Large-scale MIMO joint beam forming and channel estimation method based on deep learning

Abstract

The invention belongs to the technical field of information and communication, and relates to a large-scale MIMO joint beam forming and channel estimation method based on deep learning. In order to solve the problem of single-user MIMO channel estimation, the invention provides a deep learning framework to improve the performance of channel estimation. The framework is composed of two main modules: a receiving and transmitting end beam generator based on a convolutional neural network (CNN), and a channel estimator of a UAMP-SBL algorithm based on parameter learning. The frame adopts a multi-iteration optimization process, a beam generator adaptively designs a beam of a current round based on virtual channel information estimated last time, and a channel estimator is responsible for updating a real-time channel estimation result. Simulation experiments show that the method provided by the invention is superior to a random beam-based baseline method in performance.

Description

Technical Field

[0001] The present invention belongs to the field of information and communication technology, and specifically relates to a large-scale MIMO joint beamforming and channel estimation method based on deep learning. Background Art

[0002] With the rapid development of wireless communications, massive multiple-input multiple-output (MIMO) technology has become one of the core technologies in 5G and future 6G communication systems. By deploying a large number of antennas between base stations and user equipment, massive MIMO can significantly improve system capacity, spectrum efficiency, and energy efficiency. However, as the size of antennas increases, hardware costs and system complexity also increase. To this end, the system usually adopts analog beamforming technology to form a narrow directional beam to concentrate the signal energy in the target direction, thereby improving the quality and reliability of the communication link. The optimal design of beamforming depends on channel state information (CSI), and obtaining high-precision channel estimation has become one of the key links to improve the performance of massive MIMO systems.

[0003] In recent years, the application of deep learning in the field of wireless communications has made significant progress, especially in beamforming and channel estimation. For example, the paper "H. Hojatian, J. Nadal, J.-F. Frigon, and F. Leduc-Primeau, 'Unsupervised Deep Learning for Massive MIMO Hybrid Beamforming, 'IEEE Trans. Wireless Commun., vol. 20, no. 11, pp. 7086-7099, Nov. 2021" proposed an unsupervised deep learning method based on received signal strength indication (RSSI) to design hybrid beamforming in massive MIMO systems. In terms of channel estimation, the paper "H. He, C.-K. Wen, S. Jin, and GY Li, 'Deep Learning-Based Channel Estimation for Beamspace mmWave Massive MIMO Systems,' IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 852-855, Oct. 2018" proposed a channel estimation method based on deep learning, using a learning denoising approximate message passing (LDAMP) network, which significantly improved the channel estimation accuracy.

[0004] Although existing research has made significant progress in beamforming and channel estimation, most methods still adopt independent optimization strategies and fail to fully utilize the intrinsic connection between beamforming and channel estimation. In large-scale MIMO systems, beamforming and channel estimation are closely related. The beamforming effect directly affects the accuracy of channel estimation, while accurate channel estimation provides more accurate prior information for beamforming. In addition, independent optimization strategies may lead to inconsistent parameter adjustments during training, resulting in performance bottlenecks. Therefore, jointly optimizing beamforming and channel estimation tasks and using deep learning frameworks for collaborative training have become the key to improving the overall performance of large-scale MIMO systems. Summary of the invention

[0005] The present invention proposes a joint beamforming optimization and channel estimation framework based on deep learning, aiming to improve the channel estimation performance in single-user MIMO systems. The framework consists of two main modules: a transceiver beamformer based on a convolutional neural network (CNN) and a channel estimator based on the UAMP-SBL algorithm with parameter learning. The framework is optimized through multiple iterations, in which the beamformer adaptively designs the beam of the current round according to the virtual channel information estimated last time, while the channel estimator is responsible for updating the channel estimation results in real time. Simulation experimental results show that compared with the baseline scheme based on random beams, the method proposed in the present invention shows significant advantages in channel estimation performance.

[0006] The technical method of the present invention is:

[0007] A large-scale MIMO joint beamforming and channel estimation method based on deep learning is defined in the system where the transmitting and receiving ends have N r and N t The array uses analog beamforming technology to form M antennas. r (<N r ) receive beams and M t (<N t ) transmit beams. The method comprises:

[0008] Use with G r (>N r ) grids of discrete Fourier transform (DFT) receiving end angle domain dictionary A r and G t (>N t ) grids of DFT origin angle domain dictionary A t For MIMO channels For sparse representation, define is a virtual channel of sparse representation;

[0009] Setting up a CNN-based multi-beamformer This module is based on the virtual channel of the previous iteration To generate the current round of receiving end beamformer and transmit beamformer The specific implementation method is:

