Optimization Of Signal Constellation And Filtering Based On End-to-End Learning

To meet the ever-increasing data capacity demand,digital communication systems are developing towards high-speed and largecapacity direction.Therefore,how to design communication systems to achieve higher channel capacity,spectrum efficiency,and transmission accuracy has become an important research topic.In this context,constellation shaping technology and pulse filtering design have attracted widespread attention.On the other hand,with its powerful data processing capability,deep learning has been widely applied in physical layer communication research,where end-to-end learning communication models provide a new perspective for communication system design.The optimization method of end-to-end learning is accordance with the joint design of constellation mapping and inverse mapping,shaping filtering and matching filtering in communication systems.It also solves the problem of suboptimal system performance caused by the modular optimization method of traditional model design.Therefore,this thesis focuses on the signal constellation and filtering optimization design of end-to-end learning communication system,and the main research contents are summarized as follows:1.To achieve optimal performance under different channel models and system data rates,an end-to-end learning constellation shaping scheme based on minimizing the symbol error rate is proposed.An autoencoder is used to construct the end-to-end learning communication system.Geometric constellation shaping network and probabilistic constellation shaping network are designed at the transmitting end,and a differentiable model for symbol sampling is established.The loss function and model training method of the system are redesigned.The results show that the proposed model can generate constellation signals with the preset rate and power,and realize better symbol error performance than the MaxwellBoltzmann distribution in additive white Gaussian noise channels and Rayleigh fading channels.2.In order to improve the channel capacity of the system,approach the Shannon limit,and achieve higher transmission rates,an end-to-end learning constellation shaping scheme based on maximizing the information rate is proposed.It is theoretically proven that the channel capacity is highest when the channel input is Gaussian distribution in additive white Gaussian noise channels.However,solving the optimal input distribution for non-Gaussian channels is a complex problem.Therefore,this thesis utilizes mutual information neural estimator to estimate the mutual information between channel input and output,and redesigns the loss function to optimize the constellation shaping network.The results show that compared with uniform distribution orthogonal amplitude modulation scheme,both the geometric and probabilistic shaping algorithms of proposed model can obtain capacity gains.3.A filter optimization method for baseband transmission model based on time delay neural network is proposed.The transmitting filter,receiving filter and sampling layer of the system are constructed by the time delay neural network.The thesis qualitatively analyzed the adaptability of the receiving filter network to different pulse waveforms and transmission channels,as well as the optimization results of the transmitting filter network for different band-limited channels.Experimental results demonstrated that the proposed model can achieve a balance between spectral efficiency and inter-symbol interference,exhibiting better system performance.