MIMO (Multiple Input Multiple Output) system optimization method and device and computer equipment

Abstract

The invention relates to the technical field of telecommunication, and discloses an MIMO system optimization method and device and computer equipment, and the method comprises the steps: obtaining a modulation constellation set of a sending signal of each user of an MIMO system and a receiving signal of each antenna of a base station, and constructing a factor graph of the MIMO system; carrying out iterative calculation based on the factor graph to obtain a symbol posterior probability of each user; the symbol posterior probability comprises the confidence coefficient of each symbol in the modulation constellation set; calculating a symbol posterior probability based on the token feature in any iteration round of the iterative calculation; the token feature is determined according to a potential state of a previous iteration round, a message passing posterior statistic of a current round and a projection residual error obtained after a residual error of an observation value and an estimation value of the current round is projected to a specified channel; selecting a symbol with the maximum confidence coefficient in the symbol posterior probability as an estimated sending symbol of the user; the estimated transmit symbol is used to generate a transmit signal of the user. The method has the beneficial effect that the symbol detection accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of telecommunications technology, and in particular to a method, apparatus and computer device for optimizing a MIMO system. Background Technology

[0002] Multiple-input multiple-output (MIMO) systems have been widely used as a core technology for improving spectrum efficiency in modern wireless networks. However, in large-scale MIMO systems, achieving reliable symbol detection remains a formidable challenge due to severe inter-symbol interference and the exponential computational complexity of maximum likelihood (ML) detection. Summary of the Invention

[0003] This application provides a MIMO system optimization method, apparatus, and computer device, which solves the technical problem of how to improve symbol detection efficiency in large-scale MIMO systems and achieves the technical effect of improving symbol detection efficiency.

[0004] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a MIMO system optimization method, the method comprising: Obtain the modulation constellation set of the transmitted signals of each user in the MIMO system and the received signals of each antenna of the base station, and construct the factor graph of the MIMO system; The symbol posterior probability of each user is calculated iteratively based on the factor graph; the symbol posterior probability includes the confidence of each symbol in the modulation constellation set; the symbol posterior probability is calculated based on the token feature in any iteration round of the iterative calculation; the token feature is determined according to the latent state of the previous iteration round, the message passing posterior statistic of the current round, and the projection residual obtained by projecting the residual of the observation and the estimate of the current round onto a specified channel; The symbol with the highest confidence among the symbol posterior probabilities is selected as the user's estimated transmission symbol; the estimated transmission symbol is used to generate the user's transmission signal.

[0005] In this embodiment, a factor graph is constructed by acquiring the transmitted signals of the MIMO system users and the received signals of the base station antenna. Based on the factor graph, the symbol posterior probability is calculated iteratively using token features. Then, the method of estimating the transmitted symbol is determined according to the symbol posterior probability, thereby improving the accuracy of symbol detection in the multi-user MIMO system and thus improving the communication effect and efficiency of the MIMO system.

[0006] Secondly, embodiments of this application provide a MIMO system optimization apparatus, the apparatus comprising: The acquisition module is used to acquire the modulation constellation set of the transmitted signals of each user in the MIMO system and the received signals of each antenna of the base station, and to construct the factor graph of the MIMO system. The calculation module is used to iteratively calculate the symbol posterior probability of each user based on the factor graph; the symbol posterior probability includes the confidence of each symbol in the modulation constellation set; in any iteration round of the iterative calculation, the symbol posterior probability is calculated based on the token feature; the token feature is determined according to the latent state of the previous iteration round, the message passing posterior statistic of the current round, and the projection residual obtained by projecting the residual of the observation value and the estimate of the current round onto a specified channel; An optimization module is used to select the symbol with the highest confidence among the symbol posterior probabilities as the estimated transmission symbol for the user; the estimated transmission symbol is used to generate the user's transmission signal.

[0007] Thirdly, embodiments of this application provide a computer device, including: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method described in any of the above embodiments.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions, which are used to cause a computer to perform the method described in any one of the above embodiments.

[0009] Fifthly, embodiments of this application provide a computer program product, including computer instructions, which are used to cause a computer to perform the method described in any of the above embodiments. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0011] Figure 1 A flowchart of a MIMO system optimization method provided in this application embodiment; Figure 2 A schematic diagram of an iterative process provided in an embodiment of this application; Figure 3A schematic diagram of an iterative process provided in an embodiment of this application; Figure 4 A schematic diagram of a comparison result provided for an embodiment of this application; Figure 5 A block diagram of a MIMO system optimization device provided in this application embodiment; Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] Multiple-input multiple-output (MIMO) systems, as a core technology for improving spectral efficiency in modern wireless networks, have been widely applied. However, in massive MIMO systems, achieving reliable symbol detection remains a significant challenge due to severe inter-symbol interference and the exponentially increasing computational complexity of maximum likelihood (ML) detection. This demand for efficient and high-performance MIMO detection is particularly urgent in applications involving high-order modulation and large-scale antenna arrays.

[0014] To address the aforementioned issues, current research primarily employs algorithms such as early neural networks, graph neural networks (GNNs), and attention mechanisms to construct detectors and achieve symbol detection.

[0015] Early neural detectors, such as Deterministic Networking (DetNet), Orthogonal Approximate Message Passing Network (AMPNet), and Multi-scale Deep Learning Network (MMNet), adopted a model-driven approach, unfolding iterative algorithms into neural network architectures and learning key parameters to enhance convergence speed and robustness.

[0016] Graph neural network detectors, such as the Approximate Message Passing Graph Neural Network (AMP-GNN) and the Graph Expectation Propagation Network (GEPNet), were proposed to refine the posterior statistics in message passing inference. However, while GNN-based detectors can effectively model local dependencies, their computational patterns rely on irregular graph computations. In terms of hardware implementation, this involves numerous small kernel startups and non-contiguous memory access patterns, failing to fully utilize parallel computing resources and resulting in low runtime efficiency, making it difficult to meet the low-latency requirements of large-scale MIMO systems.

[0017] Attention-based detectors, such as Reconfigurable Intelligent MIMO (RE-MIMO), can capture inter-symbol correlations using attention mechanisms and leverage the global modeling and parallelism of the Transformer architecture, thus providing runtime efficiency superior to GNNs. However, while attention-based RE-MIMO solves the computational efficiency problem of GNNs, it redundantly injects global channel and residual information into the network, leading to model over-parameterization, increasing training difficulty and inference latency, and limiting practical application deployment and further performance improvements.

[0018] Specifically, high-precision detection methods often come with high computational overhead and inefficient runtime performance, while efficient methods fail to fully realize their potential due to redundant model design, resulting in limited performance improvements. Therefore, how to significantly improve computational efficiency while maintaining high accuracy is a major technical problem that urgently needs to be solved in the current MIMO detection field. To this end, this application aims to provide a more compact, scalable, and expressive MIMO signal detection method.

[0019] Building upon existing deep learning-based MIMO detection techniques, this application proposes a novel neural detector (AttentionMasked and Permuted Pre-training for LanguageUnderstandingNet, AttnMPNet) that integrates message passing iterative ideas with attention mechanisms. This detector achieves a highly parallelizable architecture and compact, efficient feature representation.

