Precoding model training, precoding method, apparatus, device, medium and product

Abstract

This invention discloses a precoding model training, precoding method, apparatus, device, medium, and product. It constructs a communication system model and defines an optimization problem: under normalized constant modulus constraints and system transmit power constraints, it aims to make the hybrid precoding of the communication system model approach the optimal precoding. A dilated convolutional neural network is constructed, and a hybrid dilated convolutional layer is introduced to capture the long-range spatial correlation of the channel matrix. Several sample channel matrices are collected as training data. The optimal precoding matrix of the sample channel matrices is used as a label, and an objective function is constructed based on the optimization problem. The dilated convolutional neural network is trained to obtain a trained hybrid precoding model, which is used to process the input channel matrix and output a hybrid precoding matrix. Using this invention, hybrid precoding of millimeter-wave channels is achieved based on deep learning. The introduction of dilated convolutions reduces the complexity of the neural network model and improves the accuracy and efficiency of hybrid precoding.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a precoding model training, precoding method, apparatus, device, medium, and product. Background Technology

[0002] Existing 5G millimeter-wave hybrid precoding methods can mainly include traditional 5G millimeter-wave hybrid precoding methods and deep learning-based 5G millimeter-wave hybrid precoding methods.

[0003] Traditional 5G millimeter-wave hybrid precoding methods primarily approximate optimal unconstrained precoding performance using methods such as sparse orthogonal matching pursuit, singular value decomposition, minimum mean square error, serial interference cancellation, zero-forcing, and generalized matrix factorization. These methods are all proposed under perfect CSI (Channel State Information) scenarios. However, these methods cannot fully utilize the sparsity characteristics of millimeter-wave channels, sacrificing system performance for low complexity. Furthermore, optimization- and greedy methods suffer from limitations in obtaining optimal solutions and computation time.

[0004] Deep learning-based hybrid precoding methods for 5G millimeter-wave networks primarily rely on deep neural networks or convolutional neural networks. By training deep neural networks, a hybrid precoder is selected, thereby optimizing the entire precoding process. This involves using activation functions to optimize multiple layers of the network and creating corresponding mapping relationships, effectively avoiding iterative computations and reducing computational complexity. However, most existing deep learning-based hybrid precoding schemes depend on traditional convolutional neural networks, which leads to a large number of network parameters, high network model complexity, and a potential loss of system spectral efficiency when the hybrid precoding residual is large. Summary of the Invention

[0005] The purpose of this invention is to provide a precoding model training, precoding method, apparatus, device, medium, and product that implements hybrid precoding based on deep learning. By introducing dilated convolutions to reduce the complexity of neural network models, it effectively improves the accuracy and efficiency of hybrid precoding for millimeter-wave channels.

[0006] To achieve the above objectives, embodiments of the present invention provide a precoding model training method, comprising: A communication system model is constructed and an optimization problem is determined; wherein, the optimization problem is: under the constraints of normalized constant modulus and system transmit power, to make the hybrid precoding of the communication system model approach the optimal precoding; Construct a dilated convolutional neural network; wherein, the dilated convolutional neural network introduces a hybrid dilated convolutional layer to capture long-range spatial correlation of the channel matrix; Several sample channel matrices are collected as training data. The optimal precoding matrix of the sample channel matrices is used as the label. An objective function is constructed based on the optimization problem. The dilated convolutional neural network is trained to obtain a trained hybrid precoding model. The hybrid precoding model is used to process the input channel matrix and output the hybrid precoding matrix.

[0007] As an improvement to the above scheme, the dilated convolutional neural network also introduces a custom Lambda layer, which calculates a hybrid precoding matrix that satisfies the normalized constant modulus constraint and the system transmit power constraint through Euler's formula and least squares operation.

