A large-scale MIMO precoding parameter generation method based on channel feature mapping

Abstract

This invention discloses a method for generating precoding parameters for large-scale MIMO based on channel feature mapping. Addressing the problem of high computational complexity and inability to meet real-time requirements of traditional iterative algorithms in large-scale MIMO precoding, this invention models large-scale MIMO precoding as a learning task from channel statistical features and environmental signal-to-noise ratio (SNR) to statistical parameters. Subsequently, an MPPM model based on a convolutional neural network is constructed. By mining the nonlinear mapping relationship between the input features (composed of channel statistical features and SNR) and statistical parameters, an intermediate set of statistical parameters is generated. Finally, the statistical parameters are substituted into the precoding generation formula, and the final precoding matrix is ​​obtained through a single reconstruction and then normalized for power. This invention replaces the cumbersome mathematical iteration process with a deep learning-based parameter generation mechanism, significantly reducing the dimensionality and computational complexity of precoding solutions while maintaining system performance.

Description

Technical Field

[0001] This invention relates to precoding techniques in massive MIMO communication systems, specifically to a method for generating massive MIMO precoding parameters based on channel feature mapping, belonging to the field of wireless communication technology. Background Technology

[0002] With the gradual commercialization of 5G mobile communication technology, the research focus in the communications field has shifted to 6G mobile communication systems, aiming to achieve ultra-high data transmission rates, ultra-dense connections, and deep coverage for the Internet of Things. Massive MIMO and its evolution technologies, as the core support of 6G, utilize spatial degrees of freedom to serve hundreds of users on the same time-frequency resources by configuring ultra-large-scale antenna arrays at the base station, greatly improving system capacity and spectral efficiency. In multi-user MIMO scenarios, precoding technology is key to suppressing co-channel interference and improving system performance. It spatially weights the signal at the transmitting end, enabling the beam to accurately target the user and eliminate inter-user interference. However, with the increasing scale of antenna dimensions and the surge in connection density in 6G scenarios, traditional precoding methods face severe challenges. On the one hand, the high-dimensional matrix inversion and iterative processes of iterative algorithms such as weighted least mean square error introduce enormous computational complexity, making it difficult to meet the ultra-low latency requirements of 6G. On the other hand, simple algorithms with low complexity, such as zero-forcing and maximum ratio transmission precoding algorithms, face performance bottlenecks and cannot meet the high-speed transmission requirements of large amounts of data. Therefore, how to reduce computational complexity while ensuring interference suppression performance has become the core challenge of current precoding technology. Summary of the Invention

[0003] Purpose of the Invention: In large-scale MIMO transmission scenarios, the computational complexity increases significantly with the increase in antenna dimension, and the real-time performance is poor due to the reliance on multiple optimization iterations in traditional algorithms. The purpose of this invention is to provide a method for generating large-scale MIMO precoding parameters based on channel feature mapping. By constructing a neural network model (MPPM) to generate key statistical parameters, the method learns the channel energy distribution characteristics and the correlation between the signal-to-noise ratio and the optimal precoding statistical parameters. The precoding matrix is ​​reconstructed through a single calculation, thereby significantly reducing computational complexity and processing latency while effectively suppressing interference and improving the overall system performance.

[0004] Technical Solution: To achieve the above objectives, this invention provides a method for generating large-scale MIMO precoding parameters based on channel feature mapping, comprising:

[0005] Large-scale MIMO precoding is modeled as a learning task from channel statistical features and environmental signal-to-noise ratio to statistical parameters. The input is the channel energy distribution matrix and the environmental signal-to-noise ratio, and the output is the statistical parameters required for precoding, including the intermediate parameter matrices required for the fixed-point iterative precoding and the statistical parameters of the constraint dual variables.

[0006] The MPPM (Multi-Parameter Precoding Model) network, built based on a convolutional neural network, achieves end-to-end mapping from the channel to the precoding parameters;

[0007] The MPPM network includes a data preprocessing module, a feature extraction module, and a precoding recovery module. The data preprocessing module performs statistical analysis on a set of continuously sampled channel matrices over a period of time, calculates the energy distribution characteristics of the channel in the beam domain, and constructs the corresponding energy distribution matrix. The feature extraction module performs feature learning on the energy distribution matrix, including a convolutional neural network backbone and a signal-to-noise ratio (SNR) embedding module. The convolutional neural network backbone extracts spatially relevant features from the energy distribution matrix, and the SNR embedding module maps and encodes environmental SNR parameters and fuses them with the output features of the backbone network. The precoding recovery module restores the output parameters to their corresponding matrix form, obtains the statistical parameters required for precoding, substitutes these statistical parameters into the precoding matrix generation formula for a single calculation, and performs power normalization to obtain the final precoding matrix.

