Priori-enhanced block sparse channel estimation method for millimeter wave massive mimo system

Abstract

The application discloses a priori enhancement block sparse channel estimation method for a millimeter wave massive MIMO system, and comprises the following steps: system initialization, construction of an angle domain dictionary matrix, uniform division of an observation angle range, calculation of a uniform linear array normalized steering vector corresponding to each grid angle and arrangement in columns; collection of a noisy observation matrix of a snapshot, calculation of a sample covariance matrix, extraction of a noise subspace through eigenvalue decomposition, calculation of a MUSIC pseudo-spectrum and normalization, and aggregation of an atomic level pseudo-spectrum into a block level prior weight vector; execution of a priori enhancement block orthogonal matching pursuit iteration, and cyclic execution of the following sub-steps until a stop condition is met. The method fully utilizes the characteristics that a millimeter wave channel has obvious sparsity in an angle domain, is suitable for a communication environment with high user density and large signal-to-noise ratio fluctuation, and realizes a channel estimation method with high detection performance and low calculation complexity under a complex communication environment.

Description

Technical Field

[0001] This invention relates to the field of wireless communication channel estimation technology, specifically to a priori enhancement block sparse channel estimation method for millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) systems, which is particularly suitable for passive random access (URA) communication scenarios with high user density and large fluctuations in signal-to-noise ratio (SNR). Background Technology

[0002] With the rapid development of fifth-generation and future sixth-generation mobile communication technologies, millimeter-wave massive MIMO systems, with their abundant spectrum resources and high spatial degrees of freedom, have become a core technology direction for meeting the demands of ultra-high-speed and ultra-large-capacity communication. In massive machine-type communication scenarios, a large number of low-power devices access the base station via random access, forming the so-called passive random access model. Its core task is to jointly complete the detection of active users and the acquisition of channel state information.

[0003] Millimeter-wave channels exhibit significant spatial sparsity: in the angular domain, only a few multipath components carry the main energy, and multiple multipath angles from the same user are highly concentrated within a continuous interval, forming a "block sparsity" structure. This characteristic provides a theoretical basis for efficient channel estimation using compressed sensing (CS) theory. Orthogonal Matching Pursuit (OMP) is the most representative greedy sparse recovery algorithm, but it uses single-atom selection granularity and fails to effectively utilize the block sparsity characteristic, thus limiting its performance in user detection and coefficient estimation.

[0004] Block Orthogonal Matching Pursuit (BOMP) uses predefined angle blocks as the basic selection unit and outperforms OMP. However, it still employs a purely residual-driven greedy strategy in each iteration, failing to utilize prior channel knowledge. Under low SNR conditions, BOMP is highly prone to misidentifying noise energy as valid signal components, leading to support set contamination and causing a peak phenomenon where the Normalized Mean Square Error (NMSE) abnormally increases as SNR decreases.

[0005] Multiple Signal Classification (MUSIC) algorithms can provide angular domain pseudospectral data by analyzing the noise subspace structure of the received signal, offering valuable prior information for block sparse recovery. However, existing studies introducing MUSIC priors into block sparse recovery generally lack robust design: the priors themselves are unreliable at low SNR, and unreasonable fusion methods can exacerbate performance degradation; furthermore, the lack of an effective iteration stopping mechanism causes the algorithm to continue iterating after identifying the real signal block, introducing noise atoms and further deteriorating the estimation quality.

[0006] In summary, existing BOMP-type algorithms and their prior enhancement variants all exhibit significant robustness deficiencies in low SNR scenarios, necessitating an improved algorithm that can maintain stable performance over a wide SNR range. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide a priori enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems, which solves the problems of abnormal peaks in NMSE, low active user detection rate and numerical instability in existing algorithms under low SNR, and achieves high-precision and robust channel estimation over a wide SNR range.

