A machine learning based joint design method of channel estimation and symbol level precoding

Abstract

The application discloses a machine learning assisted symbol level precoding and channel estimation joint design method, comprising: generating a data set containing multiple user received signals; using a convolutional neural network to process the received signals to obtain an estimated channel matrix by taking minimizing the mean square error between the real channel and the estimated channel as the criterion; combining the estimated channel matrix with user symbol information to construct channel data containing virtual and real parts; using a Transform neural network to jointly process the channel matrix and the user symbol information to calculate key low-dimensional variables of symbol level precoding optimization; and combining the estimated channel information and the key low-dimensional variables to calculate symbol level precoding closed form to obtain a sending signal. The application can realize joint optimization of channel estimation and symbol level precoding, and significantly improve system performance with low calculation complexity.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of large-scale MIMO communication, and particularly relates to a joint design method of channel estimation and symbol-level precoding based on machine learning. BACKGROUND

[0002] In recent years, mobile communication technology has developed rapidly, and higher requirements have been put forward for the spectrum efficiency and energy efficiency of communication systems. As a key technology of cellular wireless communication, large-scale MIMO technology significantly improves system performance through spatial multiplexing and diversity gain, and becomes the core technology of the new generation of mobile communication standards.

[0003] Precoding is a key signal processing technology in wireless communication systems. By performing complex transformation on data at the transmitting end, it can effectively improve the reliability and spectrum efficiency of signal transmission. Precoding technology mainly includes linear precoding and nonlinear precoding.

[0004] Linear precoding uses a linear transformation matrix to accurately map the transmission signal to a multi-antenna array, achieving spatial diversity gain. This method effectively suppresses multi-user interference in a multi-antenna system through algebraic transformation, improving system capacity and communication reliability.

[0005] Nonlinear precoding has significant advantages over linear precoding. Its core feature is to break through the constraints of linear transformation and achieve more complex signal processing through nonlinear mapping. This technology can more flexibly adjust the power and phase distribution of the transmitted signal, significantly improving system performance. Symbol-level precoding, as a representative method of nonlinear precoding, innovatively realizes beneficial interference exploitation in multi-user systems by specially processing data symbols.

[0006] The core feature of symbol-level precoding is beneficial interference exploitation. Unlike traditional precoding, this method actively utilizes interference signals in multi-user systems. Through carefully designed symbol mapping strategies, it converts harmful interference into usable signal resources, thereby breaking through the performance limits of traditional precoding. The key lies in innovatively reconstructing the spatial structure of the transmitted signal, so that the signals of different users can produce beneficial interactions during transmission.

[0007] Accurate acquisition of channel state information is the key to optimizing the performance of wireless communication systems. Traditional channel estimation methods insert pilot signals at the transmitting end and estimate channel characteristics based on pilot signals at the receiving end. However, multipath fading, time-varying channels, and complex noise environments pose great challenges to channel estimation.

[0008] Deep learning provides an innovative technical paradigm for wireless communication systems. Through massive data training and adaptive network structure, machine learning algorithms can accurately model complex nonlinear channel characteristics, breaking through the limitations of traditional signal processing methods. In particular, in the field of symbol-level precoding and channel estimation, deep learning shows strong modeling and optimization potential.

[0009] However, existing symbol-level precoding and channel estimation methods still face many challenges. Traditional channel estimation techniques mainly rely on pilot signals and linear estimation models, making it difficult to accurately capture the nonlinear characteristics and time-varying nature of complex wireless channels. Existing precoding methods are mostly limited to linear transformations, and cannot fully utilize the interference information in multi-user systems. In addition, traditional methods usually use iterative algorithms for channel estimation and precoding design, which have high computational complexity and severely restrict real-time performance. In scenarios where pilot resources are limited, existing technologies are more difficult to accurately estimate channel states and perform effective symbol-level precoding, resulting in significant degradation in system performance. SUMMARY

[0010] The present invention focuses on analyzing the impact of imperfect channel state information errors in channel estimation on the performance of symbol-level precoding in large-scale MIMO systems. Based on this, a joint design of symbol-level precoding and channel estimation based on deep learning is proposed, which optimizes both to improve the performance of downlink transmission, thereby solving the problems in the above background art.