[0008] As an improvement to the above scheme, the dilated convolutional neural network includes an input layer, the hybrid dilated convolutional layer, a pooling layer, a fully connected layer, and the custom Lambda layer connected thereto. The input layer is used to input the channel matrix into the hybrid dilated convolutional layer; The hybrid dilated convolutional layer is used to extract spatial features of the channel matrix; The pooling layer is used to perform redundancy removal and dimensionality reduction on the spatial features to obtain abstract features; The fully connected layer is used to perform feature fusion on the abstract features and map the learned feature information into a standardized value in the range of (0, 1); the standardized value represents the phase ratio of each element of the simulated precoding matrix. The custom Lambda layer includes a first Lambda layer and a second Lambda layer. The first Lambda layer is used to calculate a complex-valued vector from the normalized numerical values ​​using Euler's formula, and then to perform matrix transformations on the complex-valued vectors to obtain an analog precoding matrix that satisfies the normalized constant modulus constraint. The second Lambda layer is used to calculate a digital precoding matrix that satisfies the system transmit power constraint by applying least squares and normalization to the analog precoding matrix, and to calculate a hybrid precoding matrix based on the analog precoding matrix and the digital precoding matrix.

[0009] As an improvement to the above scheme, the sample channel matrix includes a perfect channel matrix and an imperfect channel matrix, wherein the imperfect channel matrix is ​​generated by adding noise to the perfect channel matrix; The optimal precoding matrix is ​​obtained by performing singular value decomposition on the sample channel matrix; The objective function is to minimize the residual between the hybrid precoding matrix and the optimal precoding matrix.

[0010] This invention also provides a precoding method, comprising: Obtain the real-time channel matrix; The real-time channel matrix is ​​input into a preset hybrid precoding model for processing to obtain the analog precoding matrix, digital precoding matrix and hybrid precoding matrix of the real-time channel matrix; The hybrid precoding model is trained using the precoding model training method described in any of the above-mentioned methods.

[0011] This invention also provides a precoding model training apparatus, comprising: The optimization problem determination module is used to construct a communication system model and determine the optimization problem; wherein, the optimization problem is: under the constraints of normalized constant modulus and system transmit power, to make the hybrid precoding of the communication system model approach the optimal precoding; A neural network building module is used to construct a dilated convolutional neural network; wherein, the dilated convolutional neural network introduces a hybrid dilated convolutional layer to capture long-range spatial correlation of the channel matrix; The precoding model training module is used to collect several sample channel matrices as training data, use the optimal precoding matrix of the sample channel matrices as the label, construct the objective function with the optimization problem, and train the dilated convolutional neural network to obtain the trained hybrid precoding model; wherein, the hybrid precoding model is used to process the input channel matrix and output the hybrid precoding matrix.

[0012] This invention also provides a precoding apparatus, comprising: The channel matrix acquisition module is used to acquire the real-time channel matrix; The precoding processing module is used to input the real-time channel matrix into a preset hybrid precoding model for processing, and obtain the analog precoding matrix, digital precoding matrix and hybrid precoding matrix of the real-time channel matrix; The hybrid precoding model is trained using the precoding model training method described in any of the above-mentioned methods.

[0013] This invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the precoding model training method as described in any of the preceding claims, or the precoding method as described in any of the preceding claims.

[0014] This invention also provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the precoding model training method as described in any one of the preceding claims, or the precoding method as described in any one of the preceding claims.

[0015] This invention also provides a computer program product, which includes a computer program or computer instructions. When the computer program or computer instructions are executed by a processor, they implement the precoding model training method as described in any one of the above claims, or the precoding method as described in any one of the above claims.

