Research On Key Technologies Of Intelligent Anti-jamming Communications ——Physical Layer Anti-jamming Technology Based On Machine Learning

Cognitive communication countermeasures have become one of the important components of modern electronic warfare.In order to obtain electromagnetic spectrum rights,the realization of intelligent anti-jamming communication is of great significance to electronic warfare.The whole signal processing flow can be summarized as the perception,recognition and processing of the signal.For the jammer,cognitive jamming needs to have the ability of waveform tracking,waveform design,dynamic change and knowledge storage of communication signals,so as to achieve targeted jamming and hinder communication of enemy signals with less power.For the anti-jamming party,cognitive anti-jamming needs to complete the accurate perception,identification and parameter extraction of the interference signal.According to the characteristics of the jamming signal,the anti-jamming party should make an intelligent anti-jamming decision with the least cost to restore the communication.The process of cognition and recognition is similar to the construction of deep neural networks.The invisible correlation between the signal and the environment and the characteristics of the signal itself can be extracted through the deep network;the decision-making process is similar to reinforcement learning,and each decision depends on the current situation.Changes in the disturbance environment and future disturbances.This thesis considers the application of machine learning in intelligent anti-jamming communication systems,aiming at the problems of unfamiliar and solidified waveform design of traditional wireless communication systems,limited design freedom,insufficient adaptive anti-jamming ability,and it is difficult to deal with the problem of unconventional and efficient interference.There is a need to design a communication endto-end and frequency-hopping anti-jamming system that can explore adaptive response to the effects of channel,noise,and interference in complex and variable electromagnetic environments.This thesis firstly constructs an end-to-end communication system model based on neural network.The system includes four neural network parameterization modules: encoder,decoder,channel estimation and channel simulation.In the forward propagation process,the encoder is responsible for generating the waveform according to the prior information,and the decoder converts it into bit stream data and calculates the decoding loss.In the parameter update process,the decoding loss transmits the gradient information to the decoder,channel and encoder through the back-propagation algorithm;the encoder further updates the parameters according to the back-propagated gradient information and generates a new waveform.In the process of backpropagation,since the channel is usually difficult to write as a display function representation to propagate the gradient,a conditional adversarial generative network is used to simulate the channel,and the gradient information is fed back to the encoder.For the channel information needed in the adversarial generation network,a channel estimation neural network based on residual network is established in this thesis.The least squares estimation of the channel is used as the initial input,and the residual network is used to denoise without prior information to achieve channel estimation with lower error.For efficient modulation targeting jamming,an intelligent anti-jamming communication system is built based on neural network.Through joint constellation design,channel coding and power allocation,it has better performance in single-carrier and multi-carrier anti-interference communication.Finally,for the two kinds of dynamic interference,linear frequency sweep interference and random interference,this thesis constructs a model for the design of the optimal anti-jamming frequency hopping pattern using reinforcement learning.The interfering signal in the medium is usually unknown,and the model mainly adopts two model-free methods of Monte Carlo and temporal difference to learn the frequency hopping pattern.