Cross-frequency-point channel graph extrapolation method, and system

Abstract

Disclosed in the present invention are a cross-frequency-point channel graph extrapolation method, and a system. In the present invention, on the basis of image style transfer technology, the correlation among channel information between different frequency points is ascertained, and a priori knowledge of a channel graph at a certain frequency point is used to estimate channel graphs at other frequency points, thereby realizing cross-frequency-point extrapolation. In the present invention, first, a channel graph in which statistical channel information is stored is modeled as a multi-channel image, and a cross-frequency-point extrapolation problem is combined with an image style transfer task; then, the mining of the correlation between frequency points, the extraction of a priori information features and the mapping of the channel graph are realized by means of designing generative neural networks; a cycle consistency loss is further introduced to perform reciprocal relationship fitting; and finally, fine-grained optimization is performed on an extrapolation result of the channel graph by means of pixel-grid conversion. The present invention can not only realize high-precision extrapolation between any frequency points, but also has a relatively strong generalization capability in different wireless communication scenarios, and can assist a wireless communication system in realizing high-rate transmission.

Description

A method and system for cross-frequency channel spectrum extrapolation Technical Field

[0001] This invention belongs to the field of communication technology and relates to a method and system for cross-frequency channel map extrapolation based on image style transfer. Background Technology

[0002] In B5G / 6G mobile communication systems, multi-band massive MIMO technology can effectively alleviate the problem of spectrum resource scarcity and is expected to improve spectrum resource utilization while providing personalized services to terminals. However, with the continuous increase in the number of antennas on the base station side and the massive growth of access devices, channel information acquisition under multi-band MIMO faces severe challenges such as a surge in pilot overhead and huge training overhead, making it difficult for traditional algorithms to match the ever-evolving new wireless communication scenarios. To address these issues, channel graph technology has emerged as a potential solution.

[0003] A channel map is a database labeled with the physical location coordinates of a terminal, storing channel information corresponding to the location information, such as line-of-sight / non-line-of-sight probability, channel gain, departure angle / arrival angle, and number of paths. This channel information can effectively assist multi-band massive MIMO transmission, avoid complex and repetitive channel information estimation, greatly alleviate pilot pressure, and improve spectrum utilization.

[0004] In recent years, academia has conducted extensive research on the construction of channel maps. However, on the one hand, most current channel maps only store simple, large-scale information, such as path loss and shadowing fading, and cannot comprehensively reflect the channel characteristics corresponding to the current location coordinates, making them difficult to apply to practical communication systems. On the other hand, most current channel map constructions are only for single-frequency systems. If these methods are directly applied to multi-band large-scale MIMO, i.e., constructing a channel map for each frequency point, the computational cost will be enormous and unacceptable. In fact, since the propagation environment (building distribution and vegetation cover, etc.) of each frequency point is basically similar, there is a certain correlation between their channel maps. Using prior information from low-frequency channel maps to assist in the construction of high-frequency channel maps will effectively reduce algorithm complexity and improve accuracy. Some current research has begun to explore the channel correlation between frequencies in multi-band MIMO systems, but the methods used are still limited to traditional numerical methods such as fitting, interpolation, and AR models. Once the physical location of the terminal in the channel map increases, the required algorithm and storage costs will surge, and the extrapolation error will also deteriorate rapidly. Summary of the Invention

[0005] Purpose of the invention: To address the shortcomings of existing technologies, the present invention aims to provide a method and system for cross-frequency channel map extrapolation. This method can extrapolate the channel maps of other frequencies based on prior knowledge of the channel map of a certain frequency by mining the correlation of channel characteristics between frequency points. While ensuring the extrapolation accuracy, it can further reduce the time complexity, algorithm complexity, and storage complexity compared with existing methods.

[0006] Technical Solution: To achieve the above-mentioned objectives, this invention proposes a method and system for cross-frequency channel map extrapolation. It utilizes grid-to-pixel transformation to model the channel map as a multi-channel image, and transforms the cross-frequency channel map extrapolation problem into an image style transfer task. A recurrent consistency generative neural network is constructed to mine inter-frequency correlations, extract prior information features, and learn the channel map mapping. A fine-grained optimization scheme for the channel map based on pixel-to-grid transformation is designed to improve prediction accuracy, thereby achieving cross-frequency channel map extrapolation. The cross-frequency channel map extrapolation method includes the following steps:

[0007] The target region is discretized and statistical channel information is mapped to each grid. Based on grid-to-pixel conversion, the channel map is transformed into a multi-channel image.