[0016] Compared with existing technologies, the precoding model training, precoding method, apparatus, device, medium, and product disclosed in this invention are based on deep learning to construct a hybrid precoding model for 5G millimeter wave hybrid precoding. Compared with traditional mathematical methods, deep learning methods can solve complex optimization problems with lower computational complexity. Applying it to hybrid precoding improves system performance and achieves fast response in system processing time. This hybrid precoding model is a novel dilated convolutional neural network model. It utilizes dilated convolutional units to extract deep features from the channel matrix. This unit reduces redundant parameters in the ACNN, thereby reducing computational complexity and learning more comprehensive channel spatial correlation information without increasing the number of network parameters, thus improving the performance of the hybrid precoding algorithm. Furthermore, this dilated convolutional neural network designs a loss function by minimizing the hybrid precoding residual. It uses optimal precoding to create network labels to train and learn the predicted hybrid precoding matrix. The trained ACNN directly provides the hybrid precoding matrix, which not only achieves good spectral efficiency but also has good robustness to channel estimation errors, effectively improving the accuracy and efficiency of millimeter wave hybrid precoding. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a precoding model training method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the communication system model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the hollow neural network in an embodiment of the present invention; Figure 4 This is a schematic diagram of the hollow convolution structure in an embodiment of the present invention; Figure 5 This is a flowchart illustrating a precoding method provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0020] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] See Figure 1 This is a flowchart illustrating a precoding model training method provided in an embodiment of the present invention. The embodiment of the present invention provides a precoding model training method, the method comprising steps S11 to S13: S11. Construct a communication system model and determine the optimization problem; wherein, the optimization problem is: under the constraints of normalized constant modulus and system transmit power, to make the hybrid precoding of the communication system model approach the optimal precoding; S12. Construct a dilated convolutional neural network; wherein, the dilated convolutional neural network introduces a hybrid dilated convolutional layer to capture long-range spatial correlation of the channel matrix; S13. Collect several sample channel matrices as training data, use the optimal precoding matrix of the sample channel matrices as the label, construct the objective function with the optimization problem, train the dilated convolutional neural network, and obtain the trained hybrid precoding model; wherein, the hybrid precoding model is used to process the input channel matrix and output the hybrid precoding matrix.

[0023] The embodiments of the present invention are applicable to application scenarios of hybrid precoding based on deep learning. A hybrid precoding model is constructed by training a deep neural network for processing hybrid precoding of millimeter-wave channels.

[0024] In this embodiment of the invention, a communication system model is first constructed based on information such as the number of antennas, radio frequency links, data streams, and channel environment at the transmitting and receiving ends. See also... Figure 2 This is a schematic diagram of the communication system model in an embodiment of the present invention, showing the transmitted signal after processing by the transmitter. x It can be written as:

[0025] in, and These represent the analog precoding matrix and the digital precoding matrix, respectively. It is a baseband transmission signal, and satisfies .

[0026] In the above system model, the baseband transmission signal s After precoding matrix processing, the signal is transmitted via the transmitting antenna and then transmitted through a millimeter-wave channel to the receiving end for merging and processing to obtain the received signal. y The mathematical expression is as follows:

[0027] in, This represents the average received signal power. and These represent the digital merging matrix and the analog merging matrix, respectively. Represents the millimeter-wave channel matrix; It is obedience The complex Gaussian noise vector, where This indicates the power of the noise signal. For the number of radio frequency chains, This refers to the number of antennas at the transmitting end. For data stream numbers, This represents the number of antennas at the receiving end.

[0028] Since the phase shifters at the transmitting and receiving ends cannot affect the signal amplitude, the analog precoding matrix... and simulation merge matrix The normalized constant modulus constraint must be satisfied, i.e. , .

[0029] To better describe the transmission characteristics of signals in millimeter-wave environments, the extended Saleh-Valenzuela channel model is typically used. This channel model is assumed to contain... Clusters, each containing Path, channel matrix H It is expressed as follows:

[0030] in, Let g be the complex gain of the g-th path in the l-th cluster; , , and These are the receiving elevation angle, receiving azimuth angle, transmitting elevation angle, and transmitting azimuth angle of the g-th path in the l-th cluster, respectively. and These are represented as array response vectors at the transmitting and receiving ends, respectively.

[0031] At both the transmitting and receiving ends, considering an M×N UPA with an element spacing of half a wavelength, the expression for the array response vector is as follows:

[0032] Where m=0,…,M-1, n=0,…,N-1.

[0033] Assuming the signal transmitted in the millimeter-wave channel follows a Gaussian distribution, the spectral efficiency of the system can be expressed as follows:

[0034] in, Let be the noise covariance matrix.

[0035] Furthermore, the optimization problem of the model needs to be defined. It should be noted that the performance of hybrid precoding algorithms is generally evaluated using the maximization of spectral efficiency criterion. The optimization problem of hybrid precoding and the merging matrix can be expressed as:

[0036] st

[0037]

[0038]

[0039] in, This represents the system transmit power constraint, where This refers to the system's transmit power. This represents the Frobenius norm.