[0008] Furthermore, the MPPM network is trained using a supervised learning method. Each training sample contains a set of energy distribution matrices, signal-to-noise ratio (SNR) parameters, and their corresponding target precoding statistical parameters. The target precoding statistical parameters are obtained by a fixed-point iterative algorithm based on statistical parameters. This algorithm generates statistical parameters by iteratively calculating multiple sets of channel samples within a preset time period and statistically processing the resulting intermediate parameter matrix. During training, the energy distribution matrix and SNR are used as network inputs, and the corresponding precoding statistical parameters are used as supervision labels. The network parameters are backpropagated and iteratively updated by minimizing the error between the output and the labels. The loss function is the mean squared error function.

[0009] Furthermore, the precoding fixed-point iterative formula is determined according to the following method:

[0010] Under the constraint of total power, a precoding optimization problem is constructed with the objective of maximizing the weighted sum rate. The constrained optimization problem is transformed into a Lagrangian form by introducing dual variables. Based on the KKT optimality condition of the optimization problem, the fixed-point update structure of the precoding matrix is ​​derived, so that the precoding matrix of each user at each base station is represented as the product of matrix inverse operation and linear combination terms.

[0011] During the fixed-point iteration process, the following intermediate parameter matrix and dual variables are constructed:

[0012] The first intermediate parameter matrix is ​​formed by the difference between the inverse matrix of the signal covariance matrix excluding the target user signal terms and the inverse matrix of the signal covariance matrix including the target user signal terms.

[0013] The second intermediate parameter matrix is ​​formed by the product of the inverse of the signal covariance matrix excluding the target user signal terms and the sum of the effective channel vectors of the target user group;

[0014] The third intermediate parameter matrix is ​​composed of the product of the first intermediate parameter matrix and the sum of the effective channel vectors of the remaining users after excluding the current target user;

[0015] The dual variable, corresponding to the Lagrange multiplier under the transmit power constraint, is used to adjust the magnitude of the precoding matrix constructed from the intermediate parameter matrix so that it meets the preset total system power limit.

[0016] The signal covariance matrix is ​​constructed by multiplying the precoding matrix of all users and the channel matrix, and includes a noise term and a power normalization term.

[0017] Furthermore, the method for determining the precoding statistical parameters is as follows: multiple sets of historical channel samples collected within a preset time period are obtained, and the corresponding instantaneous parameters are calculated using a fixed-point iterative method for each set of instantaneous channel samples; finally, the multiple sets of instantaneous parameters are statistically averaged to obtain statistical parameters, which are then used as supervision labels for neural network training.

[0018] Furthermore, during the training phase, the signal-to-noise ratio (SNR) embedding module maps the one-dimensional SNR parameter to a vector of a preset dimension through a multilayer perceptron and concatenates and fuses it with the convolutional features output by the backbone of the convolutional neural network. During the execution phase, it uses the actual environmental SNR as input for mapping to achieve feature enhancement. Finally, the fused data is output as parameters through a multilayer perceptron.

[0019] Furthermore, the backbone of the convolutional neural network consists of multiple layers of convolutional units connected in series. Except for the first layer, which is the output of the data preprocessing module, the other layers use the output of the previous layer as input to extract multi-scale spatial features from the energy distribution matrix.

[0020] Furthermore, during the online execution phase, the base station receives the current channel matrix and transmits it to the network controller. The network controller inputs the channel matrix and the current user's environmental signal-to-noise ratio data into the MPPM network. The MPPM network generates the statistical parameters based on the pre-trained weight parameters and performs a single reconstruction calculation using the precoding recovery module to obtain the final precoding matrix. Finally, the generated final precoding matrix is ​​transmitted to the base station, and the final precoding matrix is ​​used to precode the downlink signal.

[0021] Furthermore, during the MPPM network training phase, each base station transmits the obtained channel matrix to the network controller. The network controller performs fixed-point iterative solutions on multiple different channel matrices to obtain the corresponding target precoding statistical parameters. The statistical parameters are then paired with the corresponding channel energy distribution matrix and signal-to-noise ratio parameters to form training samples for supervised learning training of the MPPM network.

[0022] The present invention also provides a computer system, including a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor. When the computer program / instructions are executed by the processor, they implement the steps of the large-scale MIMO precoding parameter generation method based on channel feature mapping.

[0023] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the large-scale MIMO precoding parameter generation method based on channel feature mapping.

[0024] Beneficial Effects: This invention proposes a large-scale MIMO precoding parameter generation method based on channel feature mapping. By extracting structured features from the original channel matrix, it maps it to a physically meaningful channel energy distribution matrix, thereby reducing the input dimensionality while enhancing the ability to express key information. Based on this, an MPPM model is constructed, representing the precoding matrix as a function of several intermediate parameters through a parameterized structure. This transforms the multivariate update process involved in traditional iterative optimization into a direct parameter generation process, avoiding repeated iterative iterations. Furthermore, the model adaptively adjusts the precoding parameters using signal-to-noise ratio (SNR) information, enabling it to adapt to different channel conditions and achieving a single computation from the channel to the precoding matrix. Compared to the classic weighted least mean square error (WMS) method, this invention can recover the precoding matrix without multiple rounds of iterative optimization. While maintaining similar system performance, it significantly reduces computational complexity and latency, thereby improving system real-time performance and engineering implementation efficiency. Attached Figure Description

[0025] Figure 1This diagram illustrates a method for generating large-scale MIMO precoding parameters based on channel feature mapping, as described in an embodiment of the present invention.