[0011] Technical solution: To achieve the above invention purpose, the technical solution adopted by the present invention is:

[0012] A machine learning-based joint design method for channel estimation and symbol-level precoding, comprising the following steps:

[0013] Generate a dataset containing multiple user receive signals;

[0014] Using a convolutional neural network to process the receive signals to obtain an estimated channel matrix, with the criterion of minimizing the mean square error between the real channel and the estimated channel;

[0015] Combine the estimated channel matrix with user symbol information to construct channel data containing real and imaginary parts;

[0016] Use a Transformer neural network to jointly process the channel matrix and user symbol information to calculate the key low-dimensional variable of symbol-level precoding optimization, i.e., the normal vector of the user symbol favorable interference region;

[0017] Combine the estimated channel information and the key low-dimensional variable to calculate the symbol-level precoding closed form to obtain the transmitted signal.

[0018] Further, the K channel vectors are jointly represented as The received signals of all users can be jointly represented as where K is the number of users and N is the number of antennas at the base station.

[0019] Further, the input layer of the convolutional neural network inputs the received signal to be processed, separates the received signal into real and imaginary parts and normalizes them, and the normalized data samples enter the neural network from the input layer; then, the input information is repeatedly convolved, pooled, and activated multiple times to extract the frequency domain and spatial features in the received signal; after multiple convolutions, pooling, and activation, the extracted feature data is input into the fully connected layer, and the extracted features are mapped to the estimated channel matrix; at the same time, the signal-to-noise ratio also participates in the full connection as a node to adjust the training results; finally, two sets of real number matrices are output from the output layer, and after merging, a complex number matrix is obtained, representing the estimated channel state information.

[0020] Further, the estimated channel matrix is spliced with the K user symbol information matrix in the data set to obtain an intermediate matrix The real and imaginary parts of M are unfolded and concatenated along the column direction to obtain

[0021] Further, the dimensionality of is reduced by the Transformer neural network to obtain the key low-dimensional variable

[0022] Further, the input encoding layer of the Transformer neural network first preprocesses the complex number matrix M, separates its real and imaginary parts, and splices them to form a real number matrix Then, position encoding is performed to preserve the spatial relationship information of the matrix elements; the Transformer encoder extracts the key features in the matrix by stacking L layers of structure, each layer containing multi-head self-attention, layer normalization, and a feedforward network, and the mapping layer uses a fully connected network to compress the output of the Transformer encoder to the target dimension 2Kx1 to obtain the low-dimensional variable

[0023] Further, in the training phase, the Transformer neural network calculates the predicted low-dimensional variable by forward propagation, and then uses a loss function to evaluate the quality of the current

[0024]

[0025] where

[0026]

[0027] respectively, Λ represents a normal boundary parameter diagonal matrix of a user symbol favorable interference area, σ 2 represents noise power, P T represents a transmission signal constraint, and x represents a single user transmission signal.

[0028] Further, a diagonal matrix Λ is calculated through a user symbol information matrix s in a data set, and a key low-dimensional variable δ and an estimated channel matrix is substituted into a transmission vector expression without power limitation, to obtain an optimal transmission vector without power limitation separated by real and imaginary parts

[0029]

[0030] wherein,

[0031]

[0032] Further, according to the obtained optimal transmission vector without power limitation real and imaginary parts are separated and represented as a complex number form x through a mapping relationship * , and a power scaling factor is used to normalize the vector, to obtain an optimal transmission signal satisfying a power constraint

[0033] The application also provides a computer program product comprising a computer program, which, when executed by a processor, implements the steps of the machine learning-based channel estimation and symbol-level precoding joint design method.