[0020] This application replaces the irregular graph computation of traditional GNNs with a parallelizable single-layer Transformer module, realizing global modeling of dependencies between symbols. This fundamentally solves the computational efficiency bottleneck of GNN methods, enabling them to achieve significantly faster inference speeds when dealing with large-scale MIMO systems.

[0021] Furthermore, this application constructs concise and expressive token features by combining the previous round's latent state, message passing posterior statistics, and low-dimensional normalized projection residuals. This avoids redundant information injection, reduces model complexity, and enhances feature expressiveness, achieving superior runtime performance while providing near-maximum likelihood detection accuracy.

[0022] According to an embodiment of this application, a method for optimizing a MIMO system is provided. It should be noted that the steps shown in the flowcharts in the accompanying drawings can be executed on a computer device via a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here. The computer device may be located at a base station.

[0023] Figure 1 A flowchart illustrating a MIMO system optimization method provided in this application embodiment, in the multi-user MIMO system of this application, such as... Figure 1 As shown, with computer equipment as the executing entity, this process includes the following steps: S101. Obtain the modulation constellation set of the MIMO system and the received signals of each antenna of the base station, and construct the factor graph of the MIMO system.

[0024] For example, the MIMO system used in this application may include multiple users. Each user may be a single-antenna user. Furthermore, the MIMO system may include a multi-antenna base station.

[0025] The computer equipment can first acquire the set of modulation constellations used by each user in the MIMO system. This set of modulation constellations may include multiple symbols that each user may use. The computer equipment can also acquire the received signals received by each antenna of the base station.

[0026] Then, the computer equipment can use the modulation constellation set and the received signal to construct a factor graph of the MIMO system. This factor graph can clearly show the correlation and interaction between the antennas of each user and the base station in the system, providing a basis for subsequent calculations and analysis.

[0027] In one implementation, the computer device can treat each symbol in the modulation constellation set as a variable node in the factor graph. This variable node can be denoted as... This set of modulated constellations can be denoted as... This variable node can also be called a variable node.

[0028] In one implementation, the modulation constellation set may include multiple symbols. These symbols are used to generate the user's transmitted information.

[0029] In one implementation, the computer device can use the received signal from each antenna of the base station as a factor node in the factor graph. Each received signal Model as factor nodes .

[0030] In one implementation, the received signal is data superimposed with multiple transmitted signals received by the base station antenna.

[0031] In one implementation, in a multi-user MIMO system, A single-antenna user simultaneously transmits to the equipped The base station transmits signals using one antenna. The baseband system model can be represented as: (1) in, This is the received signal from the base station. The tensor representation can be . This is the channel matrix. The signal matrix In the expression, each element independently follows a complex Gaussian distribution with zero mean and unit variance. This is the transmit symbol. This transmit symbol... It can be represented as The sending symbol Each symbol in the text is independently derived from a discrete constellation. . This is additive white Gaussian noise (AWGN). It can be written as .

[0032] In one implementation, the computer device can estimate the symbolic posterior probability of each user through approximate reasoning. This method estimates the symbols used by each user. It avoids direct hard decision-making and improves the accuracy of determining the symbols used by users.

[0033] In one implementation, the probabilistic model for calculating the symbolic posterior probability of each user can be represented as an iterative message passing of a factor graph. Therefore, the computer device can achieve efficient inference of the symbolic posterior probability of each user based on the fully connected bipartite factor graph constructed from the aforementioned variable nodes and factor nodes.

[0034] In one implementation, a computer device can encode local likelihood constraints based on the factor graph. The reasoning in the factor graph is achieved through iterative message passing between variable nodes and factor nodes, specifically including the following steps: (2) in, Indicates from variable node To factor node The news. Used to represent integrated variable nodes Adjacent factor nodes The message does not include factor nodes. .

[0035] in, Characteristic factor node The local likelihood function. Used to represent integration factor nodes Adjacent variable nodes The message does not include variable nodes. .

[0036] in, The calculation formula can be written as: (3) in, For the base station The received signal corresponding to each antenna. Channel matrix The first in Each element. Indicates the first The variable node corresponding to each user. This represents the desired signal term. This represents the noise variance.

[0037] After sufficient iterations, the symbol The marginal posterior probability is approximately: (4) S102. The symbol posterior probability of each user is calculated iteratively based on the factor graph. The symbol posterior probability includes the confidence level of each symbol in the modulation constellation set. In any iteration round of the iterative calculation, the symbol posterior probability is calculated based on the token feature. The token feature is determined based on the latent state of the previous iteration round, the message passing posterior statistic of the current round, and the projection residual obtained by projecting the residual of the observation and the estimate of the current round onto the specified channel.

[0038] For example, the computer device, based on the constructed factor graph, uses a preset signal detector to perform iterative calculations and outputs the symbol posterior probability for each user. This symbol posterior probability includes the confidence level of each symbol in the modulation constellation set. This confidence level reflects the likelihood of using the symbol to generate transmitted information. The higher the confidence level, the better the effect of using the symbol to generate transmitted information, and the greater the likelihood that the user will use the symbol to generate transmitted information.

[0039] In each iteration, the computer device can use a token feature to calculate the symbol posterior probability. This token feature is determined based on the latent state of the previous iteration, the posterior statistic after message passing in the current iteration, and the projected residual obtained by projecting the residual between the observed value and the estimated value of the current iteration onto a specified channel. During this iteration, the computer device can iteratively approximate the accurate symbol posterior probability.

[0040] In one implementation, the modulation constellation set is a set of discrete symbols used to represent information in a communication system; different modulation schemes correspond to different symbol sets. Each transmitted signal typically corresponds to one symbol in this modulation constellation set. This modulation constellation set can also be called a constellation.

[0041] In one implementation, the latent state is a feature calculated during the iteration process, based on the latent state of the previous iteration and information such as the refined token features, log odds, the symbolic posterior statistic of the symbolic posterior probability, and the residual change in the current iteration. This feature is used to capture historical inference context and temporal correlations across iterations. In each iteration, each variable node can correspond to a latent state. The latent state calculated in each iteration can be used for the calculation in the next iteration.

[0042] In one implementation, the message passing posterior statistic is statistical information about the symbol probability calculated during the current round of message passing. This message passing posterior statistic may include the mean and variance.

[0043] In one implementation, the observed value is the received signal actually acquired by the base station's antenna.

[0044] In one implementation, the estimated value is the received signal of each antenna of the base station, which is estimated by the computer device based on the symbolic posterior statistics and signal matrix calculated in the previous iteration.

[0045] In one implementation, the projection residual is the symbol-level reconstruction error obtained by projecting the residual between the observation and the estimate onto a specified channel.

[0046] In one implementation, attention mechanisms are widely used to model pairwise dependencies between input elements. During the iterative process of this model, since there are mutual influences between the variable nodes, the computer device can employ a multi-head attention (MHA) architecture to efficiently extract structured features between the variable nodes. A given input order can be denoted as: (5) in, For sequence length, The input feature dimension. For the first User characteristic data.

[0047] MHA first uses a head-specific weighting to... Projection as A group of queries, keys, and values. This query, key, and value can be represented as: (6) in, These are learnable parameters. . It is a single-head dimension. For attention.

[0048] Each attention head The output calculation formula can be: (7) in, These are learnable or input-related attention biases that can be used to encode prior structural information. .