[0026] Figure 2 This is a diagram of the MPPM network architecture in an embodiment of the present invention.

[0027] Figure 3 The diagram shows the algorithm and rate performance of the simulation experiment in this embodiment of the invention. Detailed Implementation

[0028] The technical solution and effects of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown in the illustration, this invention discloses a method for generating large-scale MIMO precoding parameters based on channel feature mapping. Large-scale MIMO precoding is modeled as a learning task from channel statistical features and environmental signal-to-noise ratio (SNR) to statistical parameters. The input data is the channel's energy distribution matrix, and the output sequence consists of the statistical parameters required for precoding. A data-driven strategy is employed, using an MPPM network constructed based on a convolutional neural network, combined with supervised learning to achieve an end-to-end mapping from channel features to precoding parameters. The statistical parameters required for precoding include the intermediate parameter matrices required for the fixed-point iterative precoding and the statistical parameters of the constraint dual variables.

[0030] like Figure 2 As shown, the MPPM network includes a data preprocessing module, a feature extraction module, and a precoding recovery module. The data preprocessing module performs statistical analysis on a set of continuously sampled channel matrices over a period of time, calculates the energy distribution characteristics of the channel in the beam domain, and constructs the corresponding energy distribution matrix. The feature extraction module performs deep feature learning on the energy distribution matrix. This module includes a convolutional neural network backbone and a signal-to-noise ratio (SNR) embedding module. The convolutional neural network backbone extracts spatially relevant features from the energy distribution matrix, and the SNR embedding module maps and encodes the environmental SNR parameters and fuses them with the output features of the backbone network. The precoding recovery module restores the generated parameters to their corresponding matrix form, obtaining the statistical parameters required for precoding. Since these statistical parameters are iteratively converged values ​​obtained through network learning, they only need to be directly substituted into the precoding matrix generation formula for a single calculation, followed by power normalization to obtain the final precoding matrix.

[0031] Furthermore, the MPPM network is trained using a supervised learning method. Each training sample contains a set of energy distribution matrices, signal-to-noise ratio (SNR) parameters, and their corresponding target precoding statistical parameters. The target precoding statistical parameters are obtained by a fixed-point iterative algorithm based on statistical parameters. This algorithm generates statistical parameters by iteratively calculating multiple sets of channel samples within a preset time period and statistically processing the intermediate parameters. During training, the energy distribution matrix and SNR are used as network inputs, and the corresponding precoding statistical parameters are used as supervision labels. The network parameters are then updated through uppropagation and iterative updates by minimizing the error between the output and the labels. The loss function is the mean squared error function.

[0032] The detailed implementation process of the method of the present invention will be illustrated below with reference to a specific system model.

[0033] I. System Model

[0034] For downlink (DL) transmission scenarios in massively multi-input multiple-output (MIMO) systems, consider a scenario containing... Base stations (BS) and A wireless communication network for individual user equipment (UE). Let and These represent the index sets for base stations and users, respectively. (Definition) To serve users The set of base stations, and define For base station A collection of users receiving the service. Each base station is equipped with... Each user equipment has a single transmitting antenna. Root receiving antenna and support A data stream. Assume all base stations are ideally synchronized in time and frequency and interconnected via links.

[0035] This section employs a geometry-based multipath channel model, representing the instantaneous channel as a superposition of a finite number of physical propagation paths. The base station... To users Complex channel matrix It can be represented as:

[0036]

[0037] in, Represents the field of complex numbers. Represents the conjugate transpose of a matrix or vector. For the number of propagation paths, It is the first Complex gain of the path, and These represent the angle of arrival (AoA) and angle of departure (AoD) for the path, respectively. For a uniform linear array (ULA) deployed at both the receiver and transmitter, its corresponding array response vector... and They can be represented as follows:

[0038]

[0039] and

[0040]

[0041] in Transpose of a matrix or vector For carrier wavelength, and These represent the antenna element spacing at the receiving and transmitting ends, respectively.

[0042] In the beam domain channel, the statistical channel covariance matrix, i.e.

[0043]

[0044] The long-term spatial correlation caused by the scattering environment was captured. Since the correlation matrix is ​​a Hermitian positive definite matrix, its eigenvalue decomposition (EVD) yields:

[0045]

[0046] in and It is a unitary matrix containing spatial eigenvectors. and Characterize the average power distribution across these spatial feature modes. By projecting the channel onto these feature spaces, the channel can be represented in the beam domain as:

[0047]

[0048] in This is the beam domain channel matrix. To characterize the power distribution of the channel, the beam domain channel matrix and energy distribution matrix are defined. :

[0049] in, Represents the real number field. Represents the conjugate of a matrix or vector. This represents the Hadamard product, which is the operation of multiplying corresponding elements of two matrices of the same dimension.