[0008] The problem of cross-frequency channel map extrapolation is modeled as an image style transfer task. A first generative network model and a corresponding, inverse second generative network model are designed to realize the mining of correlations between frequency points, the extraction of prior information features, and the mapping of channel maps. The first / second generative network models take the channel map images of the first / second frequency points as input and the channel map images of the second / first frequency points as output targets.

[0009] Supervised network training is performed on the dataset using the base loss functions of two generative network models, or by introducing cycle consistency loss on the base loss functions, as the joint loss function for training the paired network models, to obtain the optimal weight parameters of the model.

[0010] Based on two jointly trained generative network models, the complete second / first frequency channel map image is extrapolated using the known first / second frequency channel map image, and the fine-grained optimized second / first frequency channel map is obtained based on pixel-to-grid conversion.

[0011] Preferably, the target region is discretized into a grid along the spatial dimension, and each grid is filled with a data vector representing statistical channel information. The grid-to-pixel transformation operation further subdivides each grid into ξ×ζ sub-grids, i.e., pixels, where each pixel's value inherits from its parent grid and has the same channel information value. Specifically, in some embodiments, when ξ=ζ=1, the grid does not need to be subdivided and directly corresponds to the pixels. Based on this, the channel map is modeled as a multi-channel image.

[0012] Preferably, pixel-level grayscale conversion is performed on each channel of the multi-channel image, and the pixel values ​​are converted to the range of 0 to 1 using the max-min normalization method.

[0013] As a preferred approach, the channel map extrapolation problem is analogous to the image style transfer task. Generative networks are designed based on the data structure characteristics of multi-channel images. The first and second generative network models have the same structure, both being generative architectures based on encoder-decoder, used to compress the channel map images of the first / second frequency points into the latent space and then map them to the channel map images of the second / first frequency points.

[0014] Preferably, the first / second generative network model mainly includes:

[0015] The input module is used to perform data shaping and preliminary feature extraction on the channel spectrum images of the first / second frequency points.

[0016] The downsampling module is used to perform further key feature extraction and dimensionality reduction, compressing the channel spectrum image of the first / second frequency point into a low-dimensional data manifold in the latent space Z;

[0017] The residual module is used to mine the correlation between different frequency points and learn the mapping relationship of key features between different frequency points;

[0018] The upsampling module is used to restore the data manifold on the latent space Z to the channel spectrum image of the second / first frequency point;

[0019] The output module is used to perform data shaping on the channel spectrum image of the second / first frequency point to realize cross-frequency channel spectrum extrapolation.

[0020] Preferably, the input module consists of a convolutional layer, a group normalization layer, and a ReLU activation function layer; the downsampling module consists of four encoding layers, each including a convolutional layer, a group normalization layer, and a ReLU activation function layer; the residual module consists of three residual blocks, each including a convolutional layer, an instance normalization layer, and a ReLU activation function layer; the upsampling module consists of four decoding layers, each including a deconvolutional layer, a group normalization layer, and a ReLU activation function layer; and the output module consists of a deconvolutional layer, a group normalization layer, and a Tanh activation function layer.

[0021] Preferably, based on the two basic loss functions, a cycle consistency loss with weights adaptively varying with the number of iteration steps is introduced to fit a reciprocal relationship between cross-frequency channel spectra. The joint loss function for training the pairwise network model is: L(Θ,Θ')=L basic (Θ)+L basic(Θ')+L wcyc (Θ,Θ′). The first and second terms represent the basic loss function of the generative network model, which can be expressed as follows: The last item L wcyc (Θ,Θ') represents the cycle consistency loss function with weights that adaptively vary with the number of iterations, and can be expressed as: Where i,j represent different frequency points, I i I represents the channel spectrum image at the i-th frequency point. j Represents the channel spectrum image at the j-th frequency point; Ψ i,j Indicates from I i To I j The channel spectrum extrapolation network, where Θ is the set of weight parameters for this network; Ψ j,i Indicates from I j To I i A channel map extrapolation network is used to fit Ψ. i,j The reciprocal function of Θ′, where Θ′ is the set of weight parameters for the network; f(ω) is the adaptive weight function of the cycle consistency loss, where ω is the normalized iteration number; E{·} represents the expectation function, ||·|| F This represents the Frobenius norm.

