An artificial intelligence-based beam domain channel estimation method

Abstract

The application discloses a kind of based on artificial intelligence's beam domain channel estimation method, comprising: 1) using compressed sampling matching pursuit CoSaMP algorithm, preliminary eliminates a part of invalid channel element, obtains low-dimensional measurement matrix;2) based on sparse bayesian learning method, using beam domain channel block sparse characteristics, the channel element of adjacent position is combined as a group of unified processing, using expectation maximization algorithm realizes the iterative update of channel parameter;3) in each iteration constraint channel block hyperparameter, prevent overfitting generation;4) according to whether the channel expectation tends to stable judgment estimation result whether convergence;5) using final channel expectation approximate channel estimation result.The application can make full use of the structured sparse characteristics of beam domain channel potential, greatly improve the estimation performance while maintaining reasonable computational complexity and pilot overhead, solve large-scale multiple-input multiple-output MIMO beam domain channel estimation problem.

Description

Technical Field

[0001] This invention belongs to the field of channel estimation technology, and in particular relates to a beam domain channel estimation method based on artificial intelligence. Background Technology

[0002] Multiple-input multiple-output (MIMO) technology, also known as multi-antenna technology, can be implemented in different ways to achieve anti-fading diversity gain and improve system capacity. However, its performance gain is greatly limited by the number of antennas configured in the base station. To address the shortcomings of traditional MIMO technology, Marzetta of Bell Labs proposed the concept of massive MIMO in 2010. In massive MIMO, the base station (BS) is configured with dozens to hundreds of antennas, fully utilizing the spatial degrees of freedom of the system to serve multiple users simultaneously using the same time-frequency resources, achieving a significant improvement in system capacity and transmission reliability. At the same time, the increased number of antennas and the increased gain significantly reduce power consumption, improving coverage and system power consumption. Massive MIMO technology reduces terminal complexity by performing all complex processing operations at the BS. However, the large number of antennas also brings extremely high computational complexity to transmission algorithms such as signal detection and channel estimation, which is detrimental to the high mobility requirements of next-generation mobile communications.

[0003] Therefore, further research and application of massive MIMO systems is an area that cannot be ignored in the next generation of mobile communication networks.

[0004] The Beam Domain Channel Model (BDCM), which models physical realities and is easy to analyze theoretically, has attracted considerable attention from researchers due to its ability to reveal new characteristics of large-scale MIMO channels. This newly proposed method constructs the channel by transforming the traditional channel model from the array domain to the beam domain. Compared to the array domain channel model, BDCM adds a beamforming preprocessor, which effectively reduces computational complexity, lowers sensitivity to array defects, and reduces the signal-to-noise ratio (SNR) resolution threshold, thus addressing the increasingly complex challenges of large-scale MIMO systems in next-generation mobile communication applications.

[0005] Due to the large aperture size of the antenna array, the massive MIMO channel is essentially non-stationary in the array domain. This means that some scattering clusters in the environment are not visible to the entire antenna array, but can only be observed by a specific area of ​​the array, thus significantly altering the channel's statistical characteristics. However, existing channel estimation algorithms designed for massive MIMO systems (BDCM) rarely consider this non-stationarity, making them ineffective for estimating massive MIMO channels. Furthermore, experimental studies show that if the number of scatterers in a wireless channel is limited, it will exhibit a sparse structure over a large bandwidth. Therefore, the channel estimation problem for massive MIMO systems can be transformed into a sparse signal recovery problem, and utilizing the inherent sparse structure of the channel for channel estimation has attracted widespread attention. In recent years, compressed sensing-based channel estimation algorithms have gradually emerged, their core idea being to utilize the channel's latent sparsity to recover the complete sparse channel with a small number of channel measurements. Typical compressed sensing algorithms include Orthogonal Matching Pursuit (OMP), Subspace Pursuit (SP), and Compressive Sampling Matching Pursuit (CoSaMP). However, all of these algorithms require prior channel information, such as signal sparsity and noise power. Furthermore, the performance of these algorithms rapidly declines as the correlation of the measurement matrix columns increases.