[0008] Technical solution: The method described in this invention includes the following steps:

[0009] A priori enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems, characterized by the following steps:

[0010] S1: System initialization. Set base station configuration. A uniform linear array antenna is used. The system has K potential users. Each user corresponds to a continuous block of width B in the angle dictionary. The total number of columns in the dictionary is... Set an upper limit for the actual number of active users. Multipath number P, multi-shot number L, maximum number of iterations ;

[0011] S2: Construct the angle domain dictionary matrix A, uniformly dividing the observation angle range [-90°, 90°] into... Calculate the angle of each grid point. The corresponding uniform linear array (ULA) normalizes the guide vectors and arranges them column-wise;

[0012] In step S2, the k-th column of the dictionary matrix A is the guiding vector. Construct according to formula (1):

[0013]

[0014] Mode middle, For the number of antennas, The first is obtained by uniformly discretizing within the observation angle range. Each grid angle, where j is the imaginary unit; the channel vector satisfies the block sparse observation model. Where y is the single snapshot observation vector and x is the block sparse coefficient vector. It is additive white Gaussian noise, CN is a complex Gaussian distribution, and I represents the identity matrix.

[0015] S3: Collect the noisy observation matrix Y of L snapshots, calculate the sample covariance matrix R, perform eigenvalue decomposition to extract the noise subspace, calculate the MUSIC pseudospectrum and normalize it, and aggregate the atomic-level pseudospectrum into a block-level prior weight vector w.

[0016] In step S3, the MUSIC pseudospectrum is calculated and normalized. The atomic-level pseudospectrum is then aggregated into a block-level prior weight vector w. The method is as follows:

[0017] (1) Construct the sample covariance matrix, where H represents the conjugate transpose of the matrix:

[0018]

[0019] (2) After eigenvalue decomposition, take the last ( - The noise subspace matrix is ​​composed of 10 eigenvectors. Construct the projection matrix:

[0020]

[0021] (3) Calculate and normalize the MUSIC pseudospectrum:

[0022]

[0023] (4) Aggregate into block-level prior weights ,in The first is obtained by uniformly discretizing within the observation angle range. Each grid angle This is the normalized MUSIC pseudospectral value corresponding to this angle; For users The set of atomic indices corresponding to the angle block. ,in The number of atoms contained in each angular block. With equation (1) Essentially, they all represent grid angles obtained by uniformly discretizing within the observation angle range, but different subscripts are used to represent them only to avoid confusion of symbols.

[0024] S4: Perform prior enhancement block orthogonal matching pursuit iterations, repeatedly executing the following sub-steps until the stopping condition is met:

[0025] S4.1: Estimate the noise variance by dividing the power of the current residual vector r by the number of antennas, and let the regularization parameter... , For regularization parameters, For noise variance, A fixed numerical stability factor;

[0026] Adaptive regularization parameters in step S4.1 According to the formula calculate:

[0027]

[0028]

[0029] In the formula, Here, r is the noise variance estimate, and r is the current residual vector. For the number of antennas, Numerical stability factor (typical value) The LMMSE ridge regression coefficient update formula in step S4.3 is as follows:

[0030]

[0031] Mode middle, For the current support set, To support the sub-dictionary of the set, I is the identity matrix and y is the single snapshot observation vector.

[0032] S4.2: Calculate the comprehensive score for all unselected blocks k. ,in The residual projection energy equal to the fusion weight factor of block k The product of For prior weights normalized to [0,1], As the prior fusion ratio parameter, select The largest block k* is a candidate block;

[0033] Step S4.2 Block Comprehensive Scoring According to the formula calculate:

[0034]

[0035] Mode middle, This is the submatrix corresponding to block k in the dictionary matrix. To normalize the prior weights, This is the prior fusion ratio parameter; recommended value. ;when It degenerates into a standard BOMP when It depends entirely on prior knowledge.