[0049] The outputs of all attention heads are obtained after concatenation and linear transformation: (8) in, This is for outputting the projection matrix. In practical applications, the settings are as follows: This is to ensure dimensional consistency.

[0050] S103. Select the symbol with the highest confidence among the symbol posterior probabilities as the user's estimated transmission symbol. The estimated transmission symbol is used to generate the user's transmission signal.

[0051] For example, after obtaining the symbol posterior probabilities for each user, the computer device selects the symbol with the highest confidence from the confidence levels included in the symbol posterior probabilities as the estimated transmission symbol for that user. This estimated transmission symbol is considered to be the symbol used to generate the transmission signal.

[0052] In this embodiment, a factor graph is constructed by acquiring the transmitted signal of the MIMO system user and the received signal of the base station antenna. Based on the factor graph, the symbol posterior probability is calculated iteratively using token features. Then, the method for estimating the transmitted symbol is determined based on the symbol posterior probability, thereby achieving the effect of accurate signal estimation and optimization in the MIMO system.

[0053] In one example, the iterative calculation in step S102 above may include the following steps in each iteration: S1021. Construct token features based on the potential state of each user in the previous iteration round, the posterior statistics of message passing in the current round, and the projection residual.

[0054] S1022. Based on the refined token features, calculate the symbolic posterior probability and latent state of each user.

[0055] For example, during the iteration process, the computer device comprehensively utilizes three aspects of data: the latent state information calculated by each user in the previous iteration, the posterior statistics generated during message passing in the current iteration, and the projected residuals, to construct token features. After constructing the token features, the computer device can refine them. Then, using the refined token features, the computer device can calculate the symbolic posterior probability and latent state corresponding to each user in the current iteration using a specific calculation method.

[0056] In one implementation, the latent state of the previous iteration refers to the state of the variable node as determined by the computer device during the previous iteration. An estimate and representation of the characteristics of can be denoted as: .in, This indicates the previous iteration.

[0057] In one implementation, the posterior statistics generated during the current round of message passing can include, under the assumption that the observations are independent, aggregating all messages to variable nodes. The variable node formed The mean and variance of Gaussian posterior beliefs. These mean and variance can be denoted as... . This indicates the current iteration.

[0058] In one implementation, the projection residual captures the projection onto the first... The symbol-level reconstruction error in each user direction. As a compact and informative feedback signal, it effectively guides iterative optimization. This projection residual can be denoted as... .

[0059] In one implementation, the token feature can be obtained by concatenating multiple features mentioned above. Optionally, this token feature can be denoted as... The construction of this token integrates the aforementioned information, providing effective data support for subsequent calculations of the symbol's posterior probability and latent state. Among these, This represents the potential state from the previous iteration. and This refers to the posterior statistics generated during the current round of message passing. This represents the projection residual.

[0060] In one implementation, the computer device can refine the token features based on an attention mechanism.

[0061] In one implementation, the computer device can refine the card features using a Transformer module. Optionally, this Transformer module includes Multi-Head Self-Attention (MHA) and a Feedforward Network (FFN). In one implementation, the computer device can obtain the posterior probability of each symbol relative to the modulated signal symbols in the modulation constellation set through posterior prediction.

[0062] In one implementation, the potential state is set to zero during initialization.

[0063] In this embodiment, the iterative process is completed by constructing token features based on the latent state of the previous iteration, the posterior statistics of message passing in the current iteration, and the projection residual. The refined token features are then used to calculate the symbol posterior probability and latent state of each user. By obtaining the symbol posterior probability of each user at the end of the iteration and estimating the symbols sent by the users, the efficiency of symbol estimation in the multi-user MIMO system is improved, thereby enhancing the communication performance and efficiency of the MIMO system.

[0064] In one example, step S1021 above, where the computer device obtains the latent states of each user in the previous iteration round, the posterior statistics of message passing in the current round, and the projected residuals needed to construct token features, may include: Step 211: Generate a factor graph by treating each transmitted symbol as a variable node and the received signal of each antenna as a factor node.

[0065] For example, a computer device can... Each user as a variable node And, computer equipment can be based on the... Each user generates variable nodes from the symbols in the transmitted signal output by the antenna. The characteristics of this. Furthermore, computer equipment can use the received signals from each antenna as factor nodes. Computer devices can use these variable nodes. and factor nodes According to variable nodes With factor nodes To establish the communication relationships between them, a factor graph is constructed.

[0066] In one implementation, the transmitted symbol is a specific symbol used to carry information in the transmitted signal output by the user through the antenna of their terminal device. In a communication system, signals can be represented using symbols. This transmitted symbol is a symbol from a constellation set.

[0067] In one implementation, the set of transmitted symbols It can be represented as . express A collection of constellations Send symbol set Each transmitted symbol is independently taken from a constellation set. . This refers to the number of users per antenna. It can be used to represent the first Sending symbols for each user.

[0068] In one implementation, the received signal is the signal received by the antenna. Each received signal may include the superposition of the transmitted signals output by each user. The base station of this MIMO system may be configured with... One antenna. The received signal of the base station can be denoted as... The received signal The tensor representation can be . For the first The received signals from each antenna. Each received signal Model as factor nodes .

[0069] Step 212: Based on the factor graph, calculate the Gaussian belief of each variable node to obtain the posterior statistic of message passing for each variable node in the current round.

[0070] For example, after generating the factor graph, the computer device calculates the Gaussian belief of each variable node based on the graph's structure and related probability calculation methods. The Gaussian belief reflects the probability distribution of the symbols represented by each variable node given the factor graph structure and known information. Through the calculated Gaussian beliefs, the computer device can extract the posterior message passing statistics for each variable node in the current round of message passing.

[0071] In one implementation, Gaussian belief is a probability representation based on a Gaussian distribution. In a factor graph, Gaussian belief is used to describe the probability distribution of the symbols represented by the variable nodes. Assuming that the probability distribution of a symbol follows a Gaussian distribution, a computer can calculate its mean, variance, and other parameters to determine the probability distribution of that symbol.

[0072] In one implementation, the posterior statistic of message passing may include the mean and variance.

[0073] In one implementation, the computer device can perform a Gaussian approximation on the standard sum-product belief propagation (BP) algorithm on the MIMO factor graph at the beginning of each iteration, for each variable node. Generate a Gaussian belief. The formula for the Gaussian belief of each variable node is: (9) in, Indicates the sequence number of the current iteration. That is, it is currently at the [number]th iteration. The next iteration. Represents variable nodes Gaussian beliefs in the current iteration. The variable node of the previous iteration The mean. The variable node of the previous iteration The variance.

[0074] This belief serves as the input feature for the attention module. This applies to the received signal corresponding to each antenna of the base station. Related factor nodes This application can receive signals The observations are denoted by the following formula: (10) in, This represents the desired signal term. Channel matrix The first in Each element. For the first A symbol. Indicates a distractor. For the base station Additive white Gaussian noise for the root antenna.

[0075] Under the assumption of Gaussian approximation reasoning, the interference term It is approximated as a Gaussian random variable with a mean of The variance is .in, and It is from the variable node in the previous iteration To factor node News The mean and variance of the symbolic posterior statistics of the included symbolic posterior probabilities.