[0022] Preferably, the pixel-to-grid conversion operation utilizes an integrated AI approach. It obtains the second / first frequency channel information value at the pixel level by parallel prediction of multiple pixels under the parent grid point. Then, it averages these pixel values ​​using an average pooling operation, and the resulting data is the second / first frequency channel information value of the parent grid point. Specifically, in some embodiments, when ξ = ζ = 1, this fine-grained optimization scheme can be omitted, and the results will remain consistent.

[0023] The present invention also provides a computer system, including a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor, wherein the computer program / instructions, when executed by the processor, implement the steps of the cross-frequency channel map extrapolation method described above.

[0024] The present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the cross-frequency channel map extrapolation method described above.

[0025] Beneficial Effects: Compared with existing technologies, this invention transforms the cross-frequency channel map extrapolation problem into an image style transfer problem. Based on an encoder-decoder generative neural network architecture, it completes the mining of correlations between frequency points, extraction of prior information features, and mapping of channel maps. A cycle consistency loss is introduced to fit the reciprocal relationship between cross-frequency channel maps. Finally, through a pixel-grid-based fine-grained optimization scheme, the extrapolation of the channel map of another frequency point is completed with the assistance of prior knowledge from one frequency point. Furthermore, this invention designs a weight for the cycle consistency loss that adaptively changes with the number of iterations to balance its ratio with the basic loss function. The proposed cross-frequency channel map extrapolation method not only achieves high-precision extrapolation between arbitrary frequency points but also has strong generalization ability in different wireless communication scenarios. It can assist wireless systems in achieving high-speed transmission, avoid complex and repetitive channel information estimation, reduce pilot overhead, and improve spectrum utilization. Attached Figure Description

[0026] Figure 1 is a flowchart of channel map cross-frequency point extrapolation according to an embodiment of the present invention;

[0027] Figure 2 is a schematic diagram of a generative network in an embodiment of the present invention;

[0028] Figure 3 is a schematic diagram of the cycle consistency characteristic in an embodiment of the present invention;

[0029] Figure 4 is a schematic diagram of the channel spectrum extrapolation accuracy at different frequency points in an embodiment of the present invention;

[0030] Figure 5 is a schematic diagram of the sum rate of channel map-assisted large-scale MIMO transmission in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0032] As shown in Figure 1, the cross-frequency channel map extrapolation method disclosed in this embodiment of the invention mainly includes the following steps:

[0033] Step S1: Discretize the target area and map the statistical channel information to each grid. Based on the grid-to-pixel conversion, convert the channel map into a multi-channel image.

[0034] In this embodiment, channel modeling is first performed using multi-band massive MIMO as a typical example of a wireless communication system:

[0035] Within the target area, there are N frequency points F = {f1, f2, ..., f...} on the base station side. N Coordinated transmission, each frequency point is equipped with M antennas. iA uniform area array (i = 1, 2, ..., N). The user side is equipped with a single antenna at each frequency point, and its physical location coordinates are denoted by q. Then, f i Channel H between the base station and the user located at frequency q q,i It can be written as

[0036] in For complex channel gain, For the rudder vector, θ and These are the elevation and azimuth angles of the departure angle, respectively. Then θ and... Discretize into and And guarantee make

[0037] Then (1) can be written as H q,i =A i β q,i Assuming that the channel gain is uncorrelated in different angular directions, i.e.

[0038] Where δ(·) represents the Dirac function, S q,i (k) is used to describe the power distribution in the beam domain, then the channel spatial covariance matrix R i (q) can be written as

[0039] in Let A represent the channel beam domain covariance matrix. When the number of antennas in a massive MIMO approaches infinity, A... i This will become a 2D-DFT matrix, therefore formula (5) can be seen as an expression for R. i The eigenvalue decomposition of (q) only requires obtaining its eigenvalue sequence. The channel covariance matrix can then be calculated.