[0040] Since directly solving the above equation requires handling multiple variable matrices { Simultaneous optimization design is extremely complex to solve. Therefore, this embodiment of the invention transforms the optimization problem into one that, under the constraints of normalized constant modulus and system transmit power, makes the hybrid precoding of the communication system model approach the optimal precoding. Specifically, it is transformed into two residual minimization subproblems, wherein the optimization problem at the transmitter end can be expressed in the following form:

[0041] st

[0042]

[0043] in, This represents the optimal precoding matrix for all numbers.

[0044] Similarly, the optimization problem at the receiver end can be expressed as:

[0045] st

[0046] in, This represents the optimal merge matrix with all numbers.

[0047] As can be seen from the above, the two optimization problems at the transmitting end and the receiving end have very similar mathematical expressions. Therefore, the embodiments of the present invention mainly focus on the optimization problem at the transmitting end of the system, and similar methods are also applicable to the optimization problem at the receiving end.

[0048] To address the difficulty of directly handling normalized constant modulus constraints in optimization problems, the channel matrix can be viewed as a special image with certain spatial correlation between adjacent elements. Neural networks excel at processing matrix data in image form and can extract important features from it. Therefore, this invention proposes an ACNN (Atrous Convolutional Neural Network) framework for training a hybrid precoding model.

[0049] In this ACNN neural network, the convolutional layer is the most crucial component. When the channel matrix is ​​used as input, the convolutional kernels compress it, learning spatial features of the channel from the data. These features help ACNN improve the system's output performance. In traditional CNNs, on the one hand, the presence of pooling layers can cause some small data points in the matrix data to be missed during computation, resulting in a loss of network prediction accuracy; on the other hand, as the network layers of a CNN deepen, using traditional convolution increases the network parameters, leading to greater network complexity. To address these issues, this embodiment of the invention introduces a hybrid dilated convolutional layer into the neural network, replacing the traditional convolutional layer. This hybrid dilated convolutional layer captures long-range spatial correlations of a large-scale channel matrix without increasing the number of parameters.

[0050] Furthermore, several sample channel matrices are collected as training data. These sample channel matrices include perfect channel matrices and imperfect channel matrices. The imperfect channel matrices are generated by adding noise to the perfect channel matrices. For example, generating perfect channel matrices for T=1000 frames. Each frame generates S=100 sets of imperfect channel matrices. .

[0051] The optimal precoding matrix of the sample channel matrix is ​​used as the label, and the optimal precoding matrix is ​​obtained by performing singular value decomposition on the sample channel matrix.

[0052] According to the channel capacity theory related to massive MIMO (Multiple Input Multiple Output), and By adjusting the channel matrix H The singular value decomposition is performed to obtain the result, specifically by analyzing the front of the left and right singular matrices of the region. The column is obtained by combining the water injection power distribution.

[0053] The mathematical expression is as follows:

[0054]

[0055] in, U and V These are the left and right singular value matrices, respectively, i.e., derived from... get; P The water injection power allocation matrix; Indicates taking the matrix V The former A matrix composed of columns.

[0056] The objective function is constructed based on the optimization problem, which is to minimize the residual between the hybrid precoding matrix and the optimal precoding matrix, specifically:

[0057] st

[0058]

[0059] in, This represents the optimal precoding matrix.

[0060] Based on the sample channel matrix, the optimal coding matrix, and the objective function, the ACNN neural network is trained and tested to obtain a trained hybrid precoding model; wherein, the hybrid precoding model is used to process the input channel matrix. H Process and output the hybrid precoding matrix. .