[0050] In downlink transmission of a frequency division duplex (FDD) massive MIMO system, in order to suppress inter-user interference, the base station will transmit signals... Multiply by the precoding matrix At this time, the base station Transmitted signal vector It can be modeled as:

[0051]

[0052] in For the precoding matrix, To meet The vector of symbols to be sent. User Received signal Represented as:

[0053]

[0054] The received signal contains the user's desired signal, intra-cell interference, inter-cell interference, and additive white Gaussian noise. Based on the above signal model, and treating multi-user interference as equivalent additive noise, the user... Interference plus noise covariance matrix at the location for:

[0055]

[0056] Based on this, the user The achievable rate is expressed as:

[0057]

[0058] II. Problem Statement

[0059] In massive MIMO systems, precoding aims to suppress inter-user interference and improve spectral efficiency. This embodiment uses a fixed-point iterative precoding algorithm, specifically determined as follows: First, a precoding optimization problem is constructed under total power constraints, with the objective of maximizing the weighted sum rate. This problem is then transformed into a Lagrangian form by introducing dual variables. Next, based on the KKT optimality conditions of the optimization problem, a fixed-point update structure for the precoding matrix is ​​derived, ensuring that the precoding matrix for each user at each base station is represented as the product of matrix inverse operations and linear combination terms.

[0060] During the fixed-point iteration process, the following intermediate parameter matrices and dual variables are constructed: The first intermediate parameter matrix is ​​the difference between the inverse of the signal covariance matrix excluding the target user's signal term and the inverse of the signal covariance matrix containing the target user's signal term; the second intermediate parameter matrix is ​​the product of the inverse of the signal covariance matrix excluding the target user's signal term and the sum of the effective channel vectors of the target user group; the third intermediate parameter matrix is ​​the product of the first intermediate parameter matrix and the sum of the effective channel vectors of the remaining users after excluding the current target user; the dual variable corresponds to the Lagrange multiplier under the transmit power constraint condition, used to adjust the amplitude of the precoding matrix constructed from the intermediate parameter matrices to meet the preset system total power limit. The signal covariance matrix is ​​constructed by the product of the precoding matrices and channel matrices of all users, and includes a noise term and a power normalization term.

[0061] A more specific formula is as follows: In the objective of maximizing the weighted sum rate under a total power constraint, an equal power constraint is often introduced to simplify the problem. In this case, the precoding problem can be expressed as:

[0062]

[0063] in These are the weighting coefficients. This represents the maximum available power of base station k. Due to the high coupling of the interference term, this optimization problem is non-convex. The process of designing an efficient solution algorithm will be discussed later by analyzing its Karush–Kuhn–Tucker (KKT) conditions.

[0064] Since the constraints take the equality sign, this equal power constraint problem can be further transformed into an unconstrained problem:

[0065]

[0066] in Represents a set The number of elements contained in the expression is used to adjust the noise level so that the expression remains equivalent to the original problem. From the first-order optimality condition of problem (12), the optimal precoding vector has the following fixed-point iterative form:

[0067]

[0068] in , , Defined as the intermediate parameter matrix of the precoding fixed-point iterative formula, namely the first intermediate parameter matrix, the second intermediate parameter matrix, and the third intermediate parameter matrix. For the corresponding base station For ease of description, the dual variables of the power constraint are denoted as the parameter matrices of base station k and user i, respectively. ,in It only relates to user i. Depending on user i and base station k, they have the following form:

[0069]

[0070] As can be seen from the above, the first intermediate parameter matrix The signal covariance matrix excluding the target user signal term The inverse matrix and the signal covariance matrix containing the target user signal term The difference between the inverse matrices constitutes the second intermediate parameter matrix. The signal covariance matrix excluding the target user signal term The inverse matrix and the effective channel vector of the target user group The product of the sums constitutes the third intermediate parameter matrix. From the first intermediate parameter matrix The effective channel vectors of the remaining users after excluding the current target user The product of the sum and the product constitutes the signal covariance matrix; where the signal covariance matrix is ​​the signal covariance matrix. , The precoding matrix of all users With channel matrix It is constructed from the product of and includes a noise term. and power normalization term .

[0071] Equation (14) gives the analytical form of the optimal precoding, where the parameter set This is the theoretical output parameter of the current channel, and it is also the learning object of the network. Finally, only power normalization is needed to obtain the actual precoding matrix.

