Message passing assisted fully-distributed Markov chain Monte Carlo detection method

Abstract

The invention discloses a message passing assisted fully-distributed Markov chain Monte Carlo detection method, and belongs to the technical field of wireless communication. The method comprises the following steps: S1, dividing a base station receiving antenna into a plurality of independent clusters, configuring a local baseband distributed unit for each cluster, and connecting the local baseband distributed unit with a central unit in a star topology manner; s2, each distributed unit executes local Markov chain Monte Carlo sampling by using a plurality of parallel samplers under cavity distribution regularization, and updates estimation; s3, the distributed units calculate local parameters, the central unit combines the parameters of the distributed units to obtain global observation statistics and issues regularization parameters, and the distributed units construct local cavity distribution; and S4, after multiple iterations, the central unit selects a candidate sample output result. According to the invention, based on a completely distributed baseband processing architecture, the interaction bandwidth is reduced by using message passing, and the Markov chain Monte Carlo sampling is combined to approach the optimal solution, so that the problems of high interaction overhead and high bit error rate in the prior art are solved.

Description

Technical Field

[0001] This invention relates to a message-passing-assisted fully distributed Markov chain Monte Carlo detection method, belonging to the field of wireless communication technology. Background Technology

[0002] In the development of wireless communication technology, Massive Multiple-Input Multiple-Output (MIMO) technology has made significant contributions to solving the problem of high bit error rate in signal detection caused by the increase in the number of users and achieving higher spectral and energy efficiency. However, as the number of antennas and users further increases, its design architecture faces more challenges.

[0003] In recent years, Decentralized Baseband Processing (DBP) has been introduced as an architecture to improve the efficiency of MIMO systems and reduce the interaction bandwidth between distributed units and the central processing unit. Traditional detection architectures based on DBP rely on multiple rounds of information exchange to achieve convergence, failing to achieve a good balance between bit error rate and interaction bandwidth. Gradient-accelerated Markov chain Monte Carlo (MCMC) high-dimensional signal detection technology offers a new approach to solving the signal detection problem. However, existing centralized Markov chain Monte Carlo schemes still involve uploading the raw detection signal and channel data to the central processing unit, resulting in significant bandwidth consumption. Some existing distributed Markov chain Monte Carlo schemes calculate gradients using local information in the distributed units and upload them to the central processing unit, which then samples the data and outputs the result after multiple iterations. This involves multiple rounds of information exchange at a high frequency, resulting in still relatively high interaction bandwidth. Summary of the Invention

[0004] Technical Problem: This invention addresses the aforementioned shortcomings of existing technologies by providing a message-passing-assisted fully distributed Markov chain Monte Carlo detection method. In the distributed units, a gradient-accelerated Markov chain Monte Carlo algorithm combined with cavity distribution information is used to achieve local sampling, generating high-density information which is then uploaded to the central processing unit. The central processing unit is only responsible for calculating cavity distribution parameters and broadcasting them to the distributed units. This achieves low-frequency, high-density information interaction, reduces interaction bandwidth, and improves performance.

[0005] Technical solution: The message-passing-assisted fully distributed Markov chain Monte Carlo detection method of the present invention includes the following steps:

[0006] S1. Divide the base station receiving antenna into multiple independent clusters, each cluster is configured with a local baseband distributed unit, and connected to the central unit in a star topology.

[0007] S2. Under cavity distribution regularization, each distributed unit performs local Markov chain Monte Carlo sampling using multiple parallel samplers to update the estimate;

[0008] S3. Distributed units calculate local parameters, the central unit merges the parameters of each distributed unit to obtain global observation statistics and distributes regularization parameters, and each distributed unit constructs local cavity distribution.

[0009] S4. After multiple iterations, the central unit selects candidate samples and outputs the results.

[0010] Furthermore, the detection method includes local Markov chain Monte Carlo sampling and global message passing; wherein, local Markov chain Monte Carlo sampling includes five steps: Nesterov gradient descent, random walk, mapping to constellation points, Metropolis-Hastington acceptance test, and determining the termination condition; global message passing includes three steps: each distributed unit calculates the local distribution and uploads the natural parameters, the central unit calculates the global distribution and downloads the natural parameters, and the distributed units calculate the cavity parameters.