[0061] Employing the technical means of this invention, a hybrid precoding model is constructed based on deep learning to achieve hybrid precoding for 5G millimeter waves. Compared to traditional mathematical methods, deep learning can solve complex optimization problems with lower computational complexity. Applying it to hybrid precoding improves system performance and achieves fast response in system processing time. This hybrid precoding model is a novel dilated convolutional neural network model. Dilated convolutional units extract deep-level features from the channel matrix. This unit reduces redundant parameters in the ACNN, thereby reducing computational complexity and learning more comprehensive channel spatial correlation information without increasing the number of network parameters, thus improving the performance of the hybrid precoding algorithm. Furthermore, this dilated convolutional neural network designs its loss function by minimizing the hybrid precoding residual. It uses optimal all-digital precoding to create network labels to train and learn the predicted hybrid precoding matrix. The trained ACNN directly provides the hybrid precoding matrix, achieving not only good spectral efficiency but also good robustness to channel estimation errors, effectively improving the accuracy and efficiency of millimeter-wave hybrid precoding.

[0062] As a preferred embodiment, the present invention further implements the above embodiments, wherein the dilated convolutional neural network also introduces a custom Lambda layer, and calculates a hybrid precoding matrix that satisfies the normalized constant modulus constraint and the system transmit power constraint through Euler's formula and least squares operation.

[0063] In this embodiment of the invention, the hybrid precoding problem can be viewed as a prediction problem. To solve the hybrid precoding matrix in the optimization problem, two network layers with custom functions, namely Lambda layers, are specially designed in this network to solve the simulated precoding matrix that satisfies the normalized constant modulus constraint. and the digital precoding matrix that satisfies the system transmit power constraint. Then, based on the analog and digital precoding matrices, the final hybrid precoding matrix is ​​solved. .

[0064] By employing the technical means of this invention, the hybrid precoding model constructed in this invention utilizes dilated convolutional units to extract deep-level features from the channel matrix and learns more comprehensive channel spatial correlation information without increasing the number of network parameters, thereby improving the performance of the hybrid precoding algorithm. Furthermore, in the Lambda layer with custom functionality, a Lambda layer satisfying the relevant constraints of the hybrid precoding problem is designed using Euler's formula and the least squares method, thereby calculating the hybrid precoding matrix. Based on this, the ACNN designs a loss function by minimizing the hybrid precoding residual, uses optimal all-digital precoding to create network labels for training and learning to predict the hybrid precoding matrix, and directly provides the hybrid precoding matrix using the trained ACNN. This not only achieves good spectral efficiency but also exhibits good robustness to channel estimation errors.

[0065] As a preferred embodiment, the present invention further implements the above embodiments, and provides a detailed description of the structure of the ACNN neural network. See [link to previous document]. Figure 3 This is a schematic diagram of the structure of a dilated neural network in an embodiment of the present invention. The dilated convolutional neural network includes an input layer, a hybrid dilated convolutional layer, a pooling layer, a fully connected layer, and a custom Lambda layer connected thereto.

[0066] The input layer is used to input the channel matrix into the hybrid dilated convolutional layer; the hybrid dilated convolutional layer is used to extract the spatial features of the channel matrix; the pooling layer is used to perform redundancy removal and dimensionality reduction on the spatial features to obtain abstract features; the fully connected layer is used to perform feature fusion on the abstract features and map the learned feature information into standardized values ​​in the range (0, 1); the standardized values ​​represent the phase ratio of each element of the analog precoding matrix; the custom Lambda layer includes a first Lambda layer and a second Lambda layer, the first Lambda layer is used to calculate the complex value vector by using Euler's formula on the standardized values, and then perform matrix transformation on the complex value vector to obtain the analog precoding matrix that satisfies the normalized constant modulus constraint; the second Lambda layer is used to calculate the digital precoding matrix that satisfies the system transmit power constraint by using the least squares method and normalization on the analog precoding matrix, and to calculate the hybrid precoding matrix based on the analog precoding matrix and the digital precoding matrix.

[0067] Specifically, the hyperparameter settings and processing procedures for each layer in the network are described below: (1) Input layer: The first layer has a dimension of The three-channel input layer. Real-valued convolutional neural networks have been widely used in many fields due to their ease of training and ability to achieve good performance. Because of the channel matrix... H It is a complex matrix, so... H Convert the complex number into a three-channel real-valued matrix, split the complex number into three channels: real part, imaginary part, and magnitude, and input it into ACNN.