[0034] Beneficial effects: the machine learning-based channel estimation and symbol-level precoding joint design method has the following advantages compared with the prior art:

[0035] 1. Compared with the traditional optimization-based symbol-level precoding scheme, the neural network-based symbol-level precoding scheme can greatly reduce the computational complexity required for symbol-level updating while maintaining good downlink transmission performance based on data-driven and non-iterative calculation characteristics. At the same time, the data-driven characteristics of the neural network greatly simplify the design of symbol-level precoding under imperfect state information.

[0036] 2. Compared with the traditional channel estimation scheme based on the compressed sensing algorithm, the neural network-based channel estimation method compensates for the performance loss that may be caused by the mismatch of the channel model in a data-driven manner, thereby reducing the computational complexity while improving the performance of the channel estimation algorithm.

[0037] 3. Under the joint design framework based on neural network, the neural network-based channel estimation method will optimize the results of channel estimation in the limited pilot scenario to enhance the symbol-level precoding performance, and the symbol-level precoding will adapt to the error in channel estimation, thereby having strong robustness. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The method of the present application is a general flowchart;

[0039] Figure 2 The detailed method flowchart of the embodiment of the present application is as follows. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. The specific embodiments described herein are only used to explain the present application, and are not used to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] The embodiment of the present application provides a machine learning-based channel estimation and symbol-level precoding joint design method. By designing a received signal model and generating a data set containing received signals of multiple users, the optimization of channel estimation and symbol-level precoding is supported. On this basis, the received signal is processed by using a convolutional neural network, aiming to minimize the mean square error between the real channel and its estimated channel, so as to obtain an accurate channel estimation matrix. The estimated channel matrix is then combined with user symbol information, and the channel matrix and user symbol information are reduced in dimension by relying on a Transformer, and effective symbol representation is extracted. Based on the dimension reduction result, an optimization problem is constructed, aiming to minimize the error between the received signal and the target symbol, and the symbol-level precoding is calculated to obtain the transmitted signal, thereby realizing effective joint design of channel estimation and symbol-level precoding.

[0042] Specifically, as shown in the figure, Figure 1 The machine learning-based channel estimation and symbol-level precoding joint design method described in the embodiment specifically includes the following steps:

[0043] (1) Design a received signal model and generate a data set containing received signals of multiple users for subsequent neural network training.

[0044] In this step, a received signal model is established using real channel data and unit pilot data. Exemplarily, consider a single-cell multi-user large-scale MIMO uplink pilot transmission link, where the number of antennas at the base station end is N, the number of users is K, and each user is equipped with 1 antenna. Let the received signal of the kth user be y k , the base station in the cell where the user is located is sent to the base station The channel vector is The noise n k is a zero-mean additive Gaussian white noise with a variance of σ 2 , and the received signal model can be represented as

[0045]

[0046] K channel vectors are jointly represented as The received signals of all users can be jointly represented as

[0047] (2) Using a convolutional neural network to process the received signal to obtain an estimated channel matrix, with the criterion of minimizing the mean square error between the real channel and the estimated channel.

[0048] In this step, the criterion for channel estimation is to minimize the mean square error between the real channel and the estimated channel. Since the deep learning-based channel estimation process requires separate channel estimation for each object, the received signal is obtained by using a convolutional neural network

[0049] In this embodiment, the convolutional neural network is divided into an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer. The input layer inputs the received signal to be processed. Considering that the received signal is a complex number and the convolutional neural network processes real numbers, the received signal needs to be separated into real and imaginary parts and normalized first. The normalized data samples enter the neural network from the input layer. Then, the input information is repeatedly convolved, pooled, and activated multiple times. The convolutional layer performs convolution operation on the input samples to extract the frequency domain and spatial features in the KxNxM-dimensional received signal. In order to reduce the complexity of the network as much as possible, the number of convolution kernels is as small as possible without affecting the performance. The pooling layer can reduce the number of features extracted by the convolutional layer and eliminate redundant features through the pooling operation. The role of the activation layer is to introduce a nonlinear factor, and ReLU(·) function is used for activation after each layer of pooling.