[0076] Thus, a factor node is obtained. To variable node The Gaussian message. The formula for the Gaussian message can be written as: (11) in, Indicates the factor nodes in this iteration To variable node information. This represents the desired signal term. For factor nodes To variable node Equivalent observations of the mean and variance of a Gaussian distribution.

[0077] The formula for calculating equivalent observations can be written as: Under the assumption that observations are independent, this application aggregates all messages into variable nodes. This forms a Gaussian posterior belief. This Gaussian posterior belief can be denoted as: (12) In one implementation, based on the above formula (12), the computer device can obtain the mean of the posterior statistic of message passing during the calculation process. and variance .

[0078] Step 213: Normalize the residuals determined by the observations and estimates based on the channel matrix for each variable node to obtain the projected residuals.

[0079] For example, for each variable node, the computer device first acquires the observed and estimated values ​​based on the channel matrix. Then, based on these observed and estimated values, the computer device calculates the residuals between the observed and estimated values. Finally, the computer device normalizes the residuals, transforming them to a specific range to obtain the projected residuals.

[0080] In one implementation, the channel matrix is ​​a matrix used to describe the transmission characteristics of a signal in a communication channel. It contains information such as channel gain and phase, and reflects the fading, interference, and other conditions experienced by the signal during transmission.

[0081] In one implementation, the observations are channel-related data obtained through actual measurements, such as the strength and phase of the received signal. These observations reflect the true state of the channel during actual transmission.

[0082] In one implementation, the estimated value is an approximation of the channel characteristics obtained by modeling and estimating the parameters of the channel. For example, the element values ​​of the channel matrix can be estimated using pilot signals.

[0083] In one implementation, the residual is the difference between the observed value and the estimated value. It represents the degree of deviation between the estimated value and the actual observed value, and the magnitude of the residual can reflect the accuracy of the estimation.

[0084] In one implementation, the projected residual is the result obtained after normalizing the residual. Normalization can transform the residual into a uniform range, which facilitates subsequent comparison and analysis with other data.

[0085] In one implementation, the formula for calculating the projection residual can be: (13) in, This is the normalized projection residual scalar. This represents the noise variance.

[0086] in, Channel matrix The conjugate transpose of . Represents the observed value Compared with the estimate of the previous iteration The residuals. Among them, the observed values... This is the received signal from the base station. (Estimated value) Based on the channel matrix The mean of each sign in the sign posterior statistic of the sign posterior probability in the previous iteration. Calculated. This indicates that a matched filtering or projection operation is performed on the residual to extract the th element. Information related to the transmitted signal.

[0087] in, Indicates channel gain Square and scaling factor The weighted sum. Where, channel gain... Channel matrix The first in Each element represents the transmitting antenna. to receiving antenna Channel gain. Scaling factor. This represents the variance of the symbolic posterior statistic of the symbolic posterior probability in the previous iteration.

[0088] Step 214: Combine the latent state of the previous iteration, the posterior statistics of message passing in the current iteration, and the projection residual to construct token features.

[0089] For example, the computer device can directly obtain the latent state stored during the execution of the previous iteration. Furthermore, the computer device can obtain the message passing posterior statistic for the current iteration according to step 212 above. The computer device can also obtain the projection residual according to step 213 above. Subsequently, the computer device can construct token features using the latent state, message passing posterior statistic, and projection residual according to preset splicing rules.

[0090] In one implementation, the token feature integrates information such as the potential state of the previous iteration, the posterior statistics of message passing in the current iteration, and the projected residual, providing richer and more accurate information for subsequent calculations and analysis.

[0091] In one implementation, the token feature serves as a compact and information-rich feedback signal that effectively guides iterative optimization. Furthermore, due to the low-dimensionality of information such as the latent state, the posterior statistics of the current round's message passing, and the projected residuals, the overhead introduced into the token representation is negligible.

[0092] In one implementation, when constructing token features, the computer device can use a concatenation method to concatenate the potential state of the previous iteration, the message passing posterior statistics of the current iteration, and the projection residual into a feature tensor according to a preset order.

[0093] This example demonstrates the effect of fully integrating information from multiple sources.

[0094] In one example, the calculation process of the posterior statistic of message passing in the current round in step 212 above includes: Step 2121: Based on the channel matrix and the symbol posterior statistic of the symbol posterior probability in the previous iteration, calculate the mean and variance of the Gaussian distribution.

[0095] For example, a computer device can combine a channel matrix Median The symbolic posterior statistic of the symbolic posterior probability in the previous iteration. and and received signals and noise variance The mean and variance of the Gaussian distribution are calculated.

[0096] The formulas for calculating the mean and variance of the Gaussian distribution can be written as: (14) in, For factor nodes To variable node The equivalent observation of the mean of the Gaussian distribution. For the base station The received signal from the antenna. Channel matrix The first in Each element. This is the mean of the symbolic posterior statistic of the symbolic posterior probability in the previous iteration. Mean of the interference terms. The middle can be accumulated except Apart from that, all other variable nodes to factor nodes The signal.

[0097] in, For factor nodes To variable node The equivalent observed values ​​of the variance of the Gaussian distribution. This represents the variance of the symbolic posterior statistic of the symbolic posterior probability in the previous iteration. The variance of the interference term. The middle can be accumulated except Apart from that, all other variable nodes to factor nodes The signal. This represents the noise variance.

[0098] Step 2122: Based on the channel matrix and the variance of the Gaussian distribution, calculate the variance of the posterior statistic of message passing for each variable node in the current round.

[0099] For example, the computer device obtains factor nodes in the current round. To variable node The variance of the Gaussian distribution Then, combined with the channel matrix Median The variance of the message passing posterior statistic of the variable node is calculated. This variable node Variance of posterior statistic of message passing The calculation formula can be written as: (15) in, This represents the variance of the posterior statistic for the current round of message passing. For factor nodes To variable node The variance of the Gaussian distribution. Channel matrix The first in Each element. The middle can be accumulated with All adjacent factor nodes The signal.

[0100] Step 2123: Calculate the mean of the message passing posterior statistic based on the channel matrix, the mean and variance of the Gaussian distribution, and the variance of the message passing posterior statistic.

[0101] For example, computer devices comprehensively utilize combined channel matrices Median Current round factor node To variable node The mean of the Gaussian distribution and variance and the variance of the message passing posterior statistic calculated in step 2122. The mean of the posterior statistic of message passing was calculated. The mean of the posterior statistic of the message passing. The calculation formula can be written as: (16) in, This represents the mean of the posterior statistic for message passing in the current round. This represents the variance of the posterior statistic for the current round of message passing. For factor nodes To variable node The variance of the Gaussian distribution. For factor nodes To variable node The mean of the Gaussian distribution. Channel matrix The first in Each element is a conjugate channel coefficient.

[0102] In this example, the Gaussian distribution parameters are calculated based on the channel matrix and the symbolic posterior statistic from the previous iteration. Then, the variance and mean of the posterior statistic for the current round of message passing are calculated using the channel matrix and the Gaussian distribution parameters. This achieves the effect of accurately quantifying the probability distribution characteristics of variable nodes in the current round of message passing.

[0103] In one example, step 213 above, the projected residual obtained by projecting the residual between the observed value and the estimated value of the current round onto a specified channel, includes: Step 2134: Calculate the difference between the observed value and the estimated value to obtain the residual.