[0036] S4.3: Construct the experimental support set, calculate the experimental coefficients using LMMSE ridge regression, and calculate the residual decay ratio. ;like Greater than the preset threshold If the previous valid support set and coefficients are restored, the iteration is terminated; otherwise, candidate block k* is formally accepted; if the stopping condition is met (e.g., reaching the threshold), the iteration is terminated. (or triggers a rollback), then the loop will exit;

[0037] In step S4.3, the residual attenuation ratio ρ is calculated according to the formula... calculate:

[0038]

[0039] Mode middle, This is the residual vector before adding the candidate block. The experimental residual vector after adding candidate blocks; Recommended threshold. ;when At that time, rollback to the previous effective support set With coefficient vector And terminate the iteration. S4.4: Update the support set, coefficient vector, and residual, and add k* to the selected block set; if the number of selected blocks reaches Then it will be terminated early;

[0040] S5: Output channel estimation vector , The final channel estimation vector. Beneficial effects: (1) Adaptive LMMSE regularization (step S4.1): The noise variance is adaptively estimated with residual power, and the regularization intensity is dynamically adjusted with SNR, which fundamentally eliminates the coefficient estimation explosion problem caused by matrix ill-conditioning under low SNR; under high SNR, the regularization approaches zero and does not affect the estimation accuracy.

[0041] (2) Residual decay early stop and rollback mechanism (step S4.3): Real-time monitoring of the residual energy ratio before and after the candidate block is added, identifying noise blocks and immediately terminating the iteration, effectively blocking the deterioration path of the noise pollution support set, is the core mechanism for eliminating low SNR NMSE peaks.

[0042] (3) Weighted normalized prior fusion (step S4.2): Normalization preprocessing prevents extreme prior weights from overwhelming residual information; The parameters achieve a dynamic balance between prior and data-driven approaches. When the SNR is high, the prior is reliable and guides and accelerates convergence, while when the SNR is low, the data items provide a safety net and reduce the misselection rate.

[0043] By combining the three improvements, this invention outperforms the OMP and BOMP benchmark algorithms in key metrics such as active user detection rate, NMSE, and spectral efficiency over a wide SNR range (0~30 dB). Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention;

[0045] Figure 2 This is a priori diagram of MUSIC;

[0046] Figure 3 This is a schematic diagram of step S4 iteration. Detailed Implementation

[0047] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0048] like Figure 1 As shown, a priori enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems is characterized by the following steps:

[0049] S1: System initialization. Set base station configuration. A uniform linear array antenna is used. The system has K potential users. Each user corresponds to a continuous block of width B in the angle dictionary. The total number of columns in the dictionary is... Set an upper limit for the actual number of active users. Multipath number P, multi-shot number L, maximum number of iterations ;

[0050] S2: Construct the angle domain dictionary matrix A, uniformly dividing the observation angle range [-90°, 90°] into... Calculate the angle of each grid point. The corresponding uniform linear array (ULA) normalizes the guide vectors and arranges them column-wise;

[0051] In step S2, the k-th column of the dictionary matrix A is the guiding vector. Construct according to formula (1):

[0052]

[0053] Mode middle, For the number of antennas, The first is obtained by uniformly discretizing within the observation angle range. Each grid angle, where j is the imaginary unit; the channel vector satisfies the block sparse observation model. Where y is the single snapshot observation vector and x is the block sparse coefficient vector. It is additive white Gaussian noise, CN is a complex Gaussian distribution, and I represents the identity matrix.

[0054] S3: Collect the noisy observation matrix Y of L snapshots, calculate the sample covariance matrix R, perform eigenvalue decomposition to extract the noise subspace, calculate the MUSIC pseudospectrum and normalize it, and aggregate the atomic-level pseudospectrum into a block-level prior weight vector w.