[0068] (2) Hybrid hollow convolutional layer: The purpose of the hybrid dilated convolutional unit is to capture the long-range spatial correlation of the channel matrix without increasing the number of parameters, avoiding the accuracy loss and parameter redundancy problems of traditional CNN pooling. The hybrid dilated convolutional unit mainly consists of three dilated convolutional layers, see [link to documentation]. Figure 4 This diagram illustrates the dilated convolutional structure in an embodiment of the invention. Since multiple dilated convolutional kernels can extract different features simultaneously, and using too many kernels increases computation and model parameters, while too few may fail to extract effective features from the channel matrix, one dilated convolutional layer consists of 32 dilated kernels of size 3, and the other two dilated convolutional layers each contain 16 dilated kernels of size 3. Furthermore, to avoid the vanishing gradient problem and better fit the proposed ACNN, all three dilated convolutional layers use LeakyReLU as the activation function.

[0069] (3) Pooling layer: The pooling layer further extracts more important features from the output of the hybrid dilated convolutional units, removing redundant features that are detrimental to the estimation of the hybrid precoding matrix. Furthermore, the pooling layer in this framework reduces the dimensionality of the hybrid precoding-related features, thereby reducing the computational cost and number of parameters in the proposed network, accelerating network training, preventing overfitting, and making the model more stable. The proposed ACNN framework includes a pooling layer with a convolutional kernel size of 2.

[0070] (4) Fully connected layer: The fully connected layer's role is to highly fuse the abstract feature information obtained from previous convolutional processes and map the learned channel matrix feature information to the hybrid precoding matrix, thereby achieving end-to-end mapping. The ACNN framework contains three fully connected layers, denoted as Fully Connected Layer 1, Fully Connected Layer 2, and Fully Connected Layer 3. Furthermore, to control gradient explosion or vanishing and accelerate network convergence, a batch normalization layer is configured after the hybrid dilated convolutional units and fully connected layers. LeakyReLU is chosen as the activation function in the first two fully connected layers to further improve ACNN's learning ability. In the final fully connected layer 3, the Sigmoid activation function is used to output data within the range (0,1) to the next custom-defined Lambda layer.

[0071] (5) Custom function Lambda layer: To solve the hybrid precoding matrix in the optimization problem, two custom-function network layers, Lambda1 and Lambda2, are specially designed in this network to solve the analog and digital precoding matrices, respectively. A detailed introduction to the two custom-function Lambda layers is as follows: 1) Lambda1 layer: After processing by the fully connected layer 3, the output data is within the range (0,1) using the Sigmoid activation function. Assuming the output data of the fully connected layer 2 is... Next, The input is fed into the Lambda1 layer. By customizing the functionality of this layer, Euler's formula is embedded within it, allowing for the application of Euler's formula to... The calculations are performed to obtain the complex-valued vector, and the specific process is as follows:

[0072] in, For simulating the precoding matrix The phase; vec{.} denotes vectorization operation.

[0073] By cleverly utilizing Euler's formula in a custom network layer, Each element in the matrix has a constant modulus, and then... Matrix transformations are performed to obtain the analog precoding matrix that satisfies the normalized constant modulus constraint. Finally, the Lambda1 layer outputs the analog precoding matrix that satisfies the constraint. Compared to previous algorithms, the advantage of using neural networks to directly predict the simulated precoding matrix is ​​that it avoids a large number of time-consuming complex iterative calculations and its performance is not limited by a fixed codebook.

[0074] 2) Lambda2 layer: To ensure that ACNN ultimately outputs a hybrid precoding matrix ,in Already obtained from the aforementioned Lambda1 layer, the following is based on the known... Under the given conditions, the digital precoding matrix is ​​solved using a custom Lambda2 layer. .

[0075] Due to the analog precoding matrix Output through Lambda1 layer, and an all-digital optimal precoding matrix. Therefore, the optimization problem can be simplified to simply finding the solution. The optimization problem can be rewritten as:

[0076] st

[0077] Without considering constraints for the time being It can be obtained using the least squares method. Assume the digital precoding matrix, without considering constraints, is... Its mathematical expression is:

[0078] To account for system transmit power constraints, it is necessary to... Perform normalization to satisfy the constraints, that is:

[0079] exist and Given that all factors are known, the hybrid precoding vector is obtained. Finally, regarding f Integer, to obtain the hybrid precoding matrix .