[0072] It should be noted that the above parameter set Depends on equivalent channel The equivalent channel is composed of the instantaneous channel matrix and the precoding matrix. In large-scale MIMO systems, due to the large number of antennas on the base station side, the gain and interference distribution characteristics of the equivalent channel are mainly dominated by the slow-varying statistical characteristics such as the channel spatial correlation structure and the large-scale fading factor. Based on this, this embodiment further proposes to perform statistical processing on multiple channel samples within a preset time period to extract a set of statistical parameters that can characterize the current channel scenario. By using these statistical parameters to replace the instantaneous parameters, this invention can significantly reduce the real-time computational overhead of the base station while maintaining the robustness of precoding performance. Specifically, the statistical parameters are calculated using equation (14) for the channel parameters within a certain period of time:

[0073]

[0074] Therefore, after obtaining the statistical parameter, substituting it into the original formula (14) allows for the recovery of high-performance precoding through one iteration. Finally, normalization satisfies the power requirements. For ease of description, this embodiment names this statistical parameter-based precoding method SP-FPI (Statistical Parameter-based Fixed Point Iteration). This method provides high-quality offline learning labels for the subsequent training of deep neural networks; secondly, it also serves as a benchmark algorithm for performance evaluation to measure the performance of subsequent models.

[0075] The large-scale MIMO precoding parameter generation method based on channel feature mapping in this embodiment differs from traditional calculation methods. It does not require complex iterative solutions but instead utilizes a neural network to find a mapping relationship. Given the input channel and signal-to-noise ratio, it outputs corresponding parameters, and these parameters are highly consistent with theoretical values. Specifically, each containing... Each base station is equipped with a neural network in a specific area. The input to this neural network is the energy distribution matrix of all users within the coverage area and the environmental signal-to-noise ratio, and the output is the set of statistical parameters corresponding to these users. , This represents all base station to user connections. Parameter set. Within this framework, our goal is to find an optimal mapping function that minimizes the output user parameters. With the corresponding optimal statistical parameters The differences between them. The above optimization problem can be expressed as:

[0076]

[0077] in This represents the set of energy distribution matrices for all users. Given the current environmental signal-to-noise ratio, For neural network mapping functions, These are network parameters. It is quantification and The loss function for the difference between them is the minimum mean square error loss function in this embodiment.

[0078] III. Algorithm Design

[0079] To address the computational complexity and real-time requirements of precoding design in large-scale MIMO systems, this embodiment proposes an MPPM network based on convolutional neural networks. This network aims to replace the traditional, cumbersome multi-step iterative process through end-to-end parameter learning. Figure 2 As shown, the MPPM network consists of three parts: a data preprocessing module, a feature extraction module, and a precoding recovery module. The data preprocessing module calculates the energy distribution matrix of the beam domain from the input channel data using the statistical channel covariance matrix and eigenvalue decomposition. The feature extraction module includes a neural network structure and a signal-to-noise ratio (SNR) embedding module. The neural network extracts spatial correlations from the energy distribution matrix, while the SNR embedding module maps the one-dimensional SNR data to high-dimensional data and incorporates it into the neural network output to adapt to varying SNR in real-world environments. Finally, parameters are generated using a multilayer perceptron. The precoding recovery module restores the generated parameters of the model to a matrix, directly substitutes them into the precoding matrix generation formula for a single calculation, and then performs power normalization to obtain the corresponding precoding matrix. The following sections will describe each module in detail.

[0080] 1. Data Preprocessing Module

[0081] This module receives data from the [received source]. The first base station Original channel matrix for each user To process it, first calculate its covariance matrix according to formula (4), that is... Subsequently, eigenvalue decomposition (EVD) is performed on the matrix using formula (5) to extract the eigenvector matrix representing the signal subspace. and the eigenvalue matrix characterizing the channel gain Finally, the beam domain channel energy distribution matrix for each user is calculated using equation (7). To all users in this community After unfolding and splicing, we obtain part of the input data for the feature extraction module. This matrix reflects the distribution intensity of signal energy across different spatial dimensions, while also reducing the input signal from a complex number to a real number. To provide effective prior data, label generation introduces a fixed-point iterative algorithm as shown in Equation (14) based on maximizing the sum and rate. The algorithm uses the original channel matrix... As input, the RZF method performs fixed-point iterative calculations using initial values ​​until the weighted sum rate converges, and then extracts the corresponding set of theoretical statistical parameters at that moment. As a supervisory label, the smaller the error between the network output and the theoretical parameters, the stronger the network's learning ability.

[0082] 2. Feature Extraction Module

[0083] The feature extraction module consists of a neural network structure and a signal-to-noise ratio embedding module, aiming to extract and map channel features to system environment parameters. First, the neural network structure consists of... The convolutional units are connected in series. Each convolutional unit extracts multi-scale spatial features from the energy distribution matrix through a convolutional layer, and then passes through a normalization layer and a nonlinear activation layer in sequence.

[0084] For the neural network structure, the input to its convolutional units comes from the output of the previous group. Then, a set of filters is used to perform convolution operations to extract the local spatial correlation in the channel energy distribution matrix. The calculation formula is as follows:

[0085]

[0086] in, For the first The weight matrix of the group unit convolution kernel, For the corresponding bias term, This represents the convolution operator.