[0055] In step S3, such as Figure 2 As shown, the method for calculating and normalizing the MUSIC pseudospectrum, and then aggregating the atomic-level pseudospectrum into a block-level prior weight vector w, is as follows:

[0056] (1) Construct the sample covariance matrix, where H represents the conjugate transpose of the matrix:

[0057]

[0058] (2) After eigenvalue decomposition, take the last ( - The noise subspace matrix is ​​composed of 10 eigenvectors. Construct the projection matrix:

[0059]

[0060] (3) Calculate and normalize the MUSIC pseudospectrum:

[0061]

[0062] (4) Aggregate into block-level prior weights ,in The first is obtained by uniformly discretizing within the observation angle range. Each grid angle This is the normalized MUSIC pseudospectral value corresponding to this angle; For users The set of atomic indices corresponding to the angle block. ,in The number of atoms contained in each angular block. With equation (1) Essentially, they all represent grid angles obtained by uniformly discretizing within the observation angle range, but different subscripts are used to represent them only to avoid confusion of symbols.

[0063] S4: As Figure 3 As shown, perform prior enhancement block orthogonal matching pursuit iteration, repeatedly executing the following sub-steps until the stopping condition is met:

[0064] S4.1: Estimate the noise variance by dividing the power of the current residual vector r by the number of antennas, and let the regularization parameter... , For regularization parameters, For noise variance, A fixed numerical stability factor;

[0065] Adaptive regularization parameters in step S4.1 According to the formula calculate:

[0066]

[0067]

[0068] In the formula, Here, r is the noise variance estimate, and r is the current residual vector. For the number of antennas, Numerical stability factor (typical value) The LMMSE ridge regression coefficient update formula in step S4.3 is as follows:

[0069]

[0070] Mode middle, For the current support set, To support the sub-dictionary of the set, I is the identity matrix and y is the single snapshot observation vector.

[0071] S4.2: Calculate the comprehensive score for all unselected blocks k. ,in The residual projection energy equal to the fusion weight factor of block k The product of For prior weights normalized to [0,1], As the prior fusion ratio parameter, select The largest block k* is a candidate block;

[0072] Step S4.2 Block Comprehensive Scoring According to the formula calculate:

[0073]

[0074] Mode middle, This is the submatrix corresponding to block k in the dictionary matrix. To normalize the prior weights, This is the prior fusion ratio parameter; recommended value. ;when It degenerates into a standard BOMP when It depends entirely on prior knowledge.

[0075] S4.3: Construct the experimental support set, calculate the experimental coefficients using LMMSE ridge regression, and calculate the residual decay ratio. ;like Greater than the preset threshold If the previous valid support set and coefficients are restored, the iteration is terminated; otherwise, candidate block k* is formally accepted; if the stopping condition is met (e.g., reaching the threshold), the iteration is terminated. (or triggers a rollback), then the loop will exit;

[0076] In step S4.3, the residual attenuation ratio ρ is calculated according to the formula... calculate:

[0077]

[0078] Mode middle, This is the residual vector before adding the candidate block. The experimental residual vector after adding candidate blocks; Recommended threshold. ;when At that time, rollback to the previous effective support set With coefficient vector And terminate the iteration. S4.4: Update the support set, coefficient vector, and residual, and add k* to the selected block set; if the number of selected blocks reaches Then it will be terminated early;

[0079] S5: Output channel estimation vector , This is the final channel estimation vector. Example:

[0080] This invention is based on Implementation instructions for a typical configuration:

[0081] System Scenarios

[0082] Base station equipped A uniform linear array (ULA) of antennas, the system has K potential users, and K antennas per time slot.a ( (Number) users can be actively connected at the same time.

[0083] Channel representation

[0084] Uplink channel vector of active user k Using a clustered geometric multipath model, it can be represented as:

[0085]

[0086] in Let be the complex gain of the p-th path, where P is the multipath number for each active user. Angle of arrival (AoA). Normalized ULA guide vector:

[0087]

[0088] Received signal and observation model

[0089] The superimposed signal received by the base station is:

[0090]

[0091] in For the guide vector dictionary matrix, Let n be a block sparse coefficient vector, and n be additive white Gaussian noise. This represents noise power.

