Research On Intelligent CSI Feedback Techniques For 6G Networks

Massive Multiple-Input Multiple-Output(MIMO)technology is one of the key technologies in the physical layer of The Fifth Generation mobile communication system(5G),and its advantages include improved spectrum efficiency,energy The advantages include improved spectral efficiency,energy efficiency and reduced latency,which effectively support the needs of 5G multiple scenarios.It is known from information theory that accurate Channel State Information(CSI)acquisition is essential for channel capacity.In Frequency Division Duplex(FDD)mode,the terminal needs to feed CSI to the base station through the uplink because the reciprocity of the uplink channel is not obvious.It is expected that Massive MIMO technology will continue to play an important role in the Sixth Generation Mobile Communications(6G),and the increase in the number of antennas will further increase the communication overhead of CSI feedback.The number of antennas will further increase the communication overhead of CSI feedback.How to guarantee the accuracy of channel reconstruction while keeping the CSI feedback overhead as low as possible is one of the bottlenecks that need to be overcome in FDDbased large-scale MIMO systems.Compared with the traditional CSI feedback scheme,the deep learning-based CSI feedback approach provides a feasible new way to reduce the CSI feedback overhead and improve the channel reconstruction performance,which has achieved significant results.However,existing deep learning-based CSI feedback techniques still face the following challenges:1)The design of existing deep learning-based CSI feedback schemes does not fully consider the physical characteristics of the channel,resulting in low channel reconstruction accuracy and high model computational complexity.2)Existing schemes usually assume that channel estimation is idealized and do not take channel estimation errors into account in the CSI feedback process.3)The performance improvement achieved in the existing deep learning-based CSI feedback schemes depends on a large amount of training data,which faces the problems of high data collection cost and long training time.Based on the above challenges,this paper conducts a study on deep learning-based CSI feedback,and the main research work and innovations are as follows.1)To address the challenge that the physical features of the channel are not fully considered in the deep learning-based CSI feedback scheme resulting in low channel reconstruction accuracy,this paper proposes a multilayer perceptron network with gating units for CSI feedback.The gating unit in the multilayer perceptron network can learn weighted weights for multiple angular domain features,which can be used to extract the correlation of the channel in the angular domain and thus improve the accuracy of channel reconstruction.In order to make the proposed model applicable to end devices with limited computing power and storage,the proposed model is lightened to reduce the computational complexity of the model and the number of parameters,and the parameters of the lightened model are only 50K and the computational complexity is only The computational complexity is only 3M FLOPs.2)To address the problem that the channel estimation error is not considered in the existing deep learning-based CSI feedback scheme,this paper proposes a joint deep learning-based channel estimation and channel state information feedback optimization scheme,which takes both the channel estimation error and quantization error into the CSI feedback process.The joint model based on deep learning is divided into a channel estimation sub-network and a channel state information feedback subnetwork.A learnable vector quantization layer is introduced into the channel state information feedback subnetwork to effectively quantize the feedback information by learning the distribution of data to reduce the quantization error and further improve the accuracy of channel reconstruction.The simulation results show that the vector quantization approach has a 1.7%performance improvement in channel reconstruction accuracy under different compression rates compared to uniform quantization.3)The generalizability of existing deep learning-based CSI feedback schemes to different channel scenes is limited,and the overhead of recollecting a large amount of data to train a new neural network when the channel scene changes is huge.Based on the above problems,this paper investigates the small-sample-based online update of the CSI feedback scheme,which allows the CSI feedback model to update the model online based on a small amount of real harvested channel data when the channel scene changes to reduce the cost of data collection.The generative model is used to generate a large number of pseudo-samples for training the CSI feedback model based on a small number of samples to alleviate the overfitting phenomenon in the training process.The simulation results show that the proposed data enhancement approach improves the channel reconstruction accuracy by 18%relative to that without data enhancement.