[0087] To ensure the numerical stability of deep networks during training and accelerate convergence, the feature matrix output by the convolutional layer... The data is fed into a Layer Normalization (LN) layer. This layer standardizes the mean and variance of the feature tensor, specifically calculated using the following formula:

[0088]

[0089] in, and Each is the current The mean and variance, To prevent extremely small offsets where the denominator is zero, and With learnable scaling and translation parameters, this operation can effectively alleviate the gradient vanishing problem in deep networks.

[0090] For the normalized features Further nonlinear activation layers are needed to enhance the model's nonlinear mapping capability. This embodiment uses the Gaussian Error Linear Unit (GELU) as the activation function. Unlike traditional hard-truncation activation functions, GELU achieves a smoother nonlinear mapping by multiplying the neuron's input by the cumulative probability of its own distribution. The specific calculation formula is as follows:

[0091]

[0092] in, This is the cumulative distribution function of the standard normal distribution. By introducing the GELU activation function, the model can better capture subtle changes in the channel matrix while preserving the amplitude of input features, utilizing its smooth nonlinearity. After the above cyclic evolution of "convolution-normalization-activation", the final output dimension of the branch is global feature vector .

[0093] For the signal-to-noise ratio (SNR) embedding module, it vectorizes the scalar SNR parameter using a multilayer perceptron structure. Here, two layers are used. It's important to note that this example assumes all users' SNRs are under the same environmental SNR conditions. First, the input dimension is... The SNR parameters first enter the first layer of the mapping space, and after linear transformation and activation function processing, they are mapped to a dimension of . The calculation process for the intermediate feature vector is as follows:

[0094]

[0095] in, This is the mapping weight matrix for the SNR parameters of the first layer. This is the first layer bias vector. The activation function is a linear rectifier, and its specific expression is:

[0096] Through this mapping layer, the original signal-to-noise ratio is expanded to a dimension of . intermediate feature vector Subsequently, this intermediate feature vector enters the second-level mapping space, and its calculation process is as follows:

[0097]

[0098] in, This is the weight matrix for the second layer mapping. This is the bias vector for the second layer. This layer mapping further projects the features onto a plane with dimension [dimensional value missing]. In the high-dimensional space, the final signal-to-noise ratio feature vector is output. It achieves in-depth parameterized characterization of environmental noise levels.

[0099] For output and Furthermore, the final mapping from fused features to pre-encoding parameters is achieved through a concatenation operator and a multilayer perceptron structure. First, the module performs concatenation and fusion processing, reducing the dimension of the neural network branch outputs to [missing value]. Channel feature vector The dimension of the output of the signal-to-noise ratio embedding module is eigenvectors By concatenating the first and last parts, we obtain the dimension as follows: Composite feature vector :

[0100]

[0101] The composite feature vector Simultaneously, it aggregates the spatial distribution information of the physical layer channel with the system's current signal-to-noise ratio (SNR) parameters. Subsequently, this composite vector is fed into a fully connected layer for global correlation modeling, and a nonlinear mapping is performed using the learned weight matrix. The final output parameter sequence :

[0102]

[0103] in, This is the weight matrix of the fully connected layer. For bias terms, For the output layer activation function, This represents the total number of parameters to be generated corresponding to the precoding matrix. The fully connected layer achieves the final mapping from input features to precoding parameters by weighting the fused features. These output parameters are directly used as the original input to the subsequent precoding recovery module to reconstruct the precoding matrix.

[0104] 3. Pre-encoding recovery module

[0105] The precoding restoration module is responsible for mapping the parameter sequence output by the feature extraction module back to the precoding parameter matrix. For parameter reshaping, this module receives the dimension of the output from the fully connected layer. parameter sequence To recover the precoded matrix structure, the module performs a reshape operation, rearranging the one-dimensional parameter sequence according to a predetermined antenna dimension and stream number configuration, thereby generating the parameter matrix. The specific calculation process can be expressed as follows:

[0106] Furthermore, the module will generate the parameter matrix mentioned above. With the original channel matrix By combining these formulas and substituting them into the preset precoding generation formula, a single calculation can be performed to obtain the precoding matrix. The calculation process can be expressed as follows:

[0107]

[0108] in The calculation process of expression (14) is shown. Finally, in order to meet the transmit power limit of the communication system, the module... The normalization operation is performed using the following expression:

[0109] The final output precoding matrix Not only does it possess low computational complexity, but by combining the advantages of rapid generation and single mathematical iteration in deep learning, it can significantly improve the system's sum-rate performance while ensuring real-time performance. All the steps described above regarding the MPPM network are logically integrated into the network controller.

[0110] In practical applications, the network controller uses its internal processor to call relevant hardware resources to execute the aforementioned algorithm process to complete the parameter prediction and precoding update steps.