[0010] For any iteration t, Take the absolute value Input to the generating function and mapped to a beamformer represented by real numbers and right and Each column of is vector 2-norm normalized and then converted into a complex beamformer and Finally, the two complex beamformers are orthogonalized, that is, W (t) and F (t) They are and

[0011] Setting up the channel estimator of the UAMP-SBL algorithm based on parameter learning Used to estimate the channel of the current round and its sparse representation The specific implementation method is:

[0012] The receiving end is based on W (t) Observe and observe the signal As a channel estimator The input of is the observation matrix, To vectorize the sparse representation channel, The vectorized mean is 0 and the variance is σ 2 Gaussian white noise;

[0013] The channel estimator first sets up a CNN The input is the vectorized virtual channel calculated in the previous round and variables (initialized to 1), used to learn the parameter γ (t) :

[0014]

[0015] Among them will and The input data is stacked into 2D data with 2 channels; then, the variables are calculated. for Calculate the variable p (t) for Where ⊙ is the element-by-element Hadamard product, s (t-1) Initialize to 0; calculate variables for Calculate variable s (t) for Calculated variables for Calculate the variable q (t) for Calculated variables for Compute the recovered vectored virtual channel for

[0016]

[0017] Finally, Restore to matrix form and its original estimated channel

[0018] The constructed CNN-based multi-beamformer and the channel estimator of the UAMP-SBL algorithm based on parameter learning Training is performed, specifically: generating random arrival angles AoA and departure angles AoD of multiple paths to construct the MIMO channel H as the true value; for the initial iteration round t = 0, the virtual channel Initialize all 0 elements and input the initialization data into the CNN-based multi-beam generator to generate the beam matrix W (t) and F (t) , and then the channel estimator of the UAMP-SBL algorithm based on parameter learning calculates the data used for the next round of training Until τ rounds of training are completed, the loss function of the end-to-end framework is the normalized mean square error (NMSE) of the channel estimation:

[0019]

[0020] in is the channel estimated in the last training round, ||·|| F is the matrix 2 norm;

[0021] In practice, a CNN-based multi-beamformer is used and the channel estimator of the UAMP-SBL algorithm based on parameter learning Perform channel estimation, specifically by establishing a joint iterative optimization model:

[0022]

[0023] By solving the above model, the transmit and receive beamformer and estimated channel for each iteration are obtained.

[0024] In the above system, consider a quasi-static single-user MIMO channel Then it can be expressed as

[0025]

[0026] Where: N c is the number of clusters, N p is the number of paths within each cluster, a i，j is the independent and identically distributed Gaussian fading coefficient vector, θ i，j is the AoA of the jth path in the ith cluster, is the AoD of the jth path in the ith cluster, is the steering vector of the ULA with N antennas, that is

[0027]

[0028] Where: λ and d are wavelength and antenna spacing respectively; in order to utilize the sparsity of the channel in the angular domain, the channel can be expressed as follows

[0029]

[0030] in: is the virtual channel of sparse representation in the angular domain, and They are receiving and sending DFT dictionaries respectively;

[0031] The transmitting and receiving ends can form M r and M t simulated beams, the received signal It can be expressed as

[0032] Y=W[HF H x+N]

[0033] in: and are the row-orthogonal receiving and transmitting beamforming matrices, x is the user's transmitted signal, is independent and identically distributed Gaussian white noise; further, let x = 1, Y can be vectorized as

[0034]

[0035] in: is the equivalent observation matrix, To vectorize the sparse representation channel, is the vectorized noise.

[0036] The beneficial effects proposed by the present invention are: using a CNN-based multi-beamformer to generate a transmit and receive beamforming matrix, reducing the complexity of online beam optimization; using a channel estimator of the UAMP-SBL algorithm based on parameter learning to achieve channel estimation and improve estimation robustness; in addition, simulation experiments show that: the channel estimation performance implemented based on this framework is better than other baseline schemes. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Iterative framework for joint optimization algorithms;

[0038] Figure 2 Schematic diagram of the neural network structure of the beamformer and channel estimator;

[0039] Figure 3 Training and validation loss function curves for end-to-end neural networks;

[0040] Figure 4 is the energy distribution curve of the channel and the receiving beam in the angular domain dictionary;

[0041] Figure 5 Simulation curves of the baseline algorithm using random beam and the proposed scheme under different SNRs;

[0042] Figure 6 Simulation curves of the baseline algorithm using random beams and the proposed scheme under different numbers of channel clusters; DETAILED DESCRIPTION

[0043] The specific implementation of the present invention is further described in detail below in conjunction with the drawings and examples.

[0044] Figure 1 This is the implementation block diagram of the joint optimization algorithm, including the generation of transmit and receive beams, the introduction of observation signals, and the implementation of the channel estimation algorithm.