[0104] For example, the computer device first needs to calculate the observed value and the estimated value separately. Then, the computer device can obtain the residual by calculating the difference between the observed value and the estimated value.

[0105] In one implementation, the observed value is specifically the received signal from the base station. This observation typically includes the combined effects of transmitted signal, channel effects, and noise.

[0106] In one implementation, the estimate can be based on the output information of the previous iteration and the channel matrix. The calculated value is denoted as: .

[0107] In one implementation, since the signal model of this application is a linear system, the computer device can calculate the difference between the observed value and the estimated value by subtracting them element by element, thereby obtaining the residual. This residual can specifically be a tensor.

[0108] In one implementation, the residual can be denoted as .

[0109] Step 2135: Based on the transpose of the channel matrix, calculate the projection of the residuals onto the symbol dimension.

[0110] For example, the computer device invokes a pre-stored channel matrix and calculates its conjugate transpose. The computer device can then multiply this conjugate transpose with a tensor of the residual, thereby projecting the residual onto the dimension space of the transmitted symbols.

[0111] In one implementation, the projection operation utilizes channel characteristics to weight and integrate the residuals, extracting the error components associated with each transmitted symbol. Symbol-dimensional projection refers to mapping the residuals to the coordinate system corresponding to the transmitted symbol through a linear transformation.

[0112] In one implementation, the projection calculation can be denoted as: .

[0113] Step 2136: Normalize the projection to obtain the projection residual.

[0114] For example, the computer device normalizes the symbolic dimension projection result by dividing it by a normalization factor to obtain the final projection residual.

[0115] In one implementation, the normalization factor consists of noise variance and a channel energy weighting term. This normalization operation can eliminate the influence of channel gain and noise level, giving the projection residual a uniform statistical scale, which facilitates subsequent signal detection or decoding processing.

[0116] In one implementation, the noise variance in the normalization factor characterizes the background noise intensity at the receiver. This noise variance can be denoted as... .

[0117] In one implementation, the weighted sum of the squared channel gains in the normalization factor reflects the signal aggregation effect of the multi-antenna system. This weighted sum of the squared channel gains can be denoted as... In this example, by calculating the residual between the observed and estimated values, projecting the residual onto the sign dimension, and normalizing the projection results, the residual information is efficiently extracted and processed, thus improving the efficiency of subsequent use of the residual information.

[0118] In one example, in step S1022 above, after obtaining the token features, the symbolic posterior probability and latent state of each user are calculated based on the refined token features, including: Step 221: Use attention mechanism and feedforward network to process token features to obtain refined token features.

[0119] For example, a computer device can capture the correlation between token features of various symbols through an attention mechanism, and perform nonlinear transformation on the token features through a feedforward network, ultimately outputting refined token features with adjusted dimensions.

[0120] In one implementation, the attention mechanism and feedforward network can be implemented using a single-layer Transformer module. The attention mechanism is a computational module that dynamically assigns weights based on feature importance; the feedforward network is a multilayer perceptron composed of linear layers and nonlinear activation functions.

[0121] In one implementation, the computer device can linearly embed and normalize the token features to form symbolic features. These symbolic features can then be used as input to an attention mechanism.

[0122] In one implementation, the refined token features after dimension adjustment have fewer data features and better eigenvalues, thereby reducing the amount of computation while ensuring computational efficiency.

[0123] In one implementation, the computer device can enhance the model's ability to express signal features by refining the token features.

[0124] In one implementation, directly using token features for computation relies on the assumption that observations are conditionally independent for each variable node. However, this assumption typically does not hold true in large-scale MIMO systems due to the cyclic nature of the underlying factor graph. Consequently, strong statistical dependencies may exist between the posterior distributions of different symbols. To address this, in this step, the computing device improves the effectiveness of token features by introducing a self-attention mechanism to extract the correlations between variable nodes.

[0125] Step 222: Input the refined token features, the latent state of the previous iteration, the posterior statistic of message passing in the current iteration, and the projection residual into the predictor to predict the log odds and symbol posterior probabilities of each variable node in the current iteration. The log odds include the probability parameter for each symbol in the modulation constellation set.

[0126] For example, the computer device concatenates multi-source information such as refined token features, the latent state of the previous iteration, the posterior statistic of message passing in the current iteration, and the projected residual, and inputs it into the predictor. It then generates the raw log-probability of each symbol in the modulation symbol set through a nonlinear mapping. The computer device can then convert this log-probability into a symbol posterior probability. The log-probability includes the probability parameter for each symbol in the modulation constellation set. The symbol posterior probability includes the probability value for each symbol in the modulation constellation set.

[0127] In one implementation, the multi-source information concatenation feature, including the refined token features, the latent state of the previous iteration, the posterior statistics of message passing in the current iteration, and the projected residual, can be denoted as: .in, These are the characteristics of the refined token. . This represents the potential state from the previous iteration. This represents the projection residual. and This is the posterior statistic of message passing in the current round.

[0128] In one implementation, the computer device can normalize the spliced ​​features of multi-source information to obtain the input features of the predictor. This process can be denoted as: (17) in, These are the input features of the normalized predictor. Representation layer normalization.

[0129] In one implementation, the predictor can employ a fully connected network or a recurrent neural network. This predictor may include a shared multilayer perceptron (MLP). The predictor is implemented as a two-layer MLP with a hidden dimension of... And the input tensor is mapped to using the GELU activation function. The log-odds (logits) output by this predictor can be denoted as... . . For constellations, which include a collection of all symbols.

[0130] In one implementation, the computer device can calculate the signed posterior probability by applying the softmax function to the log-odds. This calculation process can be written as: (18) in, Let be the posterior probability of the sign. . It is a logarithmic probability.

[0131] In one implementation, the initial log odds are obtained by estimating the minimum mean square error (MMSE).

[0132] In one implementation, the logarithmic probability output in the Tth iteration It is used to make hard or soft judgments.

[0133] Step 223: Calculate the latent state of the current iteration based on the latent state of the previous iteration, the refined token features of the current iteration, the log odds, the symbolic posterior statistic of the symbolic posterior probability, and the residual change.

[0134] For example, the computer device integrates the latent state of the previous iteration with the refined token features, log odds, symbolic posterior statistics of the symbolic posterior probability, residual change and other multi-dimensional information of the current iteration through the state update module, and finally generates a latent state containing the latest context information, providing a memory basis for the next iteration.

[0135] In one implementation, the latent state of the current iteration captures historical reasoning context and temporal relevance across iterations. This latent state can both aggregate long-range iterative information beyond immediate messages and enable the network to refine or revise its beliefs based on evolving residuals and posterior changes. This temporal memory is particularly beneficial in systems with cyclic factor graphs, such as MIMO.

[0136] In one implementation, the computer device can first obtain the input information for latent state calculation by concatenating and normalizing the latent state of the previous iteration, the refined token features of the current iteration, the log odds, the symbolic posterior statistic of the symbolic posterior probability, and the residual change. This input information for latent state calculation can be denoted as: (19) in, Input information for calculating the potential state. These are the characteristics of the refined token. It is a logarithmic probability. and Let be the symbolic posterior statistic of the symbolic posterior probability. This represents the change in residuals. This represents the potential state from the previous iteration. Representation layer normalization.