[0006] Sparse Bayesian Learning (SBL) is a regression and classification machine learning algorithm based on Bayesian inference. Originally proposed by Tipping around 2001, it was subsequently introduced into the fields of sparse signal recovery and compressed sensing. SBL assumes a probability distribution for samples, uses hyperparameters to establish correlations between different probability distributions, and learns iteratively through this dependency. Therefore, SBL can perform sparse signal recovery without prior information about noise and signal sparsity. However, according to research, existing SBL channel estimation methods designed for beam-domain large-scale MIMO channels are still lacking. Furthermore, most existing SBL channel algorithms rarely consider computational complexity. Even when designs exist to reduce complexity, the trade-off with estimation performance is not considered. In other words, beam-domain SBL channel estimation technology is still in its early stages. Summary of the Invention

[0007] The purpose of this invention is to provide an artificial intelligence-based beam domain channel estimation method for reconstructing sparse channels and reducing the computational complexity in the channel estimation process. This method utilizes the structured sparsity of the channel to achieve high estimation accuracy.

[0008] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0009] A beam domain channel estimation method based on artificial intelligence, the method comprising:

[0010] Step S1: Use Compressive Sampling Matching Pursuit (CoSaMP) to pre-filter the valid atoms corresponding to the non-zero channel elements in the measurement matrix A, thereby reducing the dimensionality of the measurement matrix A, and updating the measurement matrix A. Λ Iterative optimization module with received signal input;

[0011] Step S2: Based on the updated measurement matrix A from S1 Λ The beam domain sparse Bayesian learning algorithm is improved into a block sparse Bayesian learning algorithm. The block sparsity is used to improve the estimation performance. The expectation-maximum (EM) algorithm is used to calculate the beam domain channel covariance matrix, expectation and channel block hyperparameters.

[0012] Step S3: Constrain channel block hyperparameters to avoid overfitting;

[0013] Step S4: Determine whether the channel estimation result has converged based on whether the channel expectations of the two iterations are consistent. If not, return to step S2.

[0014] Step S5: Output the channel estimation results of Compressive Sampling Matching Pursuit-assisted Block Sparse Bayesian Learning (CoSaMP-BSBL) in the beam domain.

[0015] Furthermore, step S1 specifically includes:

[0016] Step S101: In the k-th iteration, search the measurement matrix A for the indices of the 2P atoms most strongly correlated with the current residual r, where Ω is the sub-index set searched each time, and P is the final number of atoms to be retained.

[0017] Ω=max{|A H r k-1 |,2P}

[0018] Among them, A H Let r denote the conjugate transpose of matrix A. k-1 This represents the residual of the (k-1)th iteration.

[0019] Step S102: Combine the updated sub-index set from this iteration with the final total index set Λ obtained from the (k-1)th iteration. k-1,t Merged into the initial total index set Λ of the k-th iteration k，i middle

[0020] Λ k，i =Λ k-1,t ∪Ω;

[0021] Where t refers to terminal and i refers to initial, the index set is updated twice in one iteration, corresponding to Λ k，i and Λ k，t .

[0022] Step S103: Based on the current index set Λ k，i Update channel estimate

[0023]

[0024] Where hB is the channel vector formed by concatenating the beam domain channel matrix along the row direction, and y is the received signal. For index set The corresponding measurement matrix composed of atoms in A.

[0025] Step S104: Based on the currently updated channel estimate The index set is re-filtered, and erroneous 2P atom indices are removed to obtain the final total index set for the k-th iteration.

[0026]

[0027] Step S105: Update the index set Λ according to the current iteration. k，t Calculate the residual r of the k-th iteration. k

[0028]

[0029] If the number of iterations has reached the set maximum, proceed to step S106; otherwise, return to step S101.

[0030] Step S106: Output the dimension-reduced measurement matrix A Λ .