[0040] The above eigenvalue sequence This is the data vector representing statistical channel information derived in this embodiment. It is understood that the present invention is not limited to this specific channel map representation. In other embodiments, any data vector that can sufficiently represent statistical channel information can be selected or derived, such as channel gain, path loss, angle information, delay information, channel covariance matrix, etc.

[0041] Secondly, based on the above discussion, a multi-frequency channel map model is established and the cross-frequency channel map extrapolation problem is characterized:

[0042] The actual environment map of size X×Y is gridded, where each grid has a length of Δx and a width of Δy. The target region is then discretized into L grids. x column, L y Draw a grid. For each grid, assign the first frequency point (set to f in this example) at the corresponding position. i The sequence of eigenvalues ​​of the channel covariance matrix of the i-th frequency point is arranged in descending order and stored in a grid, forming the basic model of the channel spectrum of the i-th frequency point, denoted as . Subsequently, a grid-to-pixel transformation operation is introduced, subdividing each grid into ξ×ζ sub-grids, i.e., pixels. Each pixel inherits its value from its parent grid and has the same channel information value. In this embodiment, pixel-level grayscale transformation is performed on each channel of the multi-channel image, and the pixel values ​​are transformed to the range of 0 to 1 using a max-min normalization method. Thus, the original channel map... It is modeled as a multi-channel image Among them I i The (h,w,m) i ) elements are expressed as

[0043] in This represents the floor function. Note that when ξ = ζ = 1, the grid does not need to be subdivided and directly corresponds to pixels. Due to the similarity of signal propagation environments (building distribution, vegetation cover, etc.), there is a correlation between the statistical channel information of each frequency point, and there is a corresponding mapping relationship between the channel maps. The prior knowledge of the channel map of the first frequency point can be used to extrapolate the second frequency point (let's call it f in this example). j The channel spectrum is defined by the frequency point (i.e., the j-th frequency point). Therefore, the channel spectrum extrapolation problem is transformed into an image style transfer problem, which utilizes the correlation between the input and output domains to achieve extrapolation, i.e., finding the I... i with I j Mapping function Ψ between i,j To make it satisfy Ψ i,j :I i →I j (7)

[0044] Among them Ψ i,j This includes a reciprocal relationship between i and j, because i and j can take any values. Considering Ψ... i,j Difficult to express analytically, we employ deep learning methods to fit it; therefore, the channel spectrum extrapolation problem can be characterized as follows:

[0045] Among them Ψ j,i Used to fit Ψ i,jThe reciprocal function is used to ensure that the channel spectrum between frequency points satisfies a one-to-one correspondence.

[0046] Step S2: Model the cross-frequency channel map extrapolation problem as an image style transfer task, and design a first generative network model and a corresponding, inverse second generative network model to realize the mining of correlations between frequency points, the extraction of prior information features, and the mapping of channel maps.

[0047] Specifically, the network Ψ is designed based on the characteristics of the channel graph extrapolation problem. i,j With Ψ j,i :

[0048] Based on the characteristics of the image style transfer problem, a generative network based on encoder-decoder is designed as shown in Figure 2. This network is similar to an "image converter," which compresses the original data into the latent space Z and then uses the correlation between multiple domains to map the data manifold to the target distribution. To achieve this task, a feature extraction layer is designed as the core component. This component first contains a convolutional layer to extract the key features of the input data. Let the input data be X, then the output data X1 can be represented as X1 = Conv2(X) (9)

[0049] Here, Conv2(·) represents the convolution operation. This convolutional layer can be configured with convolutional kernels of different sizes, utilizing different receptive fields to achieve different levels of feature extraction. A grouping normalization layer is then connected to shape the data. It's important to note that the grouping normalization operation GN(·) normalizes the data by group; it is independent of batch size, which is beneficial for the stability of model training. Let the output be X2, then the g-th group of X2 can be represented as...

[0050] Where μ g and σ g These represent [X1] respectively. g The mean and standard deviation, γ g and b g This represents the learnable parameters, where ε is a small constant used to enhance the stability of model training. Finally, an activation function layer is connected; the selection of a non-linear activation function enhances the model's generalization ability. Based on the above feature extraction components, network Ψ... i,j For example, the network is divided into five modules: input module, downsampling module, residual module, upsampling module and output module.