[0137] Among them, residual change The calculation formula can be written as: (20) in, This represents the projection residual for the current iteration round. This is the projected residual from the previous iteration. The change in this residual... It captures the instantaneous evolution of reconstruction errors, providing dynamic clues for state updates.

[0138] In one implementation, the computer device can feed the input information for calculating the latent state into a two-layer MLP with a hidden layer size of 2. And use the GELU activation function to generate . Updated potential state This is used to construct the posterior feature input in the next iteration, thus forming a closed-loop refining cycle.

[0139] In this example, by using refined token features to calculate symbol posterior probabilities and latent states, redundant information injection is effectively avoided. This not only reduces the computational complexity during the iteration process but also enhances the expressive power of the features, improving the accuracy and effectiveness of symbol posterior probabilities and latent states.

[0140] In one example, the specific process of refining the token features in step 221 above includes: Step 2211: Process the token features of each variable node to obtain symbol features.

[0141] For example, the computer device can first process each token feature through linear embedding and normalization to obtain symbol features.

[0142] In one implementation, the process can be as follows: (twenty one) in, It is a symbolic feature. Representation layer normalization. It is a learnable embedding projection matrix. It is a token characteristic.

[0143] Step 2212: Combine all variable node symbol features into a feature sequence.

[0144] For example, the computer device arranges the symbolic features of each variable node into a sequence structure according to preset rules to obtain a feature sequence.

[0145] In one implementation, the computer device can display all symbolic features. Stacked into a sequence .

[0146] Step 2213: Process the feature sequence using a multi-head attention mechanism to obtain attention features.

[0147] For example, a computer device processes feature sequences through a multi-head attention mechanism to capture long-range dependencies between feature sequences from different subspaces, thereby obtaining attention features containing multi-dimensional correlation information.

[0148] In one implementation, the multi-head attention mechanism is a parallelized self-attention computation framework that allows the model to simultaneously focus on different feature dimensions of the sequence.

[0149] In one implementation, the attention feature is an enhanced representation that integrates the results of multi-head analysis.

[0150] In one implementation, the process may include: (twenty two) in, This indicates attentional characteristics. Representation layer normalization. The symbolic features of all variable nodes form a feature sequence. This represents the multi-head attention mechanism layer.

[0151] In one implementation, to incorporate prior structural information, an attention bias matrix can be added during the attention calculation process. The attention bias matrix The coupling strength between symbols is reflected by the inner product of the channel tensors corresponding to the transmitted symbols. Attention bias matrix. The The element is used to adjust the first element. The and the first Attention bias between symbols. The formula for calculation is as follows: (twenty three) in, Attention bias matrix The The element is used to adjust the first element. The and the first Attention bias between symbols. It is a learnable scaling factor. Channel matrix The first in Each element is a conjugate channel coefficient. Channel matrix The first in Each element. The denominator represents all pairs of signs. Similar to product summation, and normalized to . This refers to the number of users per antenna. This refers to the number of antennas at the base station.

[0152] In one implementation, the bias matrix is ​​shared among all attention heads. The setting of this bias matrix encourages assigning higher attention weights to symbol pairs with strong mutual interference. Furthermore, the bias matrix remains robust to different channel implementations by normalizing the average channel interaction.

[0153] Step 2214: Process the attention features using a feedforward network to obtain refined features. The refined features contain the refined token features corresponding to each variable node.

[0154] For example, the computer device inputs attention features into a feedforward network consisting of linear transformations and nonlinear activations. The nonlinear transformations further refine the feature abstraction levels, ultimately outputting refined features corresponding to each variable node. These refined features are the refined token features.

[0155] In one implementation, the feedforward network is a stacked network containing at least one layer of fully connected and nonlinear activation.

[0156] In one implementation, the refined feature is a token feature resulting from multiple rounds of nonlinear transformation. The dimension of this refined feature is consistent with the number of symbols in the constellation.

[0157] In one implementation, the feedforward network uses a two-layer fully connected network with GELU activation. The first layer shifts the input dimension from... Extending to hidden dimension 4 The GELU activation function is then applied. The second layer then projects it back. .

[0158] In one implementation, the process includes: (twenty four) in, These are the characteristics of the refined token. . Representation layer normalization. This indicates attentional characteristics. This indicates a feedforward network layer.

[0159] In this example, token features are optimized by projecting them onto the token features and processing them using a multi-head attention mechanism and a feedforward network. This reduces the amount of data used for token features and improves their effectiveness.

[0160] In one example, in step 223 above, to reduce computational complexity, this application employs a simplified message passing mechanism for all... Sharing the same mean-variance pair Instead of calculating independent messages for each factor node, this means-variance is the symbolic posterior statistic of the symbolic posterior probability. Its calculation process may include: Step 2231: Calculate the product of the vector of symbol posterior probability and the transpose of the point vector of the modulated signal symbol to obtain the mean in the symbol posterior statistic.

[0161] For example, the computer device first transposes and aligns the vector of symbolic posterior probabilities with the point vector of the modulated signal symbols. Then, the computer device performs vector multiplication to calculate the mean of the symbolic posterior statistic.

[0162] In one implementation, the symbol posterior probability tensor is a multidimensional array containing the confidence distributions of each symbol. The confidence of each symbol is its probability value.

[0163] In one implementation, the point vectors of the modulated signal symbols are deterministic vectors corresponding to the symbol set. This symbol set is the constellation.

[0164] In one implementation, the mean of the symbolic posterior statistic is a specific numerical value.

[0165] In one implementation, the formula for calculating the mean of the posterior statistic of the sign can be written as: (25) in, Indicates the number of iterations in the current iteration round. The mean of the signed posterior statistics for each variable node. Indicates the number of iterations in the current iteration round. The posterior probability of each symbol for each variable node. This includes the confidence level of the i-th variable node corresponding to each symbol in the set. The transpose of the point vector representing the modulated signal symbol. .

[0166] Step 2232: Take the transpose of the square of the point vector of the modulated signal symbol as the first vector.

[0167] For example, the first vector can be denoted as .

[0168] Step 2233: Calculate the product of the vector of the symbolic posterior probability and the first vector to obtain the second vector.

[0169] For example, the second vector can be denoted as .

[0170] Step 2234: Use the squared modulus of the mean in the symbolic posterior statistic as the third vector.

[0171] For example, the third vector can be denoted as .

[0172] Step 2235: Calculate the difference between the second and third tensors to obtain the variance in the symbolic posterior statistic.

[0173] For example, the computer device performs element-wise subtraction between the second and third tensors to generate the variance tensor in the symbolic posterior statistic. This variance characterizes the dispersion of the symbol values ​​and reflects the combined effects of channel noise and estimation uncertainty.

[0174] In one implementation, the calculation formula can be: (26) In this example, by calculating the mean and variance of the symbolic posterior statistic of the symbolic posterior probability, the computational efficiency of the message passing mechanism is simplified by utilizing this symbolic posterior statistic.

[0175] Figure 2 This application provides an iterative process diagram, in which... Figure 1 Based on the illustrated embodiments, as Figure 2 As shown, the neural detection framework of AttnMPNet in this application includes... A fixed number of iterative steps.

