Research On Channel Estimation And Feedback Approaches Based On Deep Learning In Time-varying 3D MIMO Systems

With the widespread adoption of the Fourth Generation(4G)and the Fifth Generation(5G)worldwide in recent years,the volume of data services in communications has increased dramatically,making it difficult for traditional technologies to meet the high load and low latency requirements of communications systems.In this context,the 3D MIMO(Three Dimension Multiple Input Multiple Output)technology has emerged as one of the key technologies for 5G.Compared to the traditional large-scale 2D MIMO,this technology introduces a pitch angle in the signal propagation angle,allowing the horizontal and vertical dimensions to be fully utilised,greatly increasing the system capacity,but also increasing the difficulty of channel estimation and feedback.At the same time,high-speed mobile environments are a common scenario in communications,where the channel is characterised by fast time-varying and non-stationary conditions,and this introduces randomness into the channel,which in turn makes channel estimation and feedback more difficult.In order to solve the problems in this scenario,we focus on the rapid development of Deep Learning(DL)techniques,which,due to their powerful computational and learning capabilities,provide new ideas for solving a range of problems in time-varying large-scale 3D MIMO communication systems.In this paper,the main focus is to consider how to use deep learning methods to accomplish the channel estimation and channel feedback problems in time-varying large-scale 3D MIMO communication systems,with the following two main research points.Channel estimation in a high-speed mobile environment is a hot problem in 3D MIMO millimeter-wave(mm Wave)systems.In this environment,the channel is characterised by fast time-varying and non-stationary conditions.The time-domain correlation coefficient is a time-varying parameter,which makes it difficult for conventional frequency-domain channel estimation methods to capture the channel variation over time and achieve the desired channel estimation performance.To effectively solve the channel estimation problem of time-varying large-scale 3D MIMO systems,this paper proposes a deep learning-based method to track the time-varying large-scale 3D MIMO mm Wave channel variations and perform channel estimation.In the design of the deep neural network,considering that the large-scale 3D MIMO mm Wave channel coefficients have sparse characteristics in the spatial and frequency domains,and the fast time-varying leads to its strong time dependence in the Doppler delay domain,this paper constructs the channel matrix from three dimensions,including the spatial dimension,the frequency dimension and the Doppler delay dimension,and uses them as the input of the deep neural network.The proposed deep learning network consists of a Convolutional Neural Network(CNN)and a Recurrent Neural Network(RNN),where the CNN is used to extract features in the spatial frequency domain of the channel matrix and the RNN is used to extract time-domain correlated features of the channel response to complete interpolation of data points,i.e.channel estimation.In addition,a maximum pooling network is used to reduce the training parameters of the network.To make full use of the channel information in the training samples,standard high-speed channel data is used to train the network offline so that the proposed network can effectively learn the fast time-varying and non-stationary characteristics of the channel.Simulation results show that the proposed method is able to track the changes in channel characteristics at fast scenarios of 50km/h and 300km/h and achieves better performance compared to conventional methods.In the context of time-varying large-scale 3D MIMO system and FDD communication mode,after the User Equipment(UE)has completed the channel estimation,it needs to send the estimated Channel State Information(CSI)of the downlink channel back to the BS(Base Station)side for subsequent The pre-coding or beamforming requirements of the BS side are then used.However,the time-varying large-scale 3D MIMO systems with large antennas and time-varying environments add additional time information to the CSI,making it difficult to perform channel feedback efficiently and accurately.In order to effectively reduce the feedback overhead and complete the feedback more accurately,this paper proposes a deep learning-based network framework:BLA-Csi Net,which enables the compression and recovery of the CSI matrix at both the transmit and receive sides of a large-scale 3D MIMO system.In the network structure and design,an encoder is first designed to fully extract the feature information in the spatial and frequency dimensions of the CSI matrix by exploiting the sparse characteristics of CSI in the spatial and frequency domains at the receiver side.The encoder consists of a Three-Dimensional Convolutional Neural Network(3D CNN)and a Bidirectional Long Short-Term Memory(Bi LSTM)network.The attention-based Bi LSTM network can fully extract the feature information of the CSI matrix in the time dimension and compress it into a low-dimensional vector,which is sent to the base station via the uplink feedback channel;after receiving the low-dimensional vector at the base station,the base station uses the decoder of BLA-Csi Net to recover the CSI matrix.Simulation results show that the proposed method is able to complete the channel feedback with high performance in a high-speed mobile environment and achieve better NMSE performance and similarity performance compared to traditional methods and other deep learning networks.