[0011] Step S2 specifically includes:

[0012] S201. Set the message passing count T, the number of local samplers S, the maximum number of samples P, and the Nesterov continuous gradient descent steps as follows: The random walk step size is Learning rate Momentum factor is And initialize the current sampling count p=0.

[0013] S202. Perform Nesterov gradient descent continuously based on the local gradient and the cavity distribution gradient to obtain the result after gradient descent.

[0014] S203. Perform a random walk on the gradient descent result and project it onto the constellation points to obtain the sample to be tested by the Metropolis-Hastington line.

[0015] S204. Perform a Metropolis-Hastings acceptance test on the sample to be tested and the sample from the previous round of sampling. Use the ratio of the likelihood probabilities of the two samples as the acceptance probability. If the probability passes the test, the sample to be tested is accepted and added to the Markov chain. If it fails, the sample from the previous round of sampling is still taken and added to the Markov chain.

[0016] S205. Determine whether the current number of sampled variables p is greater than the maximum number of samples P. If not, jump to step S202; if yes, end step S2.

[0017] Step S3 specifically includes:

[0018] S301. Each distributed unit calculates the mean based on the sample, and the variance is given by the local scaling factor approximating the scaled identity matrix. Based on the above data, the natural mean and natural covariance of the skew distribution are calculated, and the natural mean and natural covariance of the cavity distribution are subtracted to obtain the locally estimated natural mean and natural covariance. The natural mean is then uploaded to the central unit.

[0019] S302. The central unit aggregates the local natural mean values ​​to obtain the global natural mean value and returns it to each distributed unit.

[0020] S303. Each distributed unit calculates the new round of cavity distribution natural mean based on the global natural mean and the local natural mean, calculates the cavity distribution natural covariance based on the scaling factor, and updates the cavity distribution natural mean and natural covariance to be used in the next round based on the suppression factor to ensure numerical stability.

[0021] In step S4, based on the Markov chain jointly constructed in steps S2 and S3, each distributed unit selects the sample with the smallest local residual norm and uploads it to the central unit. The central unit then selects and outputs the result with the smallest global residual norm.

[0022] Beneficial effects: Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0023] This invention proposes a message-passing-assisted fully distributed Markov chain Monte Carlo detection method within a distributed baseband processing architecture. The sampling process and related gradient calculations are completely decentralized to each distributed unit, and cluster-wide computing resources are fully utilized through parallel sampling chains. Simultaneously, a message propagation mechanism that only exchanges low-dimensional statistics is designed to avoid transmitting the complete received signal and channel matrix, significantly reducing inter-cluster interconnection bandwidth overhead while maintaining detection accuracy. Compared to existing centralized and distributed detection methods, this invention achieves near-maximum likelihood detection error rate performance in typical large-scale MIMO scenarios with fewer iterations, and achieves a better trade-off between computational complexity and interconnection load, improving the system's practicality and scalability in scenarios with large-scale antennas and dense deployments. Attached Figure Description

[0024] Figure 1 The flowchart of the message-passing-assisted fully distributed Markov chain Monte Carlo detection method is shown in this embodiment of the invention.

[0025] Figure 2 This is a structural diagram of the fully distributed detection method according to an embodiment of the present invention;

[0026] Figure 3 This is a comparison chart of the bit error rate performance of the method of the present invention and the comparative method. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] I. System Model Used in This Embodiment

[0029] Considering the uplink transmission signal detection scenario in a large-scale MIMO system, Each single-antenna user transmits uplink signals, and the base station, acting as a receiver, is equipped with... One antenna. Receiving signals. It can be represented as: ,in, It is receiving signals. It is the channel matrix. yes Signals transmitted by a single-antenna user It is additive noise with a power of .