[0031] Furthermore, step S2 specifically includes:

[0032] Step S201: Based on Bayes' theorem, give the beam domain channel vector h. B Posterior probability P(h) B |y), prior probability P(h) B ) and likelihood function P(y|h B The relationship between )

[0033] P(h B |y)P(y)=P(ylh B )P(h B )

[0034] Where P(y) is a normalization constant. The channel estimation problem is transformed into a maximum a posteriori (MAP) probability estimation problem. Beam-domain channel vector estimate. It can be represented as

[0035]

[0036] Here, argmax(·) represents the parameter that maximizes the function. Assume the likelihood term follows a variance of σ. 2 Gaussian distribution

[0037]

[0038] Where exp(·) denotes an exponential function with base e. By minimizing the negative log-likelihood function, we obtain...

[0039]

[0040] For h B The prior probability P(h) B Consider a generalized prior, assuming h B Each element h B，i All follow a pattern with a mean of zero and a variance of γ. i If the channel follows a Gaussian distribution, then the prior probability of the channel can be expressed as:

[0041]

[0042] Where M is the number of BS antennas, and the hyperparameter Γ = diag(γ1, γ2, ..., γ) M ), i.e., γ i The resulting diagonal matrix leads to the SBL channel estimation method.

[0043]

[0044] This step transforms the channel estimation problem into a problem of finding the optimal solution for the hyperparameter Γ.

[0045] Step S202: Based on the structured sparsity of the beam domain channel, the Block Sparse Bayesian Learning (BSBL) algorithm is used to merge channel elements at adjacent positions into a unified processing group.

[0046] Preserve the hyperparameter γ in the SBL algorithmi This is generalized to a hyperparameter in the BSBL algorithm that characterizes the features of channel elements within the entire block. If γ i If the amplitude is zero, then the amplitude of all channel elements within the i-th block is zero. To characterize the relationship between channel elements within a block, a correlation matrix B is introduced to represent the intra-block correlation. Based on the SBL channel estimation algorithm, assuming the BSBL algorithm h... B Distribution

[0047]

[0048] in B i Let represent the correlation matrix of the i-th block.

[0049] Step S203: Using the EM algorithm, the beam domain channel vector h B As latent variables, hyperparameters are implemented. Iterative updates.

[0050] E-Step:

[0051]

[0052]

[0053] M-step:

[0054]

[0055]

[0056] in and Let L and G represent the covariance matrix and expectation of hB, respectively, and let L and G represent the channel block length and the number of channel blocks, respectively. I is the identity matrix.

[0057] Furthermore, step S3 specifically includes:

[0058] By constraining the correlation matrix B i This ensures that all channel blocks have the same correlation structure, thereby avoiding overfitting. Specifically, for any i-th channel block...

[0059] B i =B

[0060] Further constrain matrix B by constructing a symmetric positive definite matrix using the Toeplitz function. Make The statement "approximates B, especially the elements on the main diagonal and secondary diagonal" means that the constructed matrix should satisfy the condition that the elements on the main diagonal and secondary diagonal are particularly close to the elements on the main diagonal and secondary diagonal of B. The following formula constructs a matrix:

[0061]

[0062] in m0 is the average of the main diagonal elements of B, m1 is the average of the secondary diagonal elements of B, and sign(·) is the sign function. The meaning is defined as follows. r is an element that constitutes the matrix and has no special meaning.

[0063] Furthermore, step S4 specifically includes:

[0064] Iterate through the E-step and M-step of the EM algorithm until h B The estimated values ​​eventually converge, and this is used here.

[0065]

[0066] Determine if the iteration has converged. Among other things, For the current iteration channel expectation, The channel expectation from the previous iteration. If the condition is not met, return to step 2. (Due to the use of...) approximate h B This value can be used to evaluate h. B Whether it converges.

[0067] Furthermore, step S5 specifically includes:

[0068] The channel estimation result of Compressive Sampling Matching Pursuit-assisted Block Sparse Bayesian Learning (CoSaMP-BSBL) can be approximated as the final channel expectation iteratively applied.

