Research On GFDM Signal Detection Technology Based On Deep Learning

In the research of modern wireless communication technologies,multi-carrier modulation technologies occupy a very important position.Generalized Frequency Division Multiplexing(GFDM)is a block structure-based non-orthogonal multi-carrier modulation scheme.It has attracted much attention due to its low peak-to-average power ratio,low out-of-band radiation,low cyclic prefix overhead,flexible structure configuration and other advantages.In wireless communication,the received signal will be affected by the propagation characteristics of the wireless fading channel.In order to recover the transmitted data symbols accurately,it is essential to obtain accurate Channel State Information(CSI).However,the cyclic shift pulse shaping process of GFDM will destroy the orthogonality between subcarriers,resulting in inherent Inter-Subcarrier Interference(ICI)and Inter-Subsymbol Interference(ISI),which seriously affect the pilot-based channel estimation performance of GFDM system.In addition,for broadband GFDM communication system,the mobility of users will bring Doppler effect,which makes the channel have time-frequency doubly selectivity.The channel estimation and signal detection of GFDM system are facing great challenges.Deep learning has been widely used in many fields such as computer vision and natural language processing due to its powerful feature expression ability and the modeling ability for complex tasks.Researchers in the field of wireless communication try to combine it with various aspects of wireless communication,and then generate intelligent communication for future communication systems,which is considered as the mainstream development direction of the B5 G.In the field of intelligent communication,channel estimation and signal detection assisted by deep learning are important research components.Since the deep learning-based communication system does not require prior channel statistical features,the deep neural network model can learn the characteristics of received signals through iteration training and complete the signal detection task.It shows significant performance advantages in reducing algorithm complexity and processing heterogeneous data.Taking advantage of these,this thesis carries out research on deep learning-based GFDM signal detection technology.This thesis first introduces the basic theory of GFDM technology and deep learning.On this basis,the deep learning technology is introduced to study the GFDM signal detection technology under the condition of multi-path Rayleigh channel and doubly selective channel respectively.Under the scenario of multi-path Rayleigh channel,the traditional GFDM detection algorithm is analyzed,and a GFDM deep receiver system is designed.The demodulation module,channel estimation module and symbol detection module of the traditional algorithm are considered jointly,so the end-to-end detection function of the received signal is realized through the fully connected neural network.The simulation system of GFDM deep receiver is designed and implemented.The simulation results demonstrate that the BER performance of the designed GFDM deep receiver is better than the MMSE-based traditional receiver,and the advantage is magnified when the number of pilot symbols is reduced.Under the scenario of doubly selective channel,a processing method of the computer vision is introduced.The channel estimation problem is modeled as a super-resolution reconstruction problem,and a GFDM signal detection framework aided by residual convolutional network is proposed.Firstly,the channel gain at the pilot position is estimated by the least squares criterion.This estimated channel containing interference and noise is formed into a low-resolution image and the residual convolutional network is used to reconstruct a high-resolution image of channel time-frequency gain.After zero forcing equilibrium and demapping,the information bits can be recovered.The simulation results demonstrate that under the condition of doubly selective channel,the GFDM signal detection system can obtain the error performance close to the minimum mean square error channel estimation by using the proposed algorithm.And the residual convolutional network-based channel estimation algorithm has strong generalization ability for different Doppler frequency shift channels.