Research On Deep Learning Based Signal Reiceiving Technology For Wireless Communication System With Mixed Noise

With the development in artificial intelligence and the improvement of hardware technology,deep learning(DL)has been used to almost all fields of wireless communication system for the past few years.The advantage of DL is that it can learn the distribution of samples from a large number of given data.Therefore,the DL-based wireless communication receiving system can extract channel characteristics from the received signals,and recover the sending symbols through the feature changes layer by layer without complex parameter estimation.Instead of the traditional communication system wich optimized module by module,the DL-based system is optimated for the whole transmitter and receiver and shows great potential.In many communication scenarios,the channel exists impulse interference besides white Gaussian noise,such as atmospheric noise in very low frequency/low frequency(VLF/LF)communication system,network interference in wireless communication,and radar clutter.The previous study of signal reception with impulse noise mainly focused on the scene of pure impulse noise,but in reality,Gaussian noise is inevitable.It is difficult to model the mixed noise channel with impulse and gauss noise and estimate its parameters.This dessertation focuses on the DL-based signal receiving technology,which can avoid those problems by training the DL network.In Chapter 1,the research background and significance are given first.Then the research status of tradition signal receiving technologies under the impulse noise and the application of DL in fields of communication is summarized.The research content and structure arrangement are introduced at the end of this chapter.In Chapter 2,for the binary phase shift keying(BPSK)signal and the minimum shift keying(MSK)signal under mixed noise channel,the statistical-based reception algorithm is discussed.Fisrt,the Symmetric ? Stable(S?S)distributed pulse noise model is introduced,and the mixed noise,whose model parameter estimator based on empirical feature function(ECF)and grid searching,is modeled as a mixure of S?S and Gaussian noise.Based on the estimated mixed noise PDF,the Maximum Likelihood(ML)receiving of BPSK and Viterbi Demodulation of MSK are given respectively.Finally,the bit error rate(BER)of BPSK and MSK under mixing noise is simulated.In Chapter 3,DL-based BPSK signal demodulation networks are studied.Fisrtly,the deep neural network(DNN),convolutional neural network(CNN)and long short-term memory(LSTM)network model used in this paper are introduced.Then according to signals' characteristics,the BPSK signal single symbol receiving and block receiving network is designed,and the BER of network is simulated.The simulation results show that when the test noise parameters are consistent with the training noise parameters,using the output of designed BPSK signal reception network to make hard decision,BPSK signal's BER is close to the traditional method,but avoids the mixed noise modeling and parameter estimating.Trianed BPSK signal reception networks'generalization ability for noise parameters is also tested in this chapter.The results show that the network has robustness for the change of noise mixing ratio and a increasing,but performance poor when ? decreasing.In Chapter 4,DL-based MSK signal demodulation networks are studied.According to the memory characteristics of MSK signal,a MSK signal direct receiving network based on LSTM is proposed.Then the proposed network is modified by taking U-net to suppress the noise before using the LSTM network to receive the pre-processed MSK signal.In addition,a DL-based branch measure is also presented for replacing the ML branch measure in the Viterbi algorithm.Next,the BER performance of three kinds of MSK signal demodulation networks is simulated and compared to Viterbi demodulation algrithom with ML branch measure based the fitting noise model.Simulation results show that Viterbi demodulation algrithom with DL-based branch measure performance best in BER.It has around 0.22dB gain compared to the ML branch measure when BER=10-3.Furthor more,trianed MSK signal reception networks'generalization ability for noise parameters is also tested.The results show that the network has robustness for the change of noise mixing ratio and ? increasing,but performance poor when ? decreasing.In Chapter 5,the work of this dissertation is summarized,and a brief description of the follow-up work is given.