[0069]

[0070] This invention presents an artificial intelligence-based beam domain channel estimation method with the following advantages: For beam domain channel models in large-scale MIMO systems, this invention proposes using the SBL algorithm as the channel estimation method and optimizes the SBL algorithm to make it applicable to non-stationary beam domain channel models. Based on the block sparsity of beam domain channels, a two-stage CoSaMP-BSBL channel estimation algorithm is further proposed. In the initialization stage, the CoSaMP algorithm is used to filter out only a small proportion of effective atoms. In the iteration stage, the beam domain SBL algorithm is improved to the BSBL algorithm, utilizing block sparsity to improve estimation performance. The proposed method significantly improves estimation performance while maintaining low computational complexity. Furthermore, this method does not require prior channel information, has low pilot overhead, and is more applicable in real-world scenarios. This invention demonstrates excellent performance in large-scale MIMO beam domain channel estimation and can provide accurate channel state information, which is of great significance for the development of modern mobile communication systems. Attached Figure Description

[0071] Figure 1 A flowchart illustrating the beam domain channel estimation method based on artificial intelligence provided by this invention;

[0072] Figure 2 This is a sparse structure diagram of a real beam domain channel in an embodiment of the present invention;

[0073] Figure 3 This is a channel sparse structure diagram reconstructed using the CoSaMP-BSBL channel estimation algorithm in an embodiment of the present invention;

[0074] Figure 4 This is a schematic diagram illustrating the NMSE performance of the CoSaMP-BSBL channel estimation algorithm and other channel estimation algorithms as a function of SNR in an indoor hotspot scenario according to an embodiment of the present invention.

[0075] Figure 5 This is a schematic diagram illustrating the NMSE performance of the CoSaMP-BSBL channel estimation algorithm and other channel estimation algorithms as a function of SNR in a rural macrocell scenario according to an embodiment of the present invention.

[0076] Figure 6 This is a schematic diagram illustrating the NMSE performance of the CoSaMP-BSBL channel estimation algorithm and other channel estimation algorithms as a function of SNR in an urban macrocell scenario according to an embodiment of the present invention.

[0077] Figure 7 This diagram illustrates the NMSE performance of the CoSaMP-BSBL channel estimation algorithm versus other channel estimation algorithms in an urban microcell scenario, as a function of SNR, according to an embodiment of the present invention. Detailed Implementation

[0078] To better understand the purpose, structure, and function of this invention, the following detailed description of an artificial intelligence-based beam domain channel estimation method is provided in conjunction with the accompanying drawings.

[0079] like Figure 1 As shown, an artificial intelligence-based beam domain channel estimation method specifically includes the following steps:

[0080] Step S1: Use Compressive Sampling Matching Pursuit (CoSaMP) to pre-filter the valid atoms corresponding to the non-zero channel elements in the measurement matrix A, thereby reducing the dimensionality of the measurement matrix A, and updating the measurement matrix A. Λ Iterative optimization module for receiving signal input.

[0081] Specifically, in this embodiment, step S1 includes:

[0082] Step S101: In the k-th iteration, search the measurement matrix A for the indices of the 2P atoms most strongly correlated with the current residual r, where Ω is the sub-index set searched each time, and P is the final number of atoms to be retained.

[0083] Ω=max{|A H r k-1 |,2P}

[0084] Among them, A H Let r denote the conjugate transpose of matrix A. k-1 This represents the residual of the (k-1)th iteration.

[0085] Step S102: Combine the updated sub-index set from this iteration with the final total index set Λ obtained from the (k-1)th iteration. k-l，t Merged into the initial total index set Λ of the k-th iteration k，i middle

[0086] Λ k，i =Λ k-1，t ∪Ω

[0087] Step S103: Based on the current index set A k，i Update channel estimate

[0088]

[0089] Where h B Let y be the channel vector formed by concatenating the beam domain channel matrix along the row direction, and let y be the received signal. For index set The corresponding measurement matrix composed of atoms in A.

[0090] Step S104: Based on the currently updated channel estimate The index set is re-filtered, and erroneous 2P atom indices are removed to obtain the final total index set for the k-th iteration.

[0091]

[0092] The backtracking mechanism of the CoSaMP algorithm improves computational efficiency while ensuring the accuracy of effective atom selection.

[0093] Step S105: Update the index set Λ according to the current iteration. k，t Calculate the residual r of the k-th iteration. k

[0094]

[0095] If the number of iterations has reached the set maximum, proceed to step S106; otherwise, return to step S101.