[0030] In the aforementioned centralized large-scale MIMO uplink model Building upon this foundation, to reduce the complexity of baseband processing and interconnect bandwidth overhead, this invention further introduces a Distributed Baseband Processing (DBP) architecture. Specifically, the large-scale antenna array on the base station side is uniformly divided into C antenna clusters, each denoted as the c-th cluster. The c-th antenna cluster contains Each antenna has a local radio frequency link connected to its corresponding distributed unit, which performs local preprocessing and detection calculations on the received signal. Therefore, the MIMO uplink model for each distributed unit is further modified as follows: , .

[0031] The channel matrix, received signal, and noise are divided into rows. , , , This represents the transpose of a matrix.

[0032] II. Specific Steps of This Embodiment

[0033] like Figure 1 As shown in the figure, this embodiment of the invention provides a flowchart of a message-passing-assisted fully distributed Markov chain Monte Carlo detection method, which includes the following steps:

[0034] S1. Divide the base station receiving antenna into multiple independent clusters, each cluster is configured with a local baseband distributed unit, and connected to the central unit in a star topology.

[0035] S2. Under cavity distribution regularization, each distributed unit performs local Markov chain Monte Carlo sampling using multiple parallel samplers to update the estimate;

[0036] S3. Distributed units calculate local parameters, the central unit merges the parameters of each distributed unit to obtain global observation statistics and distributes regularization parameters, and each distributed unit constructs local cavity distribution.

[0037] S4. After multiple iterations, the central unit selects candidate samples and outputs the results.

[0038] Markov chain Monte Carlo (MCM) is a sampling method for complex posterior inference. It constructs a Markov chain with a stationary target distribution and iteratively generates samples according to the Metropolis-Hastings (MH) acceptance criterion to approximate the expected and optimal results. Combined with gradient acceleration using global information regularization, it can efficiently explore high-dimensional spaces and output the required statistics.

[0039] Distributed architecture is an advanced system design approach that divides the entire system into multiple cooperating modules, thereby distributing functionalities and reducing the computational burden on individual system units. In a distributed architecture, each submodule has independent processing capabilities, enabling it to autonomously complete specific tasks and interact with the central processing unit through efficient communication mechanisms to achieve the complex functions of the entire system.

[0040] Furthermore, distributed architectures are divided into partially decentralized and fully decentralized schemes. Partially distributed schemes distribute computational tasks to each local baseband unit and perform collaborative detection between the central processing unit and each local baseband unit, avoiding the transmission of original samples, but still need to bear a large interconnection load of exchanging intermediate results. Fully distributed schemes, on the other hand, completely transfer the computational burden to each local baseband unit, and only occasionally exchange consensus. This invention adopts a fully distributed architecture, fully utilizing its advantages to reduce the system's bit error rate and interaction costs.

[0041] Furthermore, the message passing described in this invention is based on expectation propagation. Expectation propagation is one of the approximate inference algorithms for product models. It decomposes the global posterior into multiple factors and approximates each factor with a solvable exponential family distribution, then updates iteratively to make the global approximation closely resemble the true posterior. Combining local observations and priors in a distributed scenario, expectation propagation uses lightweight natural parameters to transmit and fuse between nodes, achieving message passing and providing stable information constraints for local Markov chain Monte Carlo sampling. Furthermore, Figure 2The diagram illustrates the structure of the fully distributed detection method in this embodiment, which includes local Markov chain Monte Carlo sampling and global message passing. Local Markov chain Monte Carlo sampling comprises five steps: Nesterov gradient descent, random walk, mapping to constellation points, Metropolis-Hastington acceptance test, and determining the termination condition. Global message passing comprises three steps: each distributed unit calculates its local distribution and uploads natural parameters; the central unit calculates the global distribution and downloads natural parameters; and the distributed units calculate cavity parameters. Further, step S2 specifically includes:

[0042] S201. Set the message passing count T, the number of local samplers S, the maximum number of samples P, and the Nesterov continuous gradient descent steps as follows: The random walk step size is Learning rate Momentum factor is And initialize the current sampling count p=0.

[0043] S202, Based on the local residual norm cavity distribution residual Computable local objective function And calculate the local target gradient. ,in , The cavity distribution parameters are then used, and Nesterov gradient descent is performed continuously to obtain the results after gradient descent.