[0051] 1) Input Module: This module consists of a feature extraction component. The convolutional layers have large kernel sizes, allowing for data shaping and initial feature extraction through a larger receptive field. The activation function is ReLU, which sets the negative values ​​to zero. For the input f... iChannel spectrum image I at frequency point i The output of the input module can be represented as: I i =ReLU(GN(Conv2(I) i (11)

[0052] 2) Downsampling Module: This module consists of several feature extraction components. The kernel size of the convolutional layers is smaller than that used in the input module, allowing for the extraction of more detailed key data features. Simultaneously, this module reduces the dimensionality of the data, compressing it to the latent space Z, thus obtaining a low-dimensional manifold representation in latent space Z. This operation is similar to the encoding process, using the ReLU activation function, whose output is represented as I. i =[ReLU(GN(Conv2(I i ′)))] n (12)

[0053] in[·] n This indicates a loop of n times.

[0054] 3) Residual Module: This module consists of several standard residual blocks, with a fixed kernel size of 3×3 and a fixed stride of 1. By stacking multiple residual blocks, the network can be further deepened, improving the efficiency of feature extraction and the ability to mine correlations, while preventing gradient vanishing and gradient exploding problems, thus avoiding network degradation. Its output is represented as I. i "′=[Res(I i ")] n′ (13)

[0055] Where Res(·) represents residual block operation, [·] n′ This indicates a loop n' times.

[0056] 4) Upsampling Module: This module consists of several feature extraction components, and its structure completely matches that of the downsampling module. The convolutional layers employ deconvolution operations, and the kernel size is the same as that used in the downsampling module. It is used to progressively recover the low-dimensional data manifold in the latent space Z, mapping the data to f. j The channel spectrum data distribution at frequency points. This operation is similar to the decoding process, selecting the activation function ReLU, whose output is represented as I. j '=[ReLU(GN(TransConv2(I i "′)))] n (14)

[0057] TransConv2(·) represents the deconvolution operation.

[0058] 5) Output Module: This module consists of a feature extraction component whose structure perfectly matches that of the input module. The convolutional layers employ deconvolution operations with kernels of the same size as those used in the input module. This is used to reconstruct the data structure and extrapolate to obtain the final f. j Channel spectrum data for each frequency point. It's important to note that the activation function here is Tanh, which maps the data to a range of -1 to 1, effectively shaping the data. Based on this, the output module's result can be expressed as...

[0059] in It is the final output of the entire generative network, i.e., f j Coarse-grained extrapolation values ​​of frequency point channel spectrum images.

[0060] Network Ψ j,i With Ψ i,j The results are consistent and will not be repeated. At this point, the extrapolation of the cross-frequency channel map is preliminarily completed, for the aforementioned network Ψ. i,j Fitting network Ψ with reciprocal function j,i Its basic loss function is expressed as

[0061] Step S3: Based on the loss function trained by the pairwise network model, perform supervised network training on the dataset to obtain the optimal weight parameters of the model.

[0062] In this embodiment, based on the one-to-one mapping property of cross-frequency channel maps, a cycle consistency loss is introduced where the weights adaptively change with the number of iteration steps:

[0063] Despite Ψ i,j and Ψ j,i Can be implemented separately and The mapping, but and I j Equality does not guarantee and I i The consistency is ensured because the network learns essentially a data distribution that satisfies additive closure over the data domain. Therefore, a cyclic consistency property, as shown in Figure 3, is introduced to force this consistency. and I i Keep it the same, that is

[0064] Based on formula (18), design the cycle consistency loss function L. cyc (Θ,Θ')

[0065] It is important to note that after introducing the cycle consistency loss, Ψ i,j and Ψ j,i It will satisfy the following properties

[0066] in I represents the composition of functions. i,i and I j,j These are two identity transformations. This also means that Ψ i,j and Ψ j,i The combination of these elements will structurally form an autoencoder, but while a regular autoencoder involves a transformation from the i-domain to the latent space and then back to the i-domain, this composite structure involves a transformation from the i-domain to the j-domain and then back to the i-domain. Furthermore, weights that adaptively change with the number of iterations are designed for the above cycle consistency loss to balance its proportion with the base loss function. In this invention, a feasible approach is to design the weights as a function f(ω) related to the number of iterations, where ω is the normalized number of iterations. Note that f(ω) can be 0 here; when it is 0, it is equivalent to not introducing a cycle consistency loss function. In summary, the cycle consistency loss function with weights adaptively changing with the number of iterations can be expressed as L... wcyc (Θ,Θ')=f(ω)L cyc (Θ,Θ′) (21)

[0067] Step S4: Based on the two jointly trained generative network models, extrapolate the complete second / first frequency channel map image using the known first / second frequency channel map image, and obtain the fine-grained optimized second / first frequency channel map based on pixel-to-grid conversion.

