Font Size: a A A

A Study On Feature-Level Identification Of Active Deception Jamming Based On Deep Learning

Posted on:2020-09-21Degree:MasterType:Thesis
Country:ChinaCandidate:H M LiuFull Text:PDF
GTID:2428330602451946Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
With the development of modern electronic spectrum jamming technology,the working environment of radar is becoming more and more complex,especially the flexible-active deception jamming brings great challenges to the normal work of radar.In order to ensure that the radar can recognize all the kinds of interference contained in the received signal accurately and quickly under the interference of the jamming signal,and implement the most correct anti-jamming method in the changeable electromagnetic environment,it is the primary purpose and research foundation of our research on radar signal classification,recognition and anti-jamming.The active deception jamming of conventional radar is studied.Based on the analysis of its generating mechanism and working characteristics,the synthetic perception method of active deception jamming of radar is discussed.Using the theory of multi-dimensional transform domain and multi-scale decomposition,the feature-level classification and recognition of three kinds of deception jamming and their mixed jamming are carried out,and the related algorithms in depth learning are combined.The classification and recognition of three kinds of deception jamming signals are studied.The follows are main research results of our thesis:1.Firstly,the generation mechanism and working characteristics of three kinds of conventional radar active deception jamming is studied,including Range-Gate and Velocity-Gate Pull Off,and range-velocity-gate pull off,and establishes a signal model during the drag period,which provides a reliable theoretical premise for the subsequent jamming recognition algorithm based on feature extraction and deep learning theory.2.A multi-domain and multi-feature combination method is introduced,which combines the mean,variance and waveform entropy characteristics of three radar active deception jamming signals,frequency domain moment kurtosis and frequency domain moment skewness,and wavelet domain features such as high-frequency components and low-frequency components decomposed by multi-scale wavelet,and extracts the distinction of features under different Jammer-to-noise ratios.The features with better degree and volatility selectivity are combined to form a feature vector.Combined with principal component analysis(PCA),the quadratic feature of effective dimension after dimensionality reduction is extracted as the feature vector for classification and recognition,and then the support vector machine(SVM)is used for training and testing.The simulation results show that the SVM trained by the second feature after dimensionality reduction can effectively classify and recognize three kinds of active deception jamming,and the classification and recognition results are better.3.In order to obtain the difference of the deep characteristics of three kinds of deception jamming signals,considering that the deep learning model has better mining ability for deep information,so the Convolutional Neural Networks(CNN)method is used to realize the effective recognition of three kinds of deception jamming signals.In this thesis,two kinds of deception jamming signals are constructed using Pytorch framework.The structure of convolution neural networks with different layers can be expressed as follows: 1.Five-layer network,with two layers of convolution layer and pooling layer and the last layer is a full connection layer;2.seven-layer network with three layers of convolution layer and pooling layer,and the last layer is also a full connection layer.The two networks are trained and tested with interference data to realize the classification and recognition of three kinds of signals.Comparing the two sets of simulation experiments,it is found that the classification and recognition of three kinds of deception jamming signals using convolutional neural networks are mainly related to the network complexity,loss function,the number of input and output channels at each level and the training times of training samples in training set.These parameters are analyzed and optimized separately.Under a certain Jammer-to-noise ratio,the two sets of convolutional neural networks can identify the three kinds of dragging jamming signals.The recognition result is better,and the classification accuracy is much higher than that of using superficial signal features directly.This shows that convolutional neural network is also effective and practical in dealing with one-dimensional signal data such as active deception jamming.
Keywords/Search Tags:Active Deception Jamming, Feature Extraction, Multi-scale Decomposition, Deep Learning, Convolutional Neural Network
PDF Full Text Request
Related items