[0050] After completing multiple convolutions, pooling and activation, the extracted feature data is input into a fully connected layer, which maps the extracted features into an estimated channel matrix. At the same time, the signal-to-noise ratio is also involved as a node in the full connection, which adjusts the training results and improves the accuracy of channel estimation. Finally, two groups of KxN-dimensional real number matrices are output from the output layer, corresponding to the real and imaginary parts of the estimated channel, respectively, and after merging, a KxN-dimensional complex matrix is obtained represents the estimated channel state information.

[0051] (3) The estimated channel matrix is combined with the user's symbol information to construct channel data containing real and imaginary parts.

[0052] In this step, the estimated channel matrix and the K user symbol information matrix in the data set are spliced together to obtain an intermediate matrix The real and imaginary parts of M are expanded and concatenated along the column direction to obtain

[0053] (4) The channel matrix and user symbol information are jointly processed using a Transformer neural network to calculate the key low-dimensional variable of symbol-level precoding optimization, i.e., the normal vector of the user symbol favorable interference region.

[0054] In this step, the Transformer neural network is used to reduce the dimension to obtain the key low-dimensional variable

[0055] The Transformer neural network mainly includes an input encoding layer, a multi-head self-attention layer, a feedforward neural network layer, and an output mapping layer.

[0056] The input encoding layer first preprocesses the complex matrix M, separates its real and imaginary parts, and splices them to form a real matrix Then, position encoding is performed to preserve the spatial relationship information of the matrix elements. Position encoding uses position embedding vectors generated by sine and cosine functions to add to the input features, allowing the network to perceive the relative position of elements in the matrix.

[0057] The multi-head self-attention mechanism is the core component of the Transformer. For the i-th attention head, the query matrix Q i , the key matrix K i and the value matrix V i are generated through a linear projection layer:

[0058]

[0059] where is a learnable projection matrix. The i-th attention head attention computation formula is:

[0060]

[0061] wherein, compute the dot product between the query and key, measure the relevance, d i is the dimension of the query and key, head i denotes the feature representation matrix after attention weighting. The multi-head mechanism computes H independent attention heads in parallel, each of which extracts information in different feature subspaces, and the output Z of the multi-head attention is obtained by concatenating the results of each head and linear transformation:

[0062] Z = Concat (head1, head2, …, head H )W o

[0063] wherein W o is a projection matrix, and the output is transmitted to the next layer after residual connection and layer normalization.

[0064] The feedforward neural network layer receives the output matrix Z of the self-attention layer, which consists of two fully connected layers. The first layer maps the input to a higher-dimensional hidden layer and introduces a nonlinear transformation using the ReLU(·) activation function, and the second layer maps the hidden layer back to the original dimension. The expression of this layer is:

[0065] FFN(Z) = max(0, ZW1 + b1)W2 + b2

[0066] wherein W1, W2 are weight matrices, and b1, b2 are bias vectors. The output of the feedforward network is also processed by layer normalization.

[0067] The Transformer encoder extracts key features in the matrix by stacking L layers of the above structure (each layer containing multi-head self-attention, layer normalization, and feedforward network), and the output mapping layer uses a fully connected network to compress the output of the Transformer encoder to the target dimension 2Kx1, obtaining the low-dimensional variable δ.

[0068] In the training phase, the network calculates the predicted low-dimensional variable δ by forward propagation, and then evaluates the quality of the current δ and using the loss function

[0069]

[0070] wherein,

[0071]

[0072] respectively represent the real and imaginary parts, Λ represents the normal boundary parameter diagonal matrix of the user symbol favorable interference region, σ 2 represents the noise power, P T represents the transmission signal constraint, x represents the single-user transmission signal, and ||·||2 is the two-norm of the vector.