[0092] Signal-to-noise ratio definition

[0093] The system signal-to-noise ratio (SNR) is defined as the ratio of signal power to noise power.

[0094]

[0095] E[.] represents the calculated average power, corresponding to the noise variance. .

[0096] dictionary matrix construction

[0097] angle range Uniformly discretized There are grid points, and the i-th column of the dictionary matrix is:

[0098]

[0099] S1: System initialization. Set base station configuration. A uniform linear array antenna is used. The system has K potential users. Each user corresponds to a continuous block of width B in the angle dictionary. The total number of columns in the dictionary is... Set an upper limit for the actual number of active users. Multipath number P, multi-shot number L, maximum number of iterations ;

[0100] S2: Construct the angle domain dictionary matrix A, uniformly dividing the observation angle range [-90°, 90°] into... Calculate the angle of each grid point. The corresponding uniform linear array (ULA) normalizes the guide vectors and arranges them column-wise;

[0101] S3: Collect the noisy observation matrix Y of L snapshots, calculate the sample covariance matrix R, perform eigenvalue decomposition to extract the noise subspace, calculate the MUSIC pseudospectrum and normalize it, and aggregate the atomic-level pseudospectrum into a block-level prior weight vector w.

[0102] S4: Perform prior enhancement block orthogonal matching pursuit iterations, repeatedly executing the following sub-steps until the stopping condition is met:

[0103] S4.1: Estimate the noise variance by dividing the power of the current residual vector r by the number of antennas, and let the regularization parameter... , For regularization parameters, For noise variance, A fixed numerical stability factor;

[0104] S4.2: Calculate the comprehensive score for all unselected blocks k. ,in The residual projection energy equal to the fusion weight factor of block k The product of For prior weights normalized to [0,1], As the prior fusion ratio parameter, select The largest block k* is a candidate block;

[0105] S4.3: Construct the experimental support set, calculate the experimental coefficients using LMMSE ridge regression, and calculate the residual decay ratio. ;like Greater than the preset threshold If the previous valid support set and coefficients are restored, the iteration is terminated; otherwise, candidate block k* is formally accepted; if the stopping condition is met (e.g., reaching the threshold), the iteration is terminated. (or triggers a rollback), then the loop will exit;

[0106] S4.4: Update the support set, coefficient vector, and residuals, and add k* to the selected block set; if the number of selected blocks reaches... Then it will be terminated early;

[0107] S5: Output channel estimation vector , This is the final channel estimation vector.

[0108] Across the entire SNR range, PE-BOMP outperforms both benchmark algorithms in active user detection rate and spectral efficiency. At SNR ≥ 10 dB, PE-BOMP's NMSE is reduced by approximately 3–8 dB compared to BOMP, fully validating the effectiveness and synergistic gain of the three improvement mechanisms. At SNR = 15 dB, In the scanning experiments with a value of 0~50, PE-BOMP still showed a relatively high detection rate of active users. The small inflection point on the left and right sides, followed by an increase, is because the guiding role of the prior weights can still effectively distinguish noise from signals during large-scale access, demonstrating the continuous support capability of this method for system capacity under high load scenarios.