[0045] Figure 2 The architecture of two neural network modules: the first module includes a 5×5 convolution layer, an adaptive average pooling layer (reducing the feature map to 16×16), and then flattened to 256-dimensional features through a Flatten layer, and then processed through two fully connected layers, with a final output size of M r ×2N r (receiver real represents the beamforming matrix) or M t ×2N t (the real representation of the beamforming matrix at the transmitting end); the second module consists of multiple 5×5 convolutional layers with an input size of G r ×G r, the number of channels is 2, the middle layer generates a feature map with 16 channels and the same size, and finally outputs G through a convolutional layer r ×G r The single channel parameters (learning parameters γ (t) ).

[0046] A specific implementation method of the present invention based on the above algorithm is given below, and the parameters of the specific method are set as follows:

[0047] Set the number of MIMO antennas to N r =N t =32, the number of beams at the transmitting and receiving ends is M r =M t =16, the antenna spacing is half a wavelength Set the number of clusters to N c =3, the number of paths within the cluster is N p =10, set AoA and AoD, the value range is The angle expansion within the cluster is 5°; the grid size of the transmit and receive DFT dictionary is set to G r =G t =48; set the maximum number of training rounds to τ = 5 and the data set size to 10 4 , where the ratio of the training set to the validation set is 9:1; the SNR range is set to 0~20dB.

[0048] According to the above parameter settings, the specific steps of the simulation are as follows:

[0049] Set the number of transmitting and receiving antennas in the system to N r =32 and N t = 32 and arranged in ULA. The array uses analog beamforming technology to form M r = 16 receive beams and M t =16 transmit beams, the method comprising:

[0050] Use with G r = DFT receiving end angle domain dictionary A of 48 grids r and G t = DFT transmitting angle domain dictionary A of 48 grids t For MIMO channels For sparse representation, define is a virtual channel of sparse representation;

[0051] Setting up a CNN-based multi-beamformer Based on the virtual channel of the previous iteration To generate the current round of receiving end beamformer and transmit beamformer The specific implementation method is:

[0052] For any iteration t, Take the absolute value Input to the generating function and mapped to a beamformer represented by real numbers and right and Each column of is vector 2-norm normalized and then converted into a complex beamformer and Finally, the two complex beamformers are orthogonalized, that is, W (t) and F (t) for and

[0053] Setting up the channel estimator of the UAMP-SBL algorithm based on parameter learning Used to estimate the channel of the current round and its sparse representation The specific implementation method is:

[0054] The receiving end is based on W (t) Observe and observe the signal As a channel estimator The input of is the observation matrix, To vectorize the sparse representation channel, The vectorized mean is 0 and the variance is σ 2 Gaussian white noise;

[0055] The channel estimator first sets up a CNN The input is the vectorized virtual channel calculated in the previous round and variables (initialized to 1), used to learn the parameter γ (t) :

[0056]

[0057] Among them will and The input data is stacked into 2D data with 2 channels; then, the variables are calculated. for Calculate the variable p (t) for Where ⊙ is the element-by-element Hadamard product, s (t-1) Initialize to 0; calculate variables for Calculate variable s (t) for Calculated variables for Calculate the variable q (t) for Calculated variables for Compute the recovered vectored virtual channel for

[0058]

[0059] Finally, Restore to matrix form and its original estimated channel

[0060] The constructed CNN-based multi-beamformer and the channel estimator of the UAMP-SBL algorithm based on parameter learning Training is performed, specifically: generating random arrival angles AoA and departure angles AoD of multiple paths to construct the MIMO channel H as the true value; for the initial iteration round t = 0, the virtual channel Initialize all 0 elements and input the initialization data into the CNN-based multi-beam generator to generate the beam matrix W (t) and F (t) , and then the channel estimator of the UAMP-SBL algorithm based on parameter learning calculates the data used for the next round of training Until τ = 5 rounds of training are completed, the loss function of the end-to-end framework is the NMSE of the channel estimation:

[0061]

[0062] in is the channel estimated in the last training round, ||·|| F is the matrix 2 norm;

[0063] In practice, a CNN-based multi-beamformer is used and the channel estimator of the UAMP-SBL algorithm based on parameter learning Channel estimation is performed, specifically establishing a joint iterative optimization model as follows:

[0064]

[0065] By solving the above model, the transmit and receive beamformer and estimated channel for each iteration are obtained.

[0066] Figure 3The graph shows the change of the loss value of the training set and the validation set with the iteration round in the end-to-end training process. The horizontal axis is the training round (Epoch), and the vertical axis is the loss value. It can be seen that the loss values ​​of the training set and the validation set both drop rapidly in the early stage, indicating that the model can quickly learn effective features. After about 50 rounds, the loss value gradually stabilizes and finally converges to a lower level. This shows that the end-to-end design of the model can effectively optimize the loss function and ensure the performance of the joint optimization task of multi-beam forming and channel estimation.