[0044] S203. Perform a random walk on the gradient descent result and project it onto the constellation points to obtain the sample to be tested by the Metropolis-Hastington line.

[0045] S204. Perform a Metropolis-Hastings acceptance test on the sample to be tested and the sample from the previous round of sampling. Use the ratio of the likelihood probabilities of the two samples as the acceptance probability. If the probability passes the test, the sample to be tested is accepted and added to the Markov chain. If it fails, the sample from the previous round of sampling is still taken and added to the Markov chain.

[0046] S205. Determine whether the current number of sampled variables p is greater than the maximum number of samples P. If not, jump to step S202; if yes, end step S2.

[0047] Furthermore, step S3 specifically includes:

[0048] S301. Each distributed unit calculates the mean based on the sample, and the variance is given by the local scaling factor pre-set based on empirical values, which is approximately the scaled unit matrix. Based on the above data, the natural mean and natural covariance of the tilted distribution are calculated, and the natural mean and natural covariance of the cavity distribution are subtracted to obtain the locally estimated natural mean and natural covariance. The natural mean is then uploaded to the central unit.

[0049] S302. The central unit aggregates the local natural mean values ​​to obtain the global natural mean value and returns it to each distributed unit.

[0050] S303. Each distributed unit calculates the new round of cavity distribution natural mean based on the global natural mean and the local natural mean, calculates the cavity distribution natural covariance based on the scaling factor, and updates the cavity distribution natural mean and natural covariance to be used in the next round based on the suppression factor to ensure numerical stability.

[0051] In step S4, based on the Markov chain jointly constructed in steps S2 and S3, each distributed unit selects the sample with the smallest local objective function and uploads it to the central unit. The central unit then selects the output with the smallest global residual norm. The result.

[0052] III. Implementation Results

[0053] To enable those skilled in the art to better understand the present invention, a performance comparison between a message-passing-assisted fully distributed Markov chain Monte Carlo detection method and a comparative method is presented below. The considered scenario uses a Rayleigh fading channel model, with data symbols employing Quadrature Amplitude Modulation (QAM) of order 16, and the number of base station receiving antennas is [missing information]. Each user is equipped with one transmitting antenna, totaling [number missing] Single-antenna user.

[0054] This embodiment evaluates the proposed detection method in terms of uncoded bit error rate (BER). The comparison methods include various centralized and distributed detection algorithms, such as centralized linear least squares, distributed expectation propagation, distributed alternating direction multiplier method, distributed large-scale MIMO approximate message passing, and mini-batch Nesterov Markov chain Monte Carlo methods. While ensuring sufficient convergence in the BER performance of each method, the message passing iterations of the proposed method remain at T=3, requiring only a small number of iterations per antenna cluster to complete the detection. Figure 3The diagram shows a comparison of the bit error rate performance between the proposed method and comparative methods. In a 64×16 scenario, the proposed method achieves a more significant performance gain compared to the baseline methods, and gradually approaches the lower bound of maximum likelihood detection as the SNR increases. This indicates that under high load conditions with large antenna size and a large number of users, the proposed method can still maintain good detection accuracy and robustness.

[0055] The results of the above implementation show that, in typical large-scale MIMO scenarios, the message-passing-assisted fully distributed Markov chain Monte Carlo detection method proposed in this invention outperforms existing distributed detection schemes in terms of uncoded bit error rate performance. It can achieve bit error rate performance close to maximum likelihood detection in fewer message-passing iterations and is more suitable for application in large-scale distributed base station systems.

[0056] The above description, in conjunction with the accompanying drawings, is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the claims. It should be understood that any equivalent changes made without departing from the spirit of the present invention are within the scope of protection covered by the claims.