[0096] Step S106: Output the dimension-reduced measurement matrix A Λ Since the CoSaMP iteration in the CoSaMP-BSBL algorithm initialization is only used to remove a portion of invalid atoms rather than to estimate the entire channel, it does not require knowledge of the channel's sparse prior. Simulation analysis compares the computational complexity of using and without initialization, where algorithm complexity is represented by runtime. Here, the total number of BS antennas is set to 4096, and P = 1024. Simulation results are shown in Table 1. It can be observed that the runtime for both cases decreases as the SNR increases. Comparing the two cases longitudinally, the initialization operation reduces the algorithm complexity by approximately 75%, demonstrating the significant effect of this design on complexity reduction.

[0097] Table 1. Impact of initialization on the complexity of the CoSaMP-BSBL algorithm.

[0098]

[0099] Step S2: Based on the measurement matrix updated in S1, use the EM algorithm to calculate the beam domain channel covariance matrix, expectation, and channel block hyperparameters.

[0100] Specifically, in this embodiment, step S2 includes:

[0101] Step S201: Based on Bayes' theorem, give the beam domain channel vector h. B Posterior probability P(h) B |y), prior probability P(h) B ) and likelihood function P(y|h B The relationship between )

[0102] P(h B|y)P(y)=P(y|h B )P(h B )

[0103] Where P(y) is a normalization constant. The channel estimation problem is transformed into a maximum a posteriori (MAP) probability estimation problem.

[0104]

[0105] Here, arg max(·) represents the parameter that maximizes the function. Assume the likelihood term follows a variance of σ. 2 Gaussian distribution

[0106]

[0107] Where exp(·) denotes an exponential function with base e. By minimizing the negative log-likelihood function, we obtain...

[0108]

[0109] For h B The prior probability P(h) B Consider a generalized prior, assuming h B Each element h B，i All follow a pattern with a mean of zero and a variance of γ. i If the channel follows a Gaussian distribution, then the prior probability of the channel can be expressed as:

[0110]

[0111] Where M is the number of BS antennas, and the hyperparameter Γ = diag(γ1, γ2, ..., γ) M ), i.e., γ i The resulting diagonal matrix leads to the SBL channel estimation method.

[0112]

[0113] This step transforms the channel estimation problem into a problem of finding the optimal solution for the hyperparameter Γ.

[0114] Step S202: Based on the structured sparsity of the beam domain channel, channel elements at adjacent positions are merged into a unified processing group to form the BSBL algorithm.

[0115] Preserve the hyperparameter γ in the SBL algorithm i This is generalized to a hyperparameter in the BSBL algorithm that characterizes the features of channel elements within the entire block. If γ iIf the amplitude is zero, then the amplitude of all channel elements within the i-th block is zero. To characterize the relationship between channel elements within a block, a correlation matrix B is introduced to represent the intra-block correlation. Based on the SBL channel estimation algorithm, assuming the BSBL algorithm h... B Distribution

[0116]

[0117] in B i Let represent the correlation matrix of the i-th block.

[0118] Step S203: Using the EM algorithm, the beam domain channel vector h B As latent variables, hyperparameters are implemented. Iterative updates.

[0119] E-Step:

[0120]

[0121]

[0122] M-step:

[0123]

[0124]

[0125] in and They represent h respectively B The covariance matrix and expectation are given by L and G, which represent the channel block length and the number of channel blocks, respectively, and I is the identity matrix.

[0126] Step S3: Constrain the channel block hyperparameters to avoid overfitting.

[0127] Specifically, in this embodiment, step S3 includes:

[0128] By constraining the correlation matrix B i This ensures that all channel blocks have the same correlation structure, thereby avoiding overfitting. Specifically, for any i-th channel block...

[0129] B i =B

[0130] Further constrain matrix B by constructing a symmetric positive definite matrix using the Toeplitz function. Make Approximates B, especially the main diagonal and secondary diagonal elements:

[0131]

[0132] in m0 is the average value of the main diagonal elements of B, m1 is the average value of the secondary diagonal elements of B, and sign(·) is the sign function.

[0133] Step S4: Determine whether the channel estimation result has converged. If it has not converged, return to step S2.