[0067] Figure 4 The channel and receive beams are shown in the angular domain dictionary A. r The energy distribution curve on the grid index is shown in Figure 1. The horizontal axis is the grid index and the vertical axis is the normalized energy. As can be seen from the figure, the energy distribution curve of the beamforming at the receiving end is aligned with the energy distribution curve of the channel at multiple grid indexes, indicating that the beamforming optimization result at the receiving end can effectively align the AoA, thereby verifying the effectiveness of the beam optimization module.

[0068] Figure 5 These are simulation curves for channel estimation using different algorithms. The horizontal axis is SNR, and the vertical axis is NMSE of channel estimation. The prefix "Proposed" corresponds to the joint optimization algorithm proposed in this invention, the name "OMP" corresponds to the algorithm in "JATropp and ACGilbert, Signal recovery from random measurements via orthogonal matching pursuit,'IEEE Trans. Inf. Theory, vol. 53, pp. 4655–4666, Dec. 2007.", and the name "UAMP-SBL" corresponds to the algorithm in "M. Luo, Q. Guo, M. Jin, YCEldar, D. Huang, and X. Meng, 'Unitary Approximate Message Passing for Sparse Bayesian Learning,'IEEE Trans. Signal Process., vol. 69, no. 1, pp. 123-135, Jan. 2021.", where these two baseline algorithms use random beamforming. As shown in the figure, the channel estimation performance of the method proposed in the present invention has obvious gain over other algorithms within the SNR range.

[0069] Figure 6 The estimated performance simulation curves under different cluster numbers, where the SNR is set to 5dB. As shown in the figure, as the number of clusters increases (the total number of paths increases), the performance of the method proposed in the present invention is much better than other baseline algorithms.

Claims

1. A massive MIMO joint beamforming and channel estimation method based on deep learning. It is defined that the system has N transmitters and receivers respectively. r and N t antennas are arranged in a uniform linear array, and the array uses analog beamforming technology to form M r <N r receive beams and M t <N t transmit beams; characterized in that, The method comprises: Use with G r >N r The DFT receiving angle domain dictionary A of the grid r and G t >N t The DFT origin angle domain dictionary A of the grid t For MIMO channels For sparse representation, define is a virtual channel of sparse representation; Setting up a multi-beamformer based on a convolutional neural network (CNN) Based on the virtual channel of the previous iteration To generate the current round of receiving end beamformer and transmit beamformer The specific implementation method is: For any iteration t ,Will Take the absolute value Input to the generating function and mapped to a beamformer represented by real numbers and right and Each column of is vector 2-norm normalized and then converted into a complex beamformer and Finally, the two complex beamformers are orthogonalized, that is, W (t) and F (t) They are and Setting up the channel estimator of the UAMP-SBL algorithm based on parameter learning Used to estimate the channel of the current round and its sparse representation The specific implementation method is: The receiving end is based on W (t) Observe and observe the signal As a channel estimator The input of is the observation matrix, To vectorize the sparse representation channel, The vectorized mean is 0 and the variance is σ 2 Gaussian white noise; The channel estimator first sets up a CNN The input is the vectorized virtual channel calculated in the previous round and variables For learning the parameter γ (t) : Among them will and The input data is stacked into 2D data with 2 channels; then, the variables are calculated. for Calculate the variable p (t) for Where ⊙ is the element-by-element Hadamard product, s (t-1) Initialize to 0; calculate variables for Calculate variable s (t) for Calculated variables for Calculated variables for Calculated variables for Compute the recovered vectored virtual channel for Finally, Restore to matrix form and its original estimated channel The constructed CNN-based multi-beamformer and the channel estimator of the UAMP-SBL algorithm based on parameter learning Training is performed, specifically: generating random arrival angles AoA and departure angles AoD of multiple paths to construct the MIMO channel H as the true value; for the initial iteration round t = 0, the virtual channel Initialize all 0 elements and input the initialization data into the CNN-based multi-beam generator to generate the beam matrix W (t) and F (t) , and then the channel estimator of the UAMP-SBL algorithm based on parameter learning calculates the data used for the next round of training Until τ The training is completed after rounds of training, where the loss function of the end-to-end framework is the normalized mean square error (NMSE) of the channel estimation: in is the channel estimated in the last training round, ||·|| F is the matrix 2 norm; In practice, a CNN-based multi-beamformer is used and the channel estimator of the UAMP-SBL algorithm based on parameter learning Perform channel estimation, specifically by establishing a joint iterative optimization model: By solving the above model, the transmit and receive beamformer and estimated channel for each iteration are obtained.

Images