Claims

1. A priori enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems, characterized in that, Includes the following steps: S1: System initialization, set base station configuration A uniform linear array antenna is used. The system has K potential users. Each user corresponds to a continuous block of width B in the angle dictionary. The total number of columns in the dictionary is... ; Set an upper limit for the number of actual active users. Multipath number P, multi-shot number L, maximum number of iterations ; S2: Construct the angle domain dictionary matrix A, uniformly dividing the observation angle range [-90°, 90°] into... Calculate the angle of each grid point. The corresponding uniform linear array normalizes the guide vectors and arranges them in columns; S3: Collect the noisy observation matrix Y of L snapshots, calculate the sample covariance matrix R, perform eigenvalue decomposition to extract the noise subspace, calculate the MUSIC pseudospectrum and normalize it, and aggregate the atomic-level pseudospectrum into a block-level prior weight vector w. S4: Perform prior enhancement block orthogonal matching pursuit iterations, repeatedly executing the following sub-steps until the stopping condition is met: S4.1: Estimate the noise variance by dividing the power of the current residual vector r by the number of antennas, and let the regularization parameter... , For regularization parameters, For noise variance, A fixed numerical stability factor; S4.2: Calculate the comprehensive score for all unselected blocks k. ,in The residual projection energy equal to the fusion weight factor of block k The product of For prior weights normalized to [0,1], As the prior fusion ratio parameter, select The largest block k* is a candidate block; S4.3: Construct the experimental support set, calculate the experimental coefficients using LMMSE ridge regression, and calculate the residual decay ratio. , These are the new and old residual vectors, respectively; if Greater than the preset threshold If the previous valid support set and coefficients are restored, the iteration is terminated; otherwise, candidate block k* is formally accepted; if the stopping condition is met, the iteration is complete. Or trigger a rollback, If the maximum number of active users is reached, the loop will exit. S4.4: Update the support set, coefficient vector, and residuals, and add k* to the selected block set; if the number of selected blocks reaches... Then it will be terminated early; S5: Output channel estimation vector , This is the final channel estimation vector.

2. The prior enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems according to claim 1, characterized in that, In step S2, the k-th column of the dictionary matrix A is the guiding vector. Construct according to formula (1): Mode middle, For the number of antennas, The first is obtained by uniformly discretizing within the observation angle range. Each grid angle, where j is the imaginary unit; the channel vector satisfies the block sparse observation model. Where y is the single snapshot observation vector and x is the block sparse coefficient vector. It is additive white Gaussian noise, CN is a complex Gaussian distribution, and I represents the identity matrix.

3. The prior enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems according to claim 1, characterized in that, In step S3, the MUSIC pseudospectrum is calculated and normalized. The atomic-level pseudospectrum is then aggregated into a block-level prior weight vector w. The method is as follows: (1) Construct the sample covariance matrix, where H represents the conjugate transpose of the matrix: (2) After eigenvalue decomposition, take the last ( - The noise subspace matrix is ​​composed of 10 eigenvectors. Construct the projection matrix: (3) Calculate and normalize the MUSIC pseudospectrum: (4) Aggregate into block-level prior weights ,in The first is obtained by uniformly discretizing within the observation angle range. Each grid angle, This is the normalized MUSIC pseudospectral value corresponding to this angle; For users The set of atomic indices corresponding to the angle block. ,in The number of atoms contained in each angular block.

4. The prior enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems according to claim 1, characterized in that, Adaptive regularization parameters in step S4.1 According to the formula calculate: In the formula, Here, r is the noise variance estimate, and r is the current residual vector. For the number of antennas, It is a numerical stability factor; The LMMSE ridge regression coefficient update formula in step S4.3 is as follows: Mode middle, For the current support set, To support the sub-dictionary of the set, I is the identity matrix and y is the single snapshot observation vector.

5. The prior enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems according to claim 4, characterized in that, Step S4.2 Block Comprehensive Scoring According to the formula calculate: Mode middle, This is the submatrix corresponding to block k in the dictionary matrix. To normalize the prior weights, For the prior fusion ratio parameter; when It degenerates into a standard BOMP when It depends entirely on prior knowledge.

6. The prior enhancement block sparse channel estimation method for millimeter-wave massive MIMO systems according to claim 5, characterized in that, In step S4.3, the residual attenuation ratio ρ is calculated according to the formula... calculate: Mode middle, This is the residual vector before adding the candidate block. Let be the experimental residual vector after adding the candidate block; when At that time, rollback to the previous effective support set With coefficient vector And terminate the iteration.

Images