[0176] In each iteration step, the received signal in the MIMO scenario is first acquired. Channel matrix and noise variance Subsequently, the computer equipment based on the received signal Channel matrix and noise variance The algorithm performs a Gaussian approximation-based message passing (MP) step to extract the message passing posterior statistics for each variable node and generate token features. Subsequently, it models the dependencies between symbols in the token features using a self-attention mechanism and refines the symbol features. Finally, it uses the refined token features to predict the symbol posterior probabilities of the variable nodes and updates the latent states used to capture inter-iteration temporal information.

[0177] Figure 3 This application provides an iterative process diagram, in which... Figure 1 and Figure 2 Based on the illustrated embodiments, as Figure 3 As shown, in each iteration of the AttnMPNet neural detection framework of this application, the computer device can receive signals... As Factor nodes .Will One transmitted signal as Variable nodes Construct a factor graph. Based on this factor graph, the log odds can be obtained through attention calculation. And based on the logarithmic probability Determine the posterior probability of the sign Furthermore, based on this logarithmic probability... The input information of the potential state can be updated. And finally calculate the potential state. .

[0178] In the attention calculation process, the computer device first generates symbol features based on the token features. Then, the symbolic features are input into MHA and FFN for processing to obtain the refined token features. .

[0179] exist Figures 1 to 3 Based on the illustrated embodiment, the computational complexity calculation process of AttnMPNet in this application may include: First, the complexity of an initial MMSE step is O(n). This expense is in The complexity is amortized across iterations. Each iteration includes Gaussian message passing, a self-attention mechanism, and multiple lightweight MLPs. The complexity of Gaussian message passing is O(log n). The complexity of the self-attention mechanism is... Multiple lightweight MLPs can include a Transformer FFN, a posterior predictor, and a latent state updater. The total complexity of these multiple lightweight MLPs is O(n log n). .

[0180] Therefore, in the AttnMPNet of this application, the total complexity of a single iteration is... This is comparable in complexity to RE-MIMO, which also employs attention and FFN modules. However, RE-MIMO embeds the complete residual tensor and channel matrix columns into each symbol token, resulting in a significant increase in token dimension and a heavier embedding layer.

[0181] In contrast, the AttnMPNet proposed in this application employs a compact design, avoiding global redundancy, thus allowing the use of smaller embedding dimensions without sacrificing modeling capabilities. This results in lower memory and computational costs, with a particularly significant advantage in large-scale MIMO scenarios, as the token dimension in this application remains almost constant with increasing antenna size. In terms of practical "wall-clock" runtime, AttnMPNet's Transformer-like design and matrix-based operations make it very GPU-friendly.

[0182] exist Figures 1 to 3 Based on the illustrated embodiment, a comprehensive simulation experiment was conducted on the execution process of the MIMO detector using the AttnMPNet network of this application.

[0183] All simulations employ the equivalent real-valued representation of complex MIMO systems and are performed under QAM modulation schemes. Signal-to-noise ratio (SNR) is defined as... The AttnMPNet model proposed in this application is implemented based on the PyTorch framework. In all experiments, the number of iterations was set. Model Dimension Attention count and potential state dimension In each epoch, 100,000 training samples and 5,000 validation samples are randomly generated. The network is trained for 500 epochs using the Adam optimizer with an initial learning rate of 0.0001 and a mini-batch size of 128. The ReduceLROnPlateau scheduler is used to adaptively adjust the learning rate. The training loss is the average cross-entropy between the predicted log odds and the one-hot target over T iterations. (twenty three) Benchmarking was performed on various neural network-based MIMO detectors, including AttnMPNet, Orthogonal Approximate Message Passing Network (OAMPNet), GEPNet, RE-MIMO, and AMP-GNN. For comprehensive comparison, the classic MMSE and AMP were also included as baselines. The ML boundary was evaluated using spherical decoding. All neural benchmark models were implemented using their official source code. All detectors were uniformly iterated to 10 iterations, consistent with the original settings of most baseline models. The average complex domain symbol error rate (SER) was reported for each signal-to-noise ratio point by accumulating detection results until 1,000 symbol errors were collected.

[0184] The comparison results can be as follows Figure 4 As shown, Figure 4 (a) Data for 8×16-MIMO system and 64-QAM modulation scenario. Figure 4 (b) Data for 32×64-MIMO system and 256-QAM modulation scenario. Figure 4 (c) Data for 48×64-MIMO system and 16-QAM modulation scenarios. Eight methods were used in each test: MMSE, AMP, OAMPNet, RE-MIMO, AMP-GNN, GEPNet, AttnMPNet, and ML. With SNR on the horizontal axis and SER on the vertical axis, the results of these eight methods in the three scenarios are shown below. Figure 4As shown, under various MIMO configurations, the AttnMPNet proposed in this application consistently outperforms MMSE, AMP, and all neural baselines, including OAMPNet, RE-MIMO, AMP-GNN, and GEPNet.

[0185] Specifically, compared to RE-MIMO, which also employs an attention mechanism, AttnMPNet achieves significant performance gains across all signal-to-noise ratios. Particularly in complex scenarios involving large-scale MIMO and high-order modulation, where RE-MIMO suffers from an error plateau problem, our application demonstrates superior performance. This indicates that our compact feature design has a significant advantage in enhancing model expressive power. Notably, AttnMPNet's performance is very close to the theoretical maximum likelihood (ML) detection boundary in all scenarios, fully demonstrating its strong robustness in large-scale MIMO detection and its effectiveness across different modulation orders.

[0186] Modern communication systems typically require parallel detection, making it possible to evaluate runtime performance at different batch sizes, providing practical insights into detector scalability. As shown in Table 1, AttnMPNet maintains low and stable inference latency across all batch sizes. This is primarily due to its Transformer-like architecture and highly parallelized matrix operations. Compared to GNN detectors that rely on irregular graph traversal and matrix inversion (such as GEPNet and AMP-GNN), AttnMPNet exhibits superior scalability. Notably, while RE-MIMO latency is similar to that of this application at smaller batch sizes, its model width... Larger batch sizes result in significantly increased runtime. In contrast, the AttnMPNet proposed in this application employs a compact feature representation that maintains a smaller token dimension even with increased antenna size, making it more suitable for deployment in wireless systems requiring large-scale, high-throughput architectures. Table 1 shows the inference runtime of the neural detector at different batch sizes. The unit is seconds per batch.

[0187] Table 1 Figure 5 A structural diagram of a MIMO system optimization device provided in this application embodiment is shown below. Figure 5 As shown, the MIMO system optimization device 500 includes: The acquisition module 501 is used to acquire the modulation constellation set of the MIMO system and the received signals of each antenna of the base station, and to construct the factor graph of the MIMO system. The calculation module 502 is used to calculate the symbol posterior probability of each user based on the factor graph iterative calculation; the symbol posterior probability includes the confidence of each symbol in the modulation constellation set; the symbol posterior probability is calculated based on the token feature in any iteration round of the iterative calculation; the token feature is determined according to the potential state of the previous iteration round, the message passing posterior statistic of the current round, and the projection residual obtained by projecting the residual of the observation value and the estimate of the current round onto the specified channel; The optimization module 503 is used to select the symbol with the highest confidence among the symbol posterior probabilities as the estimated transmission symbol of the user; the estimated transmission symbol is used to generate the transmission signal of the user.