[0134] Specifically, in this embodiment, step S4 includes:

[0135] Iterate through the E-step and M-step of the EM algorithm until h B The estimated values ​​eventually converge, and this is used here.

[0136]

[0137] Determine if the iteration has converged. Among other things, For the current iteration channel expectation, The channel expectation from the previous iteration. If the condition is not met, return to step 2.

[0138] Step S5: Output the channel estimation results.

[0139] Specifically, in this embodiment, step S5 includes:

[0140] The beam-domain CoSaMP-BSBL channel estimation result can be approximated as the final channel expectation obtained through iteration, i.e.

[0141]

[0142] In this embodiment, a beam domain channel matrix with a dimension of 64×64 is randomly generated, and each channel block contains 8 adjacent channel elements. In an indoor hotspot scenario, the sparse structure of the real beam domain channel and the sparse structure of the channel reconstructed using the CoSaMP-BSBL channel estimation algorithm are shown below. Figure 2 and Figure 3 As shown, the x-axis and y-axis represent the beam indices in the vertical and horizontal directions, respectively, while the z-axis represents the true and estimated values ​​of the channel's normalized amplitude. Simulation results show that the estimation results reconstructed by the CoSaMP-BSBL algorithm can accurately recover the true channel, with a sparse structure almost identical to the true channel. It can completely recover low-power paths without introducing additional noise. Furthermore, the CoSaMP-BSBL algorithm considers the channel's structural characteristics and successfully recovers the power leakage component in the channel.

[0143] In four scenarios—indoor hotspot, rural macrocell, urban macrocell, and urban microcell—the normalized mean square error (NMSE) of the CoSaMP-BSBL channel estimation algorithm versus other channel estimation algorithms varies with SNR as follows: Figures 4-7 As shown in the figure, the x-axis represents the signal-to-noise ratio (SNR), and the y-axis represents the normalized mean square error (MSE). OMP and BOMP channel estimation algorithms require accurate prior channel information during the estimation process. However, the prior channel sparsity varies in different environments, so they cannot perform well in all scenarios. While the SBL channel estimation algorithm does not require prior channel information, it does not fully utilize the potential structured sparsity of the channel, and therefore does not achieve higher performance gains as SNR increases. In contrast, the CoSaMP-BSBL algorithm fully exploits the sparse structure of the channel, and its NMSE performance is superior to the aforementioned algorithms in both low and high SNR ranges. Furthermore, this superiority is evident in all scenarios, verifying the algorithm's universality and robustness. Therefore, the CoSaMP-BSBL algorithm significantly reduces algorithm complexity while ensuring excellent estimation performance.

[0144] In summary, the beam domain channel estimation method based on artificial intelligence established in this invention comprises two stages. The initialization stage uses the CoSaMP algorithm to pre-filter out a large number of elements with zero amplitude in the channel, reducing computational complexity. The iteration stage improves the beam domain SBL algorithm to the BSBL algorithm, utilizing block sparsity to enhance estimation performance. This invention significantly improves estimation performance while maintaining reasonable computational complexity and pilot overhead.