[0080] The difference from the usual approach is that the least squares method is used to solve it. BBFThe process is accomplished by the custom-function Lambda layer in ACNN. Thus, for the trained model, only the channel matrix needs to be input into ACNN, and then processed through the aforementioned hybrid dilated convolutional units, fully connected layers, and custom-function Lambda layers. Finally, ACNN can output a result that satisfies the constraints. , and .

[0081] Based on the above analysis, the method steps of this invention can be summarized as follows:

[0082] See Figure 5 This is a flowchart illustrating a precoding method provided in an embodiment of the present invention. The embodiment of the present invention provides a precoding method, the method comprising steps S21 to S22: S21. Obtain the real-time channel matrix; S22. Input the real-time channel matrix into a preset hybrid precoding model for processing to obtain the analog precoding matrix, digital precoding matrix and hybrid precoding matrix of the real-time channel matrix.

[0083] In this embodiment of the invention, during the model training phase, a hybrid precoding model is trained using the precoding model training method described in any of the above embodiments. During the model application phase, a segment of real-time millimeter-wave channel matrix to be processed is first acquired. H The real-time channel matrix is ​​input into the trained hybrid precoding model, and through ACNN forward propagation, three results are directly output: the simulated precoding matrix. Data precoding matrix and the final hybrid precoding matrix .

[0084] It should be noted that the hybrid precoding model in this embodiment of the invention is trained using the precoding model training method described in the above embodiments. All the process steps of the two are one-to-one, and their working principles and beneficial effects are the same, so they will not be described again.

[0085] This invention also provides a precoding model training apparatus, comprising: The optimization problem determination module is used to construct a communication system model and determine the optimization problem; wherein, the optimization problem is: under the constraints of normalized constant modulus and system transmit power, to make the hybrid precoding of the communication system model approach the optimal precoding; A neural network building module is used to construct a dilated convolutional neural network; wherein, the dilated convolutional neural network introduces a hybrid dilated convolutional layer to capture long-range spatial correlation of the channel matrix; The precoding model training module is used to collect several sample channel matrices as training data, use the optimal precoding matrix of the sample channel matrices as the label, construct the objective function with the optimization problem, and train the dilated convolutional neural network to obtain the trained hybrid precoding model; wherein, the hybrid precoding model is used to process the input channel matrix and output the hybrid precoding matrix.

[0086] It should be noted that the precoding model training device provided in this embodiment of the invention is used to execute all the process steps of the precoding model training method in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0087] This invention also provides a precoding apparatus, comprising: The channel matrix acquisition module is used to acquire the real-time channel matrix; The precoding processing module is used to input the real-time channel matrix into a preset hybrid precoding model for processing, to obtain the analog precoding matrix, digital precoding matrix and hybrid precoding matrix of the real-time channel matrix; wherein the hybrid precoding model is trained using the precoding model training method described in any of the above embodiments.

[0088] It should be noted that the precoding apparatus provided in this embodiment of the invention is used to execute all the process steps of the precoding method in the above embodiment. The working principle and beneficial effects of the two correspond one-to-one, so they will not be described again.

[0089] This invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the precoding model training method as described in any of the above embodiments, or the precoding method as described in any of the above embodiments.

[0090] This invention also provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the precoding model training method as described in any of the above embodiments, or the precoding method as described in any of the above embodiments.

[0091] This invention also provides a computer program product, which includes a computer program or computer instructions. When the computer program or computer instructions are executed by a processor, they implement the precoding model training method as described in any of the above embodiments, or the precoding method as described in any of the above embodiments.

