Research On Data-Model Dual-Drive Advanced Transceive

As the commercial deployment of the 5th Generation(5G)mobile communication sys-tems has entered a critical stage,the advantages and disadvantages of the main core enabling technologies have become increasingly apparent.Taking into account the explosive growth of mobile data traffic and the demand for business diversification,the global wireless R&D community has begun the layout and construction of Beyond 5G(B5G)and even the 6th gen-eration(6G)mobile communication systems.Among the candidate technologies for B5G and6G,intelligent communication technology which employs artificial intelligence(AI)algorithms to improve wireless transmission efficiency has become the focus of attention in the academic and industry circles.Although there have been some successful cases in the application of AI algorithms among different layers of wireless networks,the research on theory and key tech-nologies of advanced transceiver design is still in its infancy.This thesis takes the data-model dual-drive as the core and conducts in-depth research on the theory and key technologies of advanced transceiver design for intelligent communication.Firstly,we summarize the basic theory of advanced transceiver design based on deep learn-ing(DL).The concepts of DL and the existing general architectures of advanced transceiver based on DL are briefly introduced.Aiming at the problems of difficulty in designing net-works,high computational overhead,lack of interpretability,and difficulty in guaranteeing performance of the black-box transceiver architecture,a data-model dual-driven framework that integrates theoretical models and data-driven DL algorithms is introduced,which lays the foundation for the research on advanced transceiver design in subsequent chapters.Secondly,we design a basic Turbo decoder based on data-model dual-drive.Aiming at the problems of high computational complexity of the traditional iterative Turbo decoding algo-rithms and difficulty in guaranteeing the reliability of the existing data-driven Turbo decoders,a Turbo decoding neural network architecture,called Turbo Net,is proposed,which is based on data-model dual-drive,and then a loss function is carefully designed to prevent the Turbo Net's gradient vanishing issue.The simulation results demonstrate that the proposed Turbo Net de-coder performs approximately 14 times faster than the traditional iterative decoders when they obtain a similar bit error rate(BER)performance,which indicates that the Turbo Net can effec-tively reduce the decoding latency.Subsequently,we propose a structured pruning scheme for the aforementioned Turbo Net decoder.Aiming at the problem that Turbo Net introduces too many redundant weights and increases the computational complexity,a pruned Turbo Net architecture,called Turbo Net+,is proposed,which is based on the structured pruning approach.Compared with the Turbo Net,the Turbo Net+can significantly reduce the number of weights without sacrificing error correction capability,thereby reducing computation and storage overhead.We further present a simple training strategy to address the overfitting issue,which enables efficient training of the proposed Turbo Net+.The simulation results demonstrate that the proposed Turbo Net+decoder can obtain a signal-to-noise ratio gain by nearly 1.5 d B compared with the data-driven decoder when the target BER equals 10^-7,and perform approximately 14 times faster than the aforementioned data-driven decoder.The over-the-air test demonstrates the Turbo Net's strong learning ability and great robustness to various scenarios.Finally,we design a constant envelope(CE)precoder based on data-model dual-drive for the massive multi-user multiple-input multiple-output(MIMO)systems.Aiming at the problem of high computational complexity of the existing iterative algorithms,a CE precoding neural network architecture,called CEPNet,is proposed,which is based on data-model dual-drive.The simulation results demonstrate the CEPNet's significant superiority in multi-user interference suppression capability,average achievable rate,BER,and computational overhead compared with the traditional iterative precoders.Furthermore,the CEPNet shows strong robustness to the channel estimation error,the channel model mismatch,and the MIMO system configuration.On this basis,the training strategy of joint CEPNet and Turbo Net+is studied.The simulation results demonstrate that the CEPNet and the Turbo Net+can perform approximately 13 and 14times faster than corresponding iterative algorithms when the systems obtain a similar BER performance,which effectively reducing computational complexity.