[0145] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A beam domain channel estimation method based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Use compressed sampling matched pursuit CoSaMP to pre-screen the measurement matrix. The effective atoms corresponding to the non-zero channel elements are used to realize the measurement matrix. Dimensionality reduction and updating the measurement matrix Iterative optimization module with received signal input; Step S2: Based on the updated measurement matrix in S1 The beam domain sparse Bayesian learning algorithm is improved into a block sparse Bayesian learning algorithm. The block sparsity is used to improve the estimation performance. The expectation-maximization EM algorithm is used to calculate the beam domain channel covariance matrix, expectation and channel block hyperparameters. Step S3: Constrain channel block hyperparameters to avoid overfitting; Step S4: Determine whether the channel estimation result has converged based on whether the channel expectations of the two iterations are consistent. If not, return to step S2. Step S5: Output beam domain compressed sampling matched pursuit auxiliary block sparse Bayesian learning CoSaMP-BSBL channel estimation results; Step S1 specifically includes the following steps: Step S101, in the In the next iteration, in the measurement matrix Search and current residual The most strongly correlated Index of atoms, For each search of a sub-index set, To select the final number of atoms to retain; in, Representation matrix The conjugate transpose of . Indicates the first The residual of the next iteration; Step S102: Combine the updated sub-index set from this iteration with the first... The final total index set obtained in the next iteration Merge into the The initial total index set for the next iteration middle Step S103: Based on the current index set Update channel estimate in The channel vector is formed by concatenating the beam domain channel matrix along the row direction. In order to receive signals, For index set corresponding Measurement matrix composed of atoms; Step S104: Based on the currently updated channel estimate Re-filter the index set and delete any erroneous entries. The index of the atom is obtained to get the _ ... The final total index set of the next iteration Step S105: Update the index set according to this iteration. Calculate the first Second iteration residual If the number of iterations has reached the set maximum number, proceed to step S106; otherwise, return to step S101. Step S106: Output the dimensionality-reduced measurement matrix ; Step S2 specifically includes the following steps: Step S201: Based on Bayes' theorem, give the beam domain channel vector. Posterior probability Prior probability and likelihood function Relationship: in To standardize constants; the channel estimation problem is transformed into a maximum a posteriori (MAP) estimation problem, and the beam domain channel vector estimate is obtained. Represented as in, This represents the parameter that maximizes the function; the likelihood term is set to follow a variance of . Gaussian distribution in Represented by natural constant An exponential function with base 1; by minimizing the negative log-likelihood function, we obtain for Prior probability Consider a generalized prior, set Each element h B,i All follow a mean of zero and a variance of . If the channel follows a Gaussian distribution, then the prior probability of the channel is expressed as: in The number of base station (BS) antennas, hyperparameters ,Right now The diagonal matrix is ​​constructed; the sparse Bayesian learning (SBL) channel estimation method is obtained. This step transforms the channel estimation problem into hyperparameters. Optimal solution problem; Step S202: Based on the structured sparsity of the beam domain channel, the Block Sparse Bayes Learning (BSBL) algorithm is used to merge channel elements at adjacent positions into a unified processing group. Preserving Hyperparameters in the Sparse Bayesian Learning (SBL) Algorithm This is generalized to a hyperparameter in the BSBL algorithm that characterizes the features of channel elements within the entire block; if If it equals zero, then the first The amplitude of all channel elements within a block is zero; to characterize the relationship between channel elements within a block, a correlation matrix is ​​introduced. Indicates intra-block correlation; based on the Sparse Bayesian Learning (SBL) channel estimation algorithm, the Block Sparse Bayesian Learning (BSBL) algorithm is set. Distribution in , Indicates the first The correlation matrix of each block; Step S203: Using the EM algorithm, the beam domain channel vector is... As a latent variable, hyperparameters are implemented. Iterative updates; E-Step: M-step: in represent The covariance matrix, represent Expectations Represents the channel block length. Represents the number of channel blocks. It is an identity matrix.

2. The beam domain channel estimation method based on artificial intelligence as described in claim 1, characterized in that, Step S3 specifically includes the following steps: By constraining the correlation matrix This ensures that all channel blocks have the same correlation structure, i.e., for any given channel block... Channel blocks constraint matrix Use the Toeplitz function to construct a symmetric positive definite matrix. , making Approximately Main diagonal elements and secondary diagonal elements: in , for The average of the main diagonal elements for The average value of the secondary diagonal elements. It is a symbolic function.

3. The beam domain channel estimation method based on artificial intelligence as described in claim 2, characterized in that, Step S4 specifically includes the following steps: Iterate through the E-step and M-step of the EM algorithm until... The estimated values ​​eventually converge, and this is used here. Determine whether the iteration has converged; where, For the current iteration channel expectation, The channel expectation from the previous iteration is given; if the condition is not met, return to step 2.

4. The beam domain channel estimation method based on artificial intelligence as described in claim 3, characterized in that, The specific steps S5 are as follows Includes the following steps: The beam-domain CoSaMP-BSBL channel estimation result is approximately equal to the final channel expectation obtained through iteration, i.e. 。

Images