A Channel Acquisition Method Based on Spatial-Frequency Correlation

Abstract

This invention discloses a channel acquisition method based on space-frequency correlation. In large-scale multiple-input multiple-output (MIMO)-orthogonal frequency division multiplexing (OFDM) systems, timely and accurate acquisition of channel information is crucial for maximizing the system's wireless transmission gain. This invention uses a complex-domain fully connected mixer neural network (CMixer) to learn the mapping from partially known space-frequency sub-channels to the complete MIMO-OFDM channel. Based on the trained CMixer, devices in the system requiring high-dimensional channel acquisition only need to acquire a few space-frequency sub-channels with minimal signaling resources to map the complete MIMO-OFDM channel. This invention enables channel acquisition based on space-frequency correlation, providing a low-signaling-overhead channel acquisition method in MIMO-OFDM systems.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and in particular to a channel acquisition method based on space-frequency correlation. Background Technology

[0002] Acquiring Channel State Information (CSI) is a crucial issue in wireless communication. In next-generation wireless communication systems, with the use of Multiple-Input Multiple-Output (MIMO) technology and wider communication bandwidths such as millimeter waves, timely and accurate CSI acquisition faces significant challenges. Multiple antennas and wider communication bandwidths greatly increase the dimensions of the channel that needs to be acquired, leading to greater signaling overhead.

[0003] To reduce the signaling overhead of pilot estimation in MIMO and Orthogonal Frequency Division Multiplexing (OFDM) systems, mapping the entire MIMO-OFDM channel to a portion of the space-frequency sub-channels has become an effective approach. Based on this channel mapping technique, the channel requiring signaling overhead for estimation is reduced from all transmission antennas and subcarriers to a few antennas and carriers, constituting a channel acquisition method based on space-frequency correlation. However, although the channels between different antennas / carriers exhibit significant correlation, this correlation is often highly nonlinear, making it difficult to accurately mine and utilize using traditional signal processing methods.

[0004] In recent years, deep learning technology has been applied to channel mapping tasks, achieving better performance than traditional signal processing methods. Current deep learning-driven channel mapping methods are mainly based on conventional convolutional neural networks (CNNs) or multilayer perceptrons (MLPs). However, since channel state information (CSI) is not entirely the same as the characteristics of images, CNN methods designed for image properties lack fitting ability and robustness when performing pattern recognition on CSI. On the other hand, while MLP methods have excellent fitting ability and robustness, they are prone to severe overfitting due to a lack of sufficient priors and regularization, limiting the final performance of the network. Therefore, current deep learning methods still have some bottlenecks, preventing the full potential of deep learning technology from being realized in current tasks. How to more effectively design deep learning methods to learn the mapping from partial space-frequency sub-channels to the complete channel and achieve signal acquisition based on space-frequency correlation has become an important problem in current MIMO-OFDM systems. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an efficient CSI acquisition scheme based on space-frequency correlation, so as to better achieve channel acquisition with low pilot overhead.

[0006] The specific technical solution adopted in this invention is as follows:

[0007] A channel acquisition method based on space-frequency correlation, the method comprising:

[0008] S1. Collect user CSI data;

[0009] S2. Construct a multi-domain fully connected mixer neural network as a mapping network, and randomly initialize the neural network parameters; the network includes a fully connected layer and K CMixer layers;

[0010] S3. Train the neural network using the collected CSI data.

[0011] S4. Input the known channels on some antennas and carriers into the trained network model, and output the complete MIMO-OFDM channel acquisition results.

[0012] Furthermore, the collected user CSI data specifically includes channels on some antennas and carriers, as well as channels on the complete antenna and carrier.

[0013] Furthermore, the complex-domain fully connected mixer neural network, as a mapping network, consists of a... A fully connected layer, a A fully connected layer, K CMixer layers, and a 2N′ t →2N t A fully connected layer and a It consists of fully connected layers, in which and These are the spatial and frequency domain dimensions of the input channel, respectively, N′ t and N′ c It is a structural hyperparameter, N t and N c It refers to the size of the spatial and frequency domain dimensions of the output channel.

[0014] Furthermore, the CMixer layer consists of a complex domain MLP for spatial domain mixing, a complex domain MLP for frequency domain mixing, and several layers of normalization and residual connections.

[0015] Furthermore, the complex domain MLP network structure consists of a 2X→S fully connected layer, an activation function, and an S→2X fully connected layer; X is the size of the last dimension of the input complex-valued data, and S is the hidden layer width.

[0016] Furthermore, the CMLP calculation process is as follows: first, the input complex-valued data is stretched into real-valued data, then input into a 2X→S fully connected layer and passed through an activation function, then input into a S→2X fully connected layer, and then reshaped into complex-valued data.

