Design And Detection Of High-order Modulation Based On Deep Learning

The sixth generation(6G)mobile communication systems will achieve the goal of “Everything Connected” with ultra-low latency,ultra-high speeds and high reliability,providing servives for the development of novel wireless applications such as extended reality(XR),smart healthcare and next-generation Internet of Things.In recent years,Artificial Intelligence(AI)is considered the most promising solution for designing high-performance 6G systems,and its powerful potential has been fully demonstrated in the wireless physical layer.Therefore,this paper focuses on some of the serious challenges in the wireless physical layer,such as nonlinear distortion from power amplifiers,constel-lation modulation design under Additive White Gaussian Noise(AWGN)channel and constellation rearrangement(Co Re)design in Hybrid automatic retransmission request(Hybrid ARQ,HARQ),etc,a series of studies have been carried out using deep learning techniques in AI.The research work and results of the paper include the following:Firstly,in view of the nonlinear distortion problem caused by the memoryless power amplifiers,this paper introduces the conventional post-distortion detection algorithm for the Rapp and Saleh mod-els.Considering the high implementation complexity of conventional post-distortion detection algo-rithm,this paper proposes a bit-level demodulator network(BLDnet)based on deep neural network,which provides the possibility for the application of deep learning in amplifier nonlinear scenarios.BLDnet uses the Sigmoid function to obtain the posterior probability distribution of each bit in the re-ceived signal,thus realizing the function of the traditional demodulator.The simulation results show that the soft value obtained by BLDnet can be used in the decoder of the forward error correction code,and when the modulated signal is 1024 QAM,compared with the conventional post-distortion detection algorithm,BLDnet can achieve a lower complexity on the Rapp and Saleh models,which can significantly improve the detection performance of the system.Then,an enhanced bit-wise Autoencoder(e Bit-AE)network based on deep neural network is proposed to solve the problem of constellation modulation design in AWGN channel.The e Bit-AE network essentially uses the neural network to implement the modulation,channel,and demodulation modules in the traditional communication system,and obtains the geometric constellation under the AWGN channel by optimizing the e Bit-AE network.Compared with the symbol-wise autoencoder network which uses multi-class cross entropy as the loss function,the e Bit-AE network can design constellations suitable for bit-interleaved coded modulation systmes.Simulation results show that when the modulation orders are 16,64 and 256 respectively,the geometric constellation designed by the proposed ebit-AE network has significant geometric shaping gain compared with APSK con-stellation,non-uniform constellation proposed in ATSC3.0 protocol of the third generation of digital broadcasting and television system in North America,and QAM constellation in 5G standard.Finally,this paper proposes a Core-AE scheme based on deep neural network to solve the Co Re scheme design problem in HARQ with limited transmission times.When the number of transmissions in HARQ is limited to two,considering the conventional Co Re solution in this case is difficult to guarantee the reliability of the symbols on each bit to achieve consistently,the Co Re-AE scheme uses two autoencoder networks to simulate a data transmission in the process.And this solution can not only optimize the constellation mapping method during multiple transmissions,but also optimize the geometric position of the constellation during multiple transmissions.The simulation results show that when the number of transmissions in HARQ is limited to two,the error rate performance of Co Re-AE scheme is better than that of conventional Co Re scheme with modulation order of 16 and 64,and the performance gain obtained by this scheme will increase with the increase of modulation order.In addition,we can know through analysis that the performance gain of the Co Re scheme comes from partial constellation rearrangement gain and geometric shaping gain.