Research On Detection Method Based On Deep Learning

There are many high requirements in peak rate,delay,spectrum efficiency,reliability and user rate for future wireless network applications.The fifth generation of mobile communication system(5G)has proposed the large-scale antenna arrays,millimeter wave technology,and ultra-density networks in order to meet these requirements.However,there are some inherent limitations of traditional communication theory in realizing big data transmission and ultra-high-speed communication in complex scenes.For example,channel modeling is difficult in complex scenarios,and module optimization can not guarantee global optimization.In recent years,the rapid development of artificial intelligence technology has provided a new opportunity to break the traditional communication system concept and solve the challenges faced by traditional communication system.In this thesis,the detection methods and detection performance based on deep learning techniques are studied based on the challenges encountered by traditional detection techniques.Firstly,a neural network detection method for the output signals of nonlinear system with memory is proposed.The system principle is described.The structure of neural network for detection is given.The selection method of training parameters and the training principle are discussed.The MATLAB neural network toolbox was used to build the neural network,and the detection performance of neural network was simulated under different nonlinear and memory parameters,as well as different modulation schemes of BPSK,QPSK and 16 QAM.The results show that the neural network detection achieves significant performance gains compared to the hard decision method.In addition,the generalization performance of neural network detection is simulated.It is verified that the network can still be used to detect the output signals of nonlinear channels with complete nonlinearity after the training of the training set composed by partial channel nonlinearity.Subsequently,the problem of large-scale multiple-input multiple-output(MIMO)signal detection based on deep learning is studied.The optimal MIMO detection problem proves to be an Non-deterministic Polynomial-time hard(NP-hard)problem.The computational complexity of the detection algorithm based on maximum likelihood(ML)and Maximum A Posteriori criteria(MAP)becomes very large as the decision variables increase.The detection performance of DetNet designed by the ML algorithm using the deep unfolding principle is simulated.The simulation scenario includes two scenarios in which the channel is fixed and the channel parameters are changed,and the detection performance under different antennas and the number of users is given.Then,for the problem that DetNet has reduced efficiency under variable channel,high modulation order and a large number of users and antennas,an improved scheme is proposed,and the performance of improved network is simulated.Finally,for high-speed mobile Orthogonal Frequency-Division Multiplexing(OFDM)communication systems and OFDM communication systems with large carrier phase noise in millimeter band,inter-carrier interference(ICI)is caused by doppler shift and phase noise.The structure and detection principle of OFDM detection network overcoming ICI are studied inspired by the MIMO deep detection network.The detection performance under different conditions of different doppler shift and different modulation schemes is simulated.Aiming at the problem of performance floor at high SNR,the improved schemes of channel preprocessing and Zero-Forcing(ZF)detection are proposed.The simulation proves that the improved scheme can improve performance.To further validate detection performance based on deep learning,a simulation environment with channel coding is built,and the performance of the proposed detection network is verified in the coded modulation environment.Then,for the OFDM detection problem of multiple subcarriers overcoming ICI,a deep sliding window detection network is proposed,and its principle,training method and simulation results of performance are introduced.