Design Of MIMO-OFDM End-to-end Receiver Based On Deep Learning

With the advent of the era of intelligent technology,mobile communications have become an indispensable part of our lives.Since the beginning of the 4G mobile communication system,MIMO and OFDM have become the key technologies of wireless communication and the technical basis for future intelligent communication.For the architecture design of the communication system,the traditional solution is to build the entire communication system based on hierarchy and modularity,that is,divide the entire system into multi-layer cascaded modules,where each module is responsible for a specific function,and an expert knowledge base is usually required.Carry out modeling and design.However,in the future intelligent communication,the scale of the system will become very large and complex.The traditional modular solution may be difficult to achieve better results due to its high complexity and low flexibility.At the same time,the development of high-efficiency computing hardware such as GPU and the explosive growth of data volume have laid a solid foundation for the application of deep learning and other technologies in communication systems.Through deep learning training,it is possible to learn the hidden laws between data transmission and reception in different wireless environments from massive communication data,thereby realizing data demodulation and recovery.This subject is mainly based on the data-driven method of deep learning,and conducts research on the end-to-end receiver design in the MIMO-OFDM system.This thesis first focuses on the single-input single-output OFDM system,introduces the three deep learning model structures MLP,Res Net and Transformer that have achieved significant results in the field of text images at this stage,and applies them to the receiving end of the OFDM system through model adaptation.The end-to-end receiver directly trains the deep learning model,which can directly restore the information bits from the radio frequency input data.At the same time,under different system parameter configurations(such as with or without cyclic prefix,with or without peak clipping,number of pilots,etc.),a large number of simulation comparisons were made between the deep learning scheme proposed in this thesis and traditional methods.Finally,the simulation results show that in different scenarios,the deep learning model structure introduced in this paper can improve the bit error rate performance of 1% ～ 14% compared with the traditional LS and MMSE algorithms in the OFDM end-to-end receiver design.This thesis further conducts deep learning training on the end-to-end receiver of the more complex MIMO-OFDM system.Based on the model optimization results under the single-input single-output OFDM system,optimization ideas are proposed for the above three deep learning network model structures.First of all The MLP structure introduces the idea of multi-size modeling to obtain more spatial dimension information;the structure is more complex The Res Net model adopts the idea of implicit channel estimation to speed up the convergence speed and improve the performance of the model;Finally,the convolution idea is introduced to the Transformer model,and the Conv Transformer model is proposed to further distinguish Noisy signal and real signal.At the same time,this paper conducts a lot of simulation comparisons on the optimized deep learning scheme,other existing deep learning schemes and traditional methods.The simulation results show that the improved deep learning model structure proposed in this paper can effectively improve the detection performance and speed up the running speed.Compared with the original unimproved model,its bit error rate performance is improved by 34% ～ 50%.Compared with the MMSE-SIC algorithm commonly used in the industry,the bit error rate performance of the proposed scheme in different scenarios is improved by 1% ～ 15%,and the calculation speed is doubled.