[0111] Specifically, during the MPPM network training phase, each base station transmits the obtained channel matrix to the network controller. The network controller performs fixed-point iterative solutions on multiple different channel matrices to obtain the corresponding target precoding statistical parameters. These statistical parameters are then paired with the corresponding channel energy distribution matrix and signal-to-noise ratio (SNR) parameters to form training samples for supervised learning training of the MPPM network. During the online execution phase, the base station receives the current channel matrix and transmits it to the network controller. The network controller inputs the channel matrix and the current user's environmental SNR data into the MPPM network. The MPPM network generates the statistical parameters based on pre-trained weight parameters and performs a single reconstruction calculation using the precoding recovery module to obtain the final precoding matrix. Finally, the generated final precoding matrix is ​​transmitted to the base station, and the final precoding matrix is ​​used to precode the downlink signal.

[0112] IV. Implementation Results

[0113] To enable those skilled in the art to better understand the present invention, the performance and computational complexity comparison of the large-scale MIMO precoding parameter generation method based on channel feature mapping in this embodiment are presented below under a specific system configuration. The simulation environment was generated by the QuaDRiGa platform, in which the base station is equipped with Antenna, equipped for each user Antenna. The training and testing sample sizes for the MPPM network were 1600 and 400, respectively. Each sample contained a set of energy distribution matrices, signal-to-noise ratio parameters, and their corresponding optimization statistical parameters, which were obtained using a fixed-point iterative algorithm. Furthermore, to better compare the performance of the MPPM network, SP-FPI, WMMSE (Weighted Minimum Mean Square Error), and RZF (Regularized Zero-Forcing) algorithms were introduced as baseline algorithms.

[0114] Figure 3 shows the number of base stations Number of transmitting antennas Number of receiving antennas and number of users Under the system configuration, the sum rate performance of different precoding algorithms was compared on the test set, with the WMMSE algorithm serving as the reference benchmark. The closer the MPPM curve is to its theoretical SP-FPI curve, the more accurate the fitting performance of the neural network. Within the full signal-to-noise ratio range, the MPPM algorithm proposed in this embodiment significantly outperforms the traditional RZF algorithm. When the SNR is low, the sum rate curve of MPPM is slightly lower than that of WMMSE and SP-FPI methods, achieving a sum rate that reaches 90% of the target result. As the SNR increases, although the performance gap between MPPM and WMMSE slightly increases, MPPM can achieve over 96% of the sum rate of WMMSE in high signal-to-noise ratio environments. Simulation results show that in large-scale MIMO systems, the MPPM-based multi-parameter generator network can effectively capture complex channel spatial characteristics, and its output performance is highly close to that of highly complex traditional iterative algorithms. Compared to RZF, SP-FPI, and WMMSE algorithms, MPPM significantly improves the spectral efficiency of the system while maintaining lower computational overhead.

[0115] Table 1 shows the runtime of each algorithm under different system configurations. The runtime of each algorithm increases with the number of users. In all system configuration scenarios, the MPPM precoding method proposed in this embodiment exhibits extremely high computational efficiency. Simulation results show that, while maintaining a high performance close to the theoretical upper limit of SP-FPI and approaching the WMMSE algorithm, this invention avoids the complex iterative optimization process through end-to-end parameter generation of a neural network, and its runtime is less than 0.5% of the WMMSE algorithm (20 iterations). Experimental results demonstrate that this algorithm significantly reduces computational overhead while maintaining high performance.

[0116] Table 1 Operation Schedule

[0117]

[0118] This invention also discloses a computer system, including a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor. When the computer program / instructions are executed by the processor, they implement the steps of the method for generating large-scale MIMO precoding parameters based on channel feature mapping.

[0119] This invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of a method for generating large-scale MIMO precoding parameters based on channel feature mapping. The program code for implementing the method of this invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of this invention to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

Claims

1. A method for generating precoding parameters for large-scale MIMO based on channel feature mapping, characterized in that, include: Large-scale MIMO precoding is modeled as a learning task from channel statistical features and environmental signal-to-noise ratio to statistical parameters. The input is the channel energy distribution matrix and the environmental signal-to-noise ratio, and the output is the statistical parameters required for precoding, including the intermediate parameter matrices required for the fixed-point iterative precoding and the statistical parameters of the constraint dual variables. An MPPM network based on a convolutional neural network is used to achieve end-to-end mapping from the channel to the precoding parameters. The MPPM network includes a data preprocessing module, a feature extraction module, and a precoding recovery module. The data preprocessing module performs statistical analysis on a set of continuously sampled channel matrices over a period of time, calculates the energy distribution characteristics of the channel in the beam domain, and constructs the corresponding energy distribution matrix. The feature extraction module performs feature learning on the energy distribution matrix, including a convolutional neural network backbone and a signal-to-noise ratio (SNR) embedding module. The convolutional neural network backbone extracts spatially relevant features from the energy distribution matrix, and the SNR embedding module maps and encodes environmental SNR parameters and fuses them with the output features of the backbone network. The precoding recovery module is used to restore the output parameters to the corresponding matrix form, obtain the corresponding statistical parameters required for precoding, substitute the statistical parameters into the precoding matrix generation formula for a single calculation, and perform power normalization to obtain the final precoding matrix.