[0188] In one example, the calculation module 502 is used for: In any iteration of the iterative calculation, token features are constructed based on the latent state of each user in the previous iteration, the posterior statistics of message passing in the current iteration, and the projection residual; and based on the refined token features, the symbolic posterior probability and latent state of each user are calculated.

[0189] In one example, the calculation module 502 is used for: The token features are processed using an attention mechanism and a feedforward network to obtain refined token features; The refined token features, the latent state of the previous iteration, the message passing posterior statistics of the current iteration, and the projection residual are input into the predictor to predict the log odds and symbol posterior probabilities of each variable node in the current iteration; the log odds include the probability parameters of each symbol in the modulation constellation set. The latent state of the current iteration is calculated based on the latent state of the previous iteration, the refined token features of the current iteration, the log odds, the symbolic posterior statistic of the symbolic posterior probability, and the residual change.

[0190] In one example, the calculation module 502 is used for: The token features of each variable node are processed to obtain the symbol features; The symbolic features of all variable nodes are combined into a feature sequence; The feature sequence is processed using a multi-head attention mechanism to obtain attention features; The attention features are processed using a feedforward network to obtain refined features; the refined features include the refined token features corresponding to each variable node.

[0191] In one example, the calculation module 502 is used for: The mean of the symbol posterior statistic is obtained by multiplying the vector of the symbol posterior probability with the transpose of the point vector of the modulated signal symbol. The transpose of the square of the point vector of the modulated signal symbol is used as the first vector; The second vector is obtained by multiplying the vector of the symbolic posterior probability with the first vector; The squared modulus of the mean in the symbolic posterior statistic is used as the third vector; Calculate the difference between the second vector and the third vector to obtain the variance in the symbolic posterior statistic.

[0192] In one example, the calculation module 502 is used for: Based on the channel matrix and the symbol posterior statistic of the symbol posterior probability described in the previous iteration, the mean and variance of the Gaussian distribution are calculated. Based on the channel matrix and the variance of the Gaussian distribution, the variance of the posterior statistic of message passing for each variable node in the current round is calculated. The mean of the message passing posterior statistic is calculated based on the channel matrix, the mean and variance of the Gaussian distribution, and the variance of the message passing posterior statistic.

[0193] In one example, the calculation module 502 is used for: The residual is obtained by calculating the difference between the observed value and the estimated value; Based on the transpose of the channel matrix, calculate the projection of the residual in the symbol dimension; The projection is normalized to obtain the projection residual.

[0194] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0195] In this embodiment, the MIMO system optimization device is presented in the form of a functional unit. Here, a unit refers to an application-specific integrated circuit (ASIC), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0196] Figure 6 A structural diagram of a computer device provided in an embodiment of this application, such as... Figure 6As shown, the computer device 600 includes one or more processors 601, memory 602, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 6 Take the 601 processor as an example.

[0197] Processor 601 may be a central processing unit, a network processor, or a combination thereof. Processor 601 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0198] The memory 602 stores instructions executable by at least one processor 601 to cause the at least one processor 601 to perform the method shown in the above embodiments.

[0199] Memory 602 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, memory 602 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, memory 602 may optionally include memory remotely located relative to processor 601, and this remote memory may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0200] Memory 602 may include volatile memory, such as random access memory; memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; memory 602 may also include combinations of the above types of memory.

[0201] The computer device also includes a communication interface 603 for communicating with other devices or communication networks.

[0202] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0203] This application provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method of any embodiment of this application.

[0204] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for optimizing a MIMO system, characterized in that, The method includes: Obtain the modulation constellation set of the MIMO system and the received signals of each antenna of the base station, and construct the factor graph of the MIMO system; The symbol posterior probability of each user is calculated iteratively based on the factor graph; the symbol posterior probability includes the confidence of each symbol in the modulation constellation set; the symbol posterior probability is calculated based on the token feature in any iteration round of the iterative calculation; the token feature is determined according to the latent state of the previous iteration round, the message passing posterior statistic of the current round, and the projection residual obtained by projecting the residual of the observation and the estimate of the current round onto a specified channel; The symbol with the highest confidence among the symbol posterior probabilities is selected as the estimated transmission symbol for the user; the estimated transmission symbol is used to generate the user's transmission signal.

2. The method according to claim 1, characterized in that, The symbolic posterior probability of each user is calculated iteratively based on the factor graph, including: In any iteration of the iterative calculation, token features are constructed based on the latent state of each user in the previous iteration, the posterior statistics of message passing in the current iteration, and the projection residual; and based on the refined token features, the symbolic posterior probability and latent state of each user are calculated.

3. The method according to claim 2, characterized in that, Based on the refined token features, the symbolic posterior probability and latent state of each user are calculated, including: The token features are processed using an attention mechanism and a feedforward network to obtain refined token features; The refined token features, the latent state of the previous iteration, the message passing posterior statistics of the current iteration, and the projection residual are input into the predictor to predict the log odds and symbol posterior probabilities of each variable node in the current iteration; the log odds include the probability parameters of each symbol in the modulation constellation set. The latent state of the current iteration is calculated based on the latent state of the previous iteration, the refined token features of the current iteration, the log odds, the symbolic posterior statistic of the symbolic posterior probability, and the residual change.

4. The method according to claim 3, characterized in that, The token features are processed using an attention mechanism and a feedforward network to obtain refined token features, including: The token features of each variable node are processed to obtain the symbol features; The symbolic features of all the variable nodes are combined into a feature sequence; The feature sequence is processed using a multi-head attention mechanism to obtain attention features; The attention features are processed using a feedforward network to obtain refined features; the refined features include the refined token features corresponding to each variable node.

5. The method according to claim 3, characterized in that, The calculation of the symbolic posterior statistic of the symbolic posterior probability includes: The mean of the symbol posterior statistic is obtained by multiplying the vector of the symbol posterior probability with the transpose of the point vector of the modulated signal symbol. The transpose of the square of the point vector of the modulated signal symbol is used as the first vector; The second vector is obtained by multiplying the vector of the symbolic posterior probability with the first vector; The squared modulus of the mean in the symbolic posterior statistic is used as the third vector; Calculate the difference between the second vector and the third vector to obtain the variance in the symbolic posterior statistic.

6. The method according to any one of claims 1-5, characterized in that, The calculation process of the posterior statistic of message passing in the current round includes: Based on the channel matrix of the MIMO system and the symbol posterior statistic of the symbol posterior probability in the previous iteration, the mean and variance of the Gaussian distribution are calculated. Based on the channel matrix and the variance of the Gaussian distribution, the variance of the posterior statistic of message passing for each variable node in the current round is calculated. The mean of the message passing posterior statistic is calculated based on the channel matrix, the mean and variance of the Gaussian distribution, and the variance of the message passing posterior statistic.

7. The method according to any one of claims 1-5, characterized in that, The projection residual, obtained by projecting the residuals of the observations and the estimates of the current round onto a specified channel, includes: The residual is obtained by calculating the difference between the observed value and the estimated value; Based on the transpose of the channel matrix of the MIMO system, the projection of the residual in the symbol dimension is calculated; The projection is normalized to obtain the projection residual.

8. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer instructions for causing a computer to perform the method of any one of claims 1 to 7.

Images