[0017] Furthermore, the loss function of a neural network is defined as follows:

[0018]

[0019] Where Θ is the parameter set of the neural network, N is the number of training data, and H is the number of training data. n This is the nth original training data. It is H n The output obtained after inputting into the CMixer neural network.

[0020] The beneficial effects of this invention are as follows: The channel mapping method based on CMixer proposed in this invention can acquire high-dimensional MIMO-OFDM channels with high accuracy. Through the overlapping and concatenated spatial and frequency domain learning in CMixer, the coupling between the inherent spatial and frequency domain characteristics of the channel is effectively enhanced, thus solving the problem that other current deep learning-based CSI mapping methods have poor generalization and cannot achieve excellent performance with limited training data. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of MIMO-OFDM channel acquisition through a channel mapping network;

[0022] Figure 2 This is a schematic diagram of the CMixer's specific structure;

[0023] Figure 3 This is a grayscale visualization comparison of the CMixer-based channel mapping method of the present invention with other commonly used deep learning-based channel mapping methods in obtaining channel quality on a random channel sample. Detailed Implementation

[0024] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] In this embodiment, the base station of the large-scale access system is equipped with 32 antennas, each user terminal is configured with 1 antenna, and the base station uses OFDM mode with 32 subcarriers to serve users.

[0026] This embodiment provides a deep learning-driven channel mapping scheme, and can be used to implement a channel acquisition scheme based on space-frequency correlation, which includes the following steps:

[0027] 1) Collect 40,000 CSI data points from historical communication data as training data for the neural network;

[0028] 2) According to Figure 2The structure shown is used to build a complex-domain fully connected mixer neural network (CMixer) as a mapping network, and the neural network parameters are randomly initialized.

[0029] 2.a) The neural network structure of CMixer is as follows. (Example: ...) Figure 2 As shown in (c), CMixer consists of a A fully connected layer, a A fully connected layer, K CMixer layers, and a 2N′ t →2N t A fully connected layer and a It consists of fully connected layers. and These are the spatial and frequency domain dimensions of the input channel, respectively, N′ t and N′ c It is a structural hyperparameter, N t and N c K represents the size of the spatial and frequency domain dimensions of the output channel, while K is a hyperparameter that determines the depth of the neural network.

[0030] The input channel is a 5-antenna, 5-carrier channel, i.e. The output complete MIMO-OFDM channel is a 32-antenna, 32-carrier channel, i.e., N t =32, N c =32. The hyperparameters of the neural network are set as follows: K=5, N′ t =32, N′ c =32.

[0031] 2.b) The network structure of the CMixer layer is as follows. (Example: ...) Figure 2 As shown in (a), CMixer consists of a complex domain MLP (CMLP) for space mixing, a CMLP for frequency mixing, and several layers of normalization and residual connections.

[0032] 2.c) The CMLP network structure consists of a 2X→S fully connected layer, a GeLU activation function, and an S→2X fully connected layer. X is the size of the last dimension of the input complex data, and S is the hidden layer width, a hyperparameter. The computational flow of CMLP is as follows: Figure 2 As shown in (b) above, for a B1*…*B n The input is X*2 dimensional, first stretched into B1*…*B n *2X dimension, then input a 2X→S fully connected layer, then pass through a GeLU activation function, then input another S→2X fully connected layer, and then reconstruct it into B1*…*B n *2X dimensions. Where S = 128.

[0033] 2.d) The computational structure of the CMixer Layer is as follows: Figure 2 As shown in (a), for an N′ c *N′ t The 2D input is first passed through a layer normalization and CMLP module, and then summed with itself to obtain an N′. c *N′ t The output is 2D. Then, the dimensions are changed to N′. t *N′ c *2D, then through a layer normalization and CMLP module, and summed with itself, to obtain an N′. t *N′ c The output is 2D. Finally, a dimension permutation is performed to obtain N′. c *N′ t *2D output.

[0034] 2.e) The calculation process of CMixer is as follows. The known channel is input first. The fully connected layer is obtained The output, in the input The fully connected layer obtains N′ c *N′ t The output is *2. Then, input a stacked K-layer CMixer layer to obtain N′. c *N′ t The output is *2. Then input 2N′. t →2N t The fully connected layer is used to obtain N′. c *N t The output of *2 is then input as 2N′. c →2N c The fully connected layer yields N c *N t *2 output.

[0035] 3) Using the CSI data collected in step 1), with CSI on a portion of the antenna and carrier as input and the complete MIMO-OFDM CSI as the target output, train the neural network in step 2). The input is known. The channel is N-dimensional, and the output is N. c *N t *2 complete channels. The parameters of the CMixer are continuously trained using gradient descent until convergence;

[0036] 3.a) The loss function (Loss) of a neural network is defined as follows:

[0037]

[0038] Where Θ is the parameter set of the neural network, N is the number of training data, and H is the number of training data. n This is the nth original training data. It is H n The output obtained after inputting into the CMixer neural network.