[0073] The loss function comprehensively considers the signal transmission error and power constraint, wherein the first term measures the error between the transmission signal and the target symbol, and the second term represents the power regularization term. Based on the calculated loss value, the gradient of the loss function with respect to the parameters of each layer of the network is calculated through the back propagation algorithm, and the weight matrix and bias vector of the Transformer are updated using the gradient descent method, so that the network gradually learns the optimal dimension reduction mapping relationship.

[0074] After sufficient training, in the actual application stage, a new input matrix The network does not need to perform loss function optimization and parameter update again, but directly outputs the optimal low-dimensional variable δ through the forward calculation process.

[0075] (5) Combining the estimated channel information and the key low-dimensional variable, the symbol-level precoding closed form is calculated to obtain the transmission signal.

[0076] In this step, the extended vector is first subjected to symbol feature extraction to generate the symbol matrix

[0077]

[0078] Wherein, the sign(·) function maps the elements to {-1, 0, 1} according to the positive and negative nature of the vector elements.

[0079] By judging whether the vector elements satisfy the modulation order threshold condition, the indication vector w

[0080]

[0081] Wherein, |·| is the absolute value of the vector element, θ min and θ max represent the lower and upper threshold values of the symbol favorable interference region, respectively.

[0082] The normal boundary parameter diagonal matrix of the user symbol information favorable interference region is is obtained by element-by-element multiplication operation

[0083]

[0084] where, represents the element-wise multiplication of two vectors, and diag(·) means to form a diagonal matrix with the elements in the parentheses as the diagonal elements.

[0085] The calculated diagonal matrix A, the vector δ obtained in step (4), and the estimated channel matrix obtained in step (2) Substituted into the expression of the power-unlimited transmission vector, the optimal power-unlimited transmission vector with real and imaginary parts separated is obtained

[0086]

[0087] where,

[0088]

[0089] According to the obtained optimal power-unlimited transmission vector The real and imaginary parts are separated by the mapping relationship and expressed as a complex number x * , the power scaling factor The vector is normalized to obtain the optimal transmission signal that satisfies the power constraint

[0090] This embodiment utilizes the advantages of deep learning in channel estimation and symbol-level precoding, and innovatively solves the limitations of traditional channel parameter estimation methods. Machine learning methods can significantly improve the accuracy and robustness of channel estimation through nonlinear feature extraction and deep representation.

[0091] In a multi-user massive MIMO communication system, the deep learning model can effectively capture complex channel features, and has stronger signal processing capability compared to traditional linear estimation methods. The introduction of advanced technologies such as convolutional neural networks and Transformers enables the channel parameter estimation to be transformed from linear mapping to nonlinear feature extraction.

[0092] Channel estimation is a key technology in modern communication systems, directly affecting the communication performance and spectral efficiency of the system. This scheme breaks through the performance bottleneck of traditional channel estimation in complex channel environments through machine learning methods, providing a new technical path for efficient and accurate signal processing.

[0093] Precoding technology is a key signal processing technology in multi-antenna communication systems, and its core goal is to effectively combat channel fading and interference through signal preprocessing at the sending end, improving the communication performance of the system. This scheme realizes end-to-end optimization of signal processing by jointly designing channel estimation and symbol-level precoding, providing an innovative solution for future intelligent communication systems.

[0094] The embodiment of the present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the method for joint design of channel estimation and symbol-level precoding based on machine learning. The program code for implementing the method of the present application can be written in any combination of one or more programming languages. These program codes can be provided to the processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, so that the program codes, when executed by the processor or controller, cause the steps of the method of the present application to be implemented. The program codes can be executed entirely on a machine, partially on a machine, partially on a machine as a separate software package, and partially on a remote machine or server. The present application does not detail the known technologies of those skilled in the art.