2. The method for generating large-scale MIMO precoding parameters based on channel feature mapping according to claim 1, characterized in that, The MPPM network is trained using a supervised learning method. Each training sample contains a set of energy distribution matrices, signal-to-noise ratio (SNR) parameters, and their corresponding target precoding statistical parameters. The target precoding statistical parameters are obtained by a fixed-point iterative algorithm based on statistical parameters. This algorithm generates statistical parameters by iteratively calculating multiple sets of channel samples within a preset time period and statistically processing the resulting intermediate parameter matrix. During training, the energy distribution matrix and SNR are used as network inputs, and the corresponding precoding statistical parameters are used as supervision labels. The network parameters are backpropagated and iteratively updated by minimizing the error between the output and the labels. The loss function is the mean squared error function.

3. The method for generating large-scale MIMO precoding parameters based on channel feature mapping according to claim 1, wherein the precoding fixed-point iterative formula is determined according to the following method: A precoded optimization problem is constructed under the total power constraint with the objective of maximizing the weighted sum rate, and the constrained optimization problem is transformed into a Lagrangian form by introducing dual variables; Based on the KKT optimality condition of the optimization problem, the fixed-point update structure of the precoding matrix is ​​derived, so that the precoding matrix of each user at each base station is represented as the product of matrix inverse operation and linear combination terms. During the fixed-point iteration process, the following intermediate parameter matrix and dual variables are constructed: The first intermediate parameter matrix is ​​formed by the difference between the inverse matrix of the signal covariance matrix excluding the target user signal terms and the inverse matrix of the signal covariance matrix including the target user signal terms. The second intermediate parameter matrix is ​​formed by the product of the inverse of the signal covariance matrix excluding the target user signal terms and the sum of the effective channel vectors of the target user group; The third intermediate parameter matrix is ​​composed of the product of the first intermediate parameter matrix and the sum of the effective channel vectors of the remaining users after excluding the current target user; The dual variable, corresponding to the Lagrange multiplier under the transmit power constraint, is used to adjust the magnitude of the precoding matrix constructed from the intermediate parameter matrix so that it meets the preset total system power limit. The signal covariance matrix is ​​constructed by multiplying the precoding matrix of all users and the channel matrix, and includes a noise term and a power normalization term.

4. The method for generating large-scale MIMO precoding parameters based on channel feature mapping according to claim 2, characterized in that, The method for determining the precoding statistical parameters is as follows: multiple sets of historical channel samples collected within a preset time period are obtained, and the corresponding instantaneous parameters are calculated using a fixed-point iterative method for each set of instantaneous channel samples. Finally, the multiple sets of instantaneous parameters are statistically averaged to obtain statistical parameters, which are then used as supervision labels for neural network training.

5. The method for generating large-scale MIMO precoding parameters based on channel feature mapping according to claim 1, characterized in that, During the training phase, the signal-to-noise ratio (SNR) embedding module maps the one-dimensional SNR parameter to a vector of a preset dimension through a multilayer perceptron and concatenates and fuses it with the convolutional features output by the backbone of the convolutional neural network. During the execution phase, it uses the actual environmental SNR as input for mapping to achieve feature enhancement. Finally, the fused data is output as parameters through a multilayer perceptron.

6. The method for generating large-scale MIMO precoding parameters based on channel feature mapping according to claim 1, characterized in that, The backbone of the convolutional neural network consists of multiple layers of convolutional units connected in series. Except for the first layer, which is the output of the data preprocessing module, the other layers use the output of the previous layer as input to extract multi-scale spatial features from the energy distribution matrix.

7. The method for generating large-scale MIMO precoding parameters based on channel feature mapping according to claim 1, characterized in that, During the online execution phase, the base station receives the current channel matrix and transmits it to the network controller, which then inputs the channel matrix and the current user's environmental signal-to-noise ratio data into the MPPM network. The MPPM network generates the statistical parameters based on the pre-trained weight parameters, and uses the precoding recovery module to perform a single reconstruction calculation to obtain the final precoding matrix. Finally, the generated final precoding matrix is ​​transmitted to the base station, and the final precoding matrix is ​​used to precode the downlink signal.

8. The method for generating large-scale MIMO precoding parameters based on channel feature mapping according to claim 1, characterized in that, During the MPPM network training phase, each base station transmits the obtained channel matrix to the network controller. The network controller performs fixed-point iterative solution on multiple different channel matrices to obtain the corresponding target precoding statistical parameters. The statistical parameters are then paired with the corresponding channel energy distribution matrix and signal-to-noise ratio parameters to form training samples for supervised learning training of the MPPM network.

9. A computer system comprising a memory, a processor, and computer programs / instructions stored in the memory and executable on the processor, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method for generating large-scale MIMO precoding parameters based on channel feature mapping as described in any one of claims 1-8.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method for generating large-scale MIMO precoding parameters based on channel feature mapping as described in any one of claims 1-8.

Images