[0039] 3.b) The neural network was trained using the Adam optimizer. The optimizer optimizes each parameter of the neural network by taking the derivative of the loss function and backpropagating it, performing gradient descent optimization on each parameter until convergence. The learning rate was set to 0.001, and the training was performed for 2000 iterations, with the learning rate decreasing to 1 / 5 of its previous value every 500 iterations after the 500th training iteration.

[0040] 4. Channel H on some antennas and carriers, known through methods such as channel estimation. 0 The input is fed into the trained CMixer model, and the output of CMixer is used as the acquisition result of the complete MIMO-OFDM channel to achieve the acquisition of the complete MIMO-OFDM channel.

[0041] Example 2

[0042] The table below compares the CMixer-based channel mapping method of this invention with other commonly used deep learning-based channel mapping methods in terms of channel quality acquisition. The performance metrics are normalized mean square error and correlation.

[0043]

[0044] Computer simulations show that, as shown in the table, the CMixer-based channel acquisition scheme proposed in this invention achieves more accurate channel acquisition results compared to MLP and Res_CNN-based schemes, resulting in lower NMSE and higher ρ performance index, where ρ represents the average cosine similarity between the mapped channel and the real channel on each carrier. Furthermore, as... Figure 3 As shown, the CMixer-based channel acquisition scheme proposed in this invention obtains a channel that is closer to the real channel than channel acquisition schemes based on MLP and Res_CNN, better preserving the internal characteristics of the channel and achieving higher channel acquisition quality. Therefore, the CMixer-based channel acquisition scheme proposed in this invention provides a highly effective channel acquisition method for MIMO-OFDM systems.

[0045] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A channel acquisition method based on space-frequency correlation, characterized in that, The method includes: S1. Collect user CSI data; S2. Construct a multi-domain fully connected mixer neural network as a mapping network, and randomly initialize the neural network parameters; the network includes a fully connected layer and K CMixer layers; The neural network structure of CMixer is as follows: CMixer consists of a... A fully connected layer, a A fully connected layer, K CMixer layers, and a 2N t → 2N t A fully connected layer and a Composed of fully connected layers; and These are the spatial and frequency domain dimensions of the input channel, respectively, N. t ′ and N c ′ is a structural hyperparameter, N t and N c It represents the size of the spatial and frequency domain dimensions of the output channel, and K is a hyperparameter that determines the depth of the neural network. The network structure of the CMixer layer is as follows: the CMixer consists of a complex domain MLP for space mixing, a CMLP for frequency mixing, and several layers of normalization and residual connections. The CMLP network structure consists of a 2X→S fully connected layer, a GeLU activation function, and an S→2X fully connected layer; X is the size of the last dimension of the input complex-valued data, S is the hidden layer width, and a hyperparameter; the computational flow of CMLP is as follows: for a B1*…*B n The input is X*2 dimensional, first stretched into B1*…*B n *2X dimension, then input a 2X→S fully connected layer, then pass through a GeLU activation function, then input another S→2X fully connected layer, and then reconstruct it into B1*…*B n *2X dimensions; The computational structure of the CMixer layer is as follows: for an N′ c *N′ t The 2D input is first passed through a layer normalization and CMLP module, and then summed with itself to obtain an N′. c *N′ t *2D output; then change the dimensions to N′ t *N′ c *2D, then through a layer normalization and CMLP module, and summed with itself, to obtain an N′. t *N′ c *2D output; finally, perform a dimension permutation to obtain N′. c *N′ t *2D output; The CMixer calculation process is as follows; The known channel is input first. The fully connected layer is obtained The output, in the input The fully connected layer obtains N′ c *N′ t The output of *2; then input the K stacked CMixer layers to obtain N′ c *N′ t *2 output; then input 2N′ t →2N t The fully connected layer is used to obtain N′. c *N t The output of *2 is then input to 2N′. c →2N c The fully connected layer yields N c *N t *2 output; S3. Train the neural network using the collected CSI data. S4. Input the known channels on some antennas and carriers into the trained network model, and output the complete MIMO-OFDM channel acquisition results.

2. The channel acquisition method based on space-frequency correlation according to claim 1, characterized in that, The collected user CSI data specifically includes channels on partial antennas and carriers, as well as channels on complete antennas and carriers.

3. The channel acquisition method based on space-frequency correlation according to claim 1, characterized in that, The loss function of a neural network is defined as follows: Where Θ is the parameter set of the neural network, N is the number of training data, and H is the number of training data. n This is the nth original training data. It is H n The output obtained after inputting into the CMixer neural network.

Images