[0095] Finally, it should be noted that: the above is only the preferred embodiment of the present application, and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced equivalently, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A joint design method for channel estimation and symbol-level precoding based on machine learning, characterized in that: Includes the following steps: Generate a dataset containing signals received by multiple users; Using the criterion of minimizing the mean square error between the real channel and the estimated channel, a convolutional neural network is used to process the received signal to obtain the estimated channel matrix; The estimated channel matrix is ​​combined with user symbol information to construct channel data containing both real and virtual parts; By using the Transformer neural network to jointly process the channel matrix and user symbol information, the key low-dimensional variable for symbol-level precoding optimization is calculated, namely the normal vector of the user symbol's favorable interference region. By combining the estimated channel information and key low-dimensional variables, the symbol-level precoding closed form is calculated to obtain the transmitted signal; This is achieved through the user symbol information matrix in the dataset. Calculate the diagonal matrix , with key low-dimensional variables and the estimated channel matrix Substituting these values ​​into the expression for an unconstrained transmit vector, we obtain the optimal unconstrained transmit vector with separate real and virtual components. : ； in, ； , These represent taking the real part and the imaginary part, respectively. This represents the diagonal matrix of normal boundary parameters for the user-symbol advantageous interference region. Indicates noise power. Indicates the constraint of sending signals. Represents a unit vector. Indicates the number of users.

2. The method for joint design of channel estimation and symbol-level precoding based on machine learning according to claim 1, characterized in that: In the dataset, The joint representation of the channel vectors is as follows: The received signals from all users are jointly represented as ;in This represents the number of antennas at the base station.

3. The method for joint design of channel estimation and symbol-level precoding based on machine learning according to claim 1, characterized in that: The input layer of the convolutional neural network takes in the received signal to be processed, separates the received signal into real and imaginary parts and normalizes it, and the normalized data sample enters the neural network from the input layer; then, the input information undergoes multiple repetitions of convolution, pooling and activation to extract the frequency domain and spatial features in the received signal. After multiple convolutions, pooling, and activations, the extracted feature data is input into the fully connected layer, which maps the extracted features to the estimated channel matrix. At the same time, the signal-to-noise ratio also participates in the fully connected layer as a node to adjust the training results. Finally, two sets of real matrices are output from the output layer, which are merged to obtain a complex matrix representing the estimated channel state information.

4. The method for joint design of channel estimation and symbol-level precoding based on machine learning according to claim 2, characterized in that: The estimated channel matrix and in the dataset User symbol information matrix By piecing them together, we obtain the intermediate matrix. ,Will Expanding the real and virtual parts, and cascading them along the column direction, we obtain .

5. The method for joint design of channel estimation and symbol-level precoding based on machine learning according to claim 4, characterized in that: Through the Transformer neural network Dimensionality reduction to obtain key low-dimensional variables .

6. The method for joint design of channel estimation and symbol-level precoding based on machine learning according to claim 1, characterized in that: The input encoding layer of the Transformer neural network first processes the complex matrix... Preprocessing is performed to separate and concatenate the real and imaginary parts to form a real matrix. Then, position encoding is performed to preserve the spatial relationship information of the matrix elements; Transformer encoders are stacked The structure consists of layers, each containing multi-head self-attention, layer normalization, and a feedforward network to extract matrices. A key feature of this process is that the output mapping layer uses a fully connected network to compress the output of the Transformer encoder to the target dimension. To obtain low-dimensional variables .

7. The method for joint design of channel estimation and symbol-level precoding based on machine learning according to claim 4, characterized in that: During the training phase, the Transformer neural network computes the predicted low-dimensional variables through forward propagation. Then, the loss function is used to evaluate the current and Quality: ； in, ； This indicates that a single user is sending a signal.

8. The method for joint design of channel estimation and symbol-level precoding based on machine learning according to claim 1, characterized in that: Based on the obtained optimal transmission vector without power limitations By using mapping relationships, the real and imaginary parts are separated and reconstructed into complex form. Using power scaling factor Normalize the vector to obtain the optimal transmitted signal that satisfies the power constraint. .

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the joint design method of channel estimation and symbol-level precoding based on machine learning as described in any one of claims 1-8.

Images