Research On Space-Time-Frequency Multi-Dimensional Channel Prediction Method For 6G

With the development of research on the sixth generation(6G)mobile communication’ techniques,extremely large-scale multiple-input multiple-output(MIMO),reconfigurable intelligent surface(RIS),terahertz communication,and other techniques have become the key enabling techniques.The networks of 6G will expand towards physical resource dimensions such as the time,frequency,and space domain of wireless channels.Faced with the expansion of channel dimensions,channel predictions that utilize the acquired channel state information(CSI)to predict the target CSI have come into the researchers’ view.However,the mapping between channels is often nonlinear,making it difficult to analyze mathematically.With the development of artificial intelligence(AI),AI techniques have solved previously intractable problems in wireless channel modeling,estimation,resource management,and other aspects through its powerful nonlinear processing capabilities.In 2016,the research team proposed that AI-based channel predictions can be studied from the perspective of channel propagation characteristics.This dissertation starts by studying multidimensional channel propagation characteristics,which deeply explores the principles and methods of AI-assisted multi-dimensional channel predictions.The details are as follows:(1)Given the problem that existing channel predictions only focus on the characteristics of the AI algorithm and lack in-depth fusion analysis of the multi-dimensional channel characteristics and AI algorithm,this dissertation divides channel prediction into three dimensions:time,frequency,and space domain.Firstly,the physical characteristics of electromagnetic wave propagation affecting channel prediction in each domain are analyzed,such as spatial consistency property,partial reciprocity,high-and low-frequency angular distribution correlation,spatial nonstationarity,etc.Then,how AI models are combined with channel propagation characteristics is explained.By giving three typical channel prediction cases from uplink(UL)to downlink(DL)channel prediction,high-and low-frequency dynamic beam prediction,and spatial cross-domain channel prediction,it is verified that AI-based channel prediction methods have powerful nonlinear mapping capabilities,intelligent adaptive optimization capabilities,and potential cross-domain experience reference capabilities,respectively.It is worth mentioning that this dissertation proposes a Transformer-based channel prediction scheme and finds the existence of crossdomain channel prediction gains.Simulation results show that the experience of the AI model cross different domains can be mutually reinforcing.(2)Aiming at the problem that the DL-CSI is quickly outdated and requires frequent estimation in high-speed scenarios,and the high-dimensional channel characteristics are difficult to measure accurately,this dissertation proposes a channel prediction method suitable for time/frequency division duplexing systems by conditional GAN.In the proposed method,taking the UL-CSI of the previous time slots as a conditional constraint,the predicted DL-CSI and the real DL-CSI are classified by a discriminator network.The discriminator network will also judge whether the channel mapping between the UL-CSI and DL-CSI is established to solve the problem that the multi-dimensional channel characteristics are difficult to measure uniformly.The generator network takes the UL-CSI of previous time slots as input and predicts the DL-CSI of future time slots.The GAN of conventional unsupervised learning realizes the model training in the way of supervised learning.Considering the characteristics of real channels,a channel prediction error indicator is raised to determine whether the generator network reaches the optimal prediction state.Simulation results show that the proposed method can perform satisfactory prediction at 300 km/h.(3)Aiming at the problem of the huge beam sweep overhead in mmWave massive MIMO systems,this dissertation proposes a dynamic beam prediction method based on deep reinforcement learning for future multi-band coexistence communication systems.A learning motivation for dynamic beam selection is raised by analyzing the distribution law of the mmWave and sub-6 GHz channels’ power angle spectrum(PAS):to learn the index offset of the mmWave channel’s optimal beam on the sub-6 GHz channel’s spatial spectrum.Multiple beam peak indexes are defined on the sub-6 GHz channel’s spatial spectrum as the mmWave beam sweep subset’s initial locations.A new action space is presented to determine the size of the beam sweep subset at each location.The proposed method can predict different mmWave beam sweep sets with dynamic sizes for different wireless channels.According to the designed reward function,the dynamic trade-off between beam selection quality and beam sweep overheads is achieved.The superior performance of the proposed method is demonstrated compared with the existing approaches by simulations and channel measurement data,which shows the advantage of deeply combining channel propagation characteristics with AI algorithms.(4)Aiming at the problem that it is difficult for RIS-assisted communications to obtain high-dimensional CSI when the transmitter has a massive antenna array,this dissertation proposes a space-domain channel prediction approach and a joint beam prediction method,respectively.To achieve a better channel/beam prediction performance,this dissertation develops the adaptive antenna selection(AAS)strategy on both the transmitter and RIS.The AAS strategy chooses different optimal antenna subsets under different wireless channels for channel estimation.This dissertation extends the probability sampling theory to realize the joint optimization of two-dimensional(transmitter and RIS)AAS strategy and channel/beam prediction network by backpropagation.The problem of designing channel AAS strategy to assist channel/beam prediction under unknown CSI is solved.Simulation results show that the channel prediction scheme’s achievable rate(AR)is very close to the upper bound of AR at a high spatial compression ratio.The beam prediction method suits scenarios with high energy consumption constraints.It can reach a satisfactory AR with less partial CSI.