduces wavelet transform to encoder-decoder model,thus proposes Sequence to Sequence Wavelet model.The comparison of power amplifier modeling performance through four kind of behavioral models,Volterra-Laguerre,Long short term memory neural network,Sequence to Sequence model and Sequence to Sequence Wavelet model.For the feature extraction of radiation source,this thesis mainly studies feature extraction methods based on deep learning.On one hand,this thesis uses auto encoder of unsupervised learning to extract features from signals directly.On the other hand,this thesis studies the algorithm of higher order cumulant and bispectrum to extracts diagonal slice of bispectrum.Convolutional neural network is introduced to the extraction of radiation source features.This thesis uses convolutional neural network of Vgg-16 and Resnet to extraction features from bispectrum figures.For the classification problems,the support vector machines classifier,the Gradient Boosting Decision Tree algorithm and the Extreme Gradient Boosting algorithm are studied.Comparing the performance of three classifiers through experimental analysis of simulated signals and measured signals,a predominant classification effect was obtained.The behavior modeling,feature extraction methods and classification of low-frequency radiation sources are studied in this thesis which has a certain value in solving the problem of individual identification.