Claims

1. A message-passing-assisted fully distributed Markov chain Monte Carlo detection method, characterized in that, Includes the following steps: S1. Divide the base station's receiving antenna array into multiple independent antenna clusters, and configure each antenna cluster with a local baseband distributed unit; All distributed units are connected to a central unit in a star topology; S2. Under the constraint of cavity distribution regularization, each distributed unit independently performs local Markov chain Monte Carlo sampling using multiple parallel samplers to update the local estimate of the transmitted signal. S3. Each distributed unit calculates local natural parameters based on local sampling results and uploads them to the central unit; the central unit aggregates the local natural parameters uploaded by all distributed units to obtain global natural parameters and broadcasts them to each distributed unit; each distributed unit constructs a local cavity distribution for the next iteration based on the received global natural parameters and local information. S4. Repeat steps S2 and S3 for multiple iterations. Each distributed unit selects candidate samples from its constructed Markov chain and uploads them to the central unit. The central unit selects the final detection result based on global criteria and outputs it.

2. The method according to claim 1, characterized in that, Step S2, the local Markov chain Monte Carlo sampling, specifically includes the following sub-steps: S201. Preset system parameters, including: number of message passing iterations, number of local parallel samplers, maximum number of samples per sampler, number of consecutive steps in Nesterov gradient descent, random walk step size, learning rate, and momentum factor. S202. For the current sampling point, combine the gradient of local observation and the gradient provided by the cavity distribution, and continuously perform Nesterov gradient descent to obtain the intermediate variables after gradient descent. S203. Perform a random walk operation on the intermediate variable and project the result onto a predetermined signal constellation point to generate a candidate sample to be accepted. S204. Perform the Metropolis-Hastington acceptance test on the candidate sample to be accepted and the sample accepted in the previous round; use the ratio of the likelihood probabilities of the two samples as the acceptance probability. If the test passes, add the candidate sample as a new sample to the Markov chain; otherwise, retain the sample from the previous round and add it to the Markov chain. S205. Determine whether the number of samples completed so far has reached the preset maximum number of samples; if not, return to step S202 to continue sampling; if it has reached the maximum number of samples, complete this local sampling.

3. The method according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S301. Each distributed unit calculates the local empirical mean based on the samples in the local Markov chain, and approximates the covariance as a scaled identity matrix using a local scaling factor; based on this, it calculates the natural mean and natural covariance of the local skew distribution, subtracts the natural mean and natural covariance of the current cavity distribution, and obtains the locally estimated natural mean and natural covariance; the locally estimated natural mean is then uploaded to the central unit. S302. The central unit performs a weighted aggregation of the local natural mean values ​​of all distributed units received to obtain the global natural mean value, and broadcasts it to all distributed units. S303. Each distributed unit calculates the natural mean of the new round of cavity distribution based on the received global natural mean and its own calculated local natural mean. The natural covariance of the cavity distribution is calculated based on the local scaling factor; The calculated cavity distribution natural parameters are updated using a suppression factor to ensure numerical stability during the iteration process.

4. The method according to claim 1, characterized in that, In step S4, each distributed unit selects the sample with the smallest local residual norm from its constructed Markov chain as a candidate sample and uploads it to the central unit; the central unit selects the sample with the smallest global residual norm from all candidate samples as the final detection result and outputs it.

5. The method according to claim 1, characterized in that, The method combines the expectation propagation framework with Markov chain Monte Carlo sampling; wherein, the expectation propagation framework is used to realize the message passing and fusion of low-dimensional natural parameters between distributed units and central units, providing regularization constraints for local sampling; the Markov chain Monte Carlo sampling is performed locally in the distributed units to approximate the posterior distribution of the signal.

6. The method according to claim 2, characterized in that, Nesterov gradient descent utilizes a momentum factor to accelerate convergence, and its update process depends simultaneously on the gradient of the local likelihood function and the regularized gradient derived from the cavity distribution.

7. The method according to claim 3, characterized in that, The suppression factor is used to prevent excessive growth of parameter values ​​when iteratively updating the natural mean and natural covariance of the cavity distribution, thereby ensuring the convergence and stability of the message passing process.

8. The method according to any one of claims 1 to 7, characterized in that, The method is applied to uplink signal detection scenarios in large-scale multiple-input multiple-output systems; the distributed architecture is a fully distributed architecture, and the central unit does not perform specific sampling calculations, but is only responsible for parameter aggregation and distribution.

Images