[0092] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0093] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for training a pre-encoding model, characterized in that, include: A communication system model is constructed and an optimization problem is determined; wherein, the optimization problem is: under the constraints of normalized constant modulus and system transmit power, to make the hybrid precoding of the communication system model approach the optimal precoding; Construct a dilated convolutional neural network; wherein, the dilated convolutional neural network introduces a hybrid dilated convolutional layer to capture long-range spatial correlation of the channel matrix; Several sample channel matrices are collected as training data. The optimal precoding matrix of the sample channel matrices is used as the label. An objective function is constructed based on the optimization problem. The dilated convolutional neural network is trained to obtain a trained hybrid precoding model. The hybrid precoding model is used to process the input channel matrix and output the hybrid precoding matrix. 2.The pre-encoding model training method of claim 1, wherein, The dilated convolutional neural network also introduces a custom Lambda layer, which calculates a hybrid precoding matrix that satisfies the normalized constant modulus constraint and the system transmit power constraint through Euler's formula and least squares operation.

3. The precoding model training method as described in claim 2, characterized in that, The dilated convolutional neural network includes an input layer, the hybrid dilated convolutional layer, a pooling layer, a fully connected layer, and the custom Lambda layer, all connected thereto. The input layer is used to input the channel matrix into the hybrid dilated convolutional layer; The hybrid dilated convolutional layer is used to extract spatial features of the channel matrix; The pooling layer is used to perform redundancy removal and dimensionality reduction on the spatial features to obtain abstract features; The fully connected layer is used to perform feature fusion on the abstract features and map the learned feature information into a standardized value in the range of (0, 1); the standardized value represents the phase ratio of each element of the simulated precoding matrix. The custom Lambda layer includes a first Lambda layer and a second Lambda layer. The first Lambda layer is used to calculate a complex-valued vector from the normalized numerical values ​​using Euler's formula, and then to perform matrix transformations on the complex-valued vectors to obtain an analog precoding matrix that satisfies the normalized constant modulus constraint. The second Lambda layer is used to calculate a digital precoding matrix that satisfies the system transmit power constraint by applying least squares and normalization to the analog precoding matrix, and to calculate a hybrid precoding matrix based on the analog precoding matrix and the digital precoding matrix.

4. The precoding model training method as described in claim 1, characterized in that, The sample channel matrix includes a perfect channel matrix and an imperfect channel matrix, wherein the imperfect channel matrix is ​​generated by adding noise to the perfect channel matrix; The optimal precoding matrix is ​​obtained by performing singular value decomposition on the sample channel matrix; The objective function is to minimize the residual between the hybrid precoding matrix and the optimal precoding matrix.

5. A precoding method, characterized in that, include: Obtain the real-time channel matrix; The real-time channel matrix is ​​input into a preset hybrid precoding model for processing to obtain the analog precoding matrix, digital precoding matrix and hybrid precoding matrix of the real-time channel matrix; The hybrid precoding model is trained using the precoding model training method described in any one of claims 1 to 4.

6. A pre-coding model training device, characterized in that, include: The optimization problem determination module is used to construct a communication system model and determine the optimization problem; wherein, the optimization problem is: under the constraints of normalized constant modulus and system transmit power, to make the hybrid precoding of the communication system model approach the optimal precoding; A neural network building module is used to construct a dilated convolutional neural network; wherein, the dilated convolutional neural network introduces a hybrid dilated convolutional layer to capture long-range spatial correlation of the channel matrix; The precoding model training module is used to collect several sample channel matrices as training data, use the optimal precoding matrix of the sample channel matrices as the label, construct the objective function with the optimization problem, and train the dilated convolutional neural network to obtain the trained hybrid precoding model; wherein, the hybrid precoding model is used to process the input channel matrix and output the hybrid precoding matrix.

7. A precoding device, characterized in that, include: The channel matrix acquisition module is used to acquire the real-time channel matrix; The precoding processing module is used to input the real-time channel matrix into a preset hybrid precoding model for processing, and obtain the analog precoding matrix, digital precoding matrix and hybrid precoding matrix of the real-time channel matrix; The hybrid precoding model is trained using the precoding model training method described in any one of claims 1 to 4.

8. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the precoding model training method as described in any one of claims 1 to 4, or the precoding method as described in claim 5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the precoding model training method as described in any one of claims 1 to 4, or the precoding method as described in claim 5.

10. A computer program product, characterized in that, The computer program product includes a computer program or computer instructions, which, when executed by a processor, implement the precoding model training method as described in any one of claims 1 to 4, or the precoding method as described in claim 5.

Images