[0068] In this embodiment, the pixel-grid optimization scheme is designed as follows:

[0069] The input data is subjected to min-max normalization and then fed into the trained model to obtain f. j Coarse-grained extrapolation values ​​of frequency channel spectrum images It is important to note that at this time The elements in the graph represent pixels rather than grid points, which cannot correspond to the channel map model, thus requiring pixel-to-grid conversion. Since each grid point is subdivided into ξ×ζ pixels during the grid-to-pixel conversion, this invention employs a simpler method: average pooling, to integrate and restore the pixels. The kernel size for this average pooling operation is defined as ξ×ζ, and the restored channel map is... but The (l) y ,l x ,m j ) elements can be represented as

[0070] In other words, the channel information value of each grid point is the average of the pixel values ​​in its sub-pixels. In essence, the grid-to-pixel transformation involves replicating a single grid into ξ×ζ copies, allowing the network to make parallel predictions on multiple copies simultaneously, and then integrating the results using the pixel-to-grid transformation. Although extrapolation errors are unavoidable in neural networks, this strategy effectively mitigates the impact of extreme errors, reduces model variance, and improves the model's generalization ability. This concept is similar to the idea of ​​neural network ensembles. Specifically, when ξ=ζ=1, executing the pixel-to-grid optimization scheme has the same effect as not executing it; omitting this fine-grained optimization scheme will still yield consistent results.

[0071] To enable those skilled in the art to better understand the present invention, the performance results of the cross-frequency channel map extrapolation method in this embodiment under several specific system configurations are given below.

[0072] The target area size is set to 32×32m. 2 The grid width and height are Δx = Δy = 2, and ξ and ζ in the grid-to-pixel conversion are both set to 2. The communication scenario is set to "38.901_UMI_LOS" supported by QuaDRiGa, where f1 = 2GHz and f2 = 3.9 / 5 / 24GHz are selected for simulation verification on the base station side, with antenna numbers M1 = 8×8 and M2 = 10×10, respectively. For the adaptive weighting of the cyclic consistency loss function, f(ω) = ω is selected. To evaluate the accuracy of channel spectrum extrapolation, the Normalized Mean Square Error (NMSE) is defined as the metric, and its mathematical expression is:

[0073] Where i∈{1,2}, [G i ] m and G i and The grayscale image of the m-th channel is shown in Figure 4. Figure 4 illustrates the cross-frequency channel extrapolation performance of the proposed scheme (named CF-CGN) and two other baseline schemes at different frequencies, including directions from f1→f2 and f2→f1. Baseline algorithm 1 is a linear interpolation extrapolation method, and baseline algorithm 2 is a cyclic GAN extrapolation method. Note that the selection of f2 includes general Sub-6GHz frequencies, Wi-Fi frequencies, and millimeter-wave frequencies, fully demonstrating the adaptability and generalization ability of the proposed scheme. It can be seen that the NMSE of the cross-frequency channel extrapolation results increases as f2 increases from 3.6GHz to 24GHz. In fact, the greater the difference between two frequencies, the less similar their propagation characteristics become, weakening the correlation of the channel spectra at the corresponding frequencies, thus leading to increased extrapolation errors. Regardless of the frequency difference, the extrapolation performance of the proposed scheme outperforms baseline schemes based on interpolation and CycleGAN, confirming its superior performance in cross-frequency channel spectrum extrapolation. Furthermore, Figure 4 shows the results of ablation experiments, comparing the scheme with a cyclic consistency loss function but without fine-grained optimization (i.e., CF-CGN without rrefining scheme, equivalent to ξ=ζ=1) and the scheme without a cyclic consistency loss function but with fine-grained optimization (i.e., CF-CGN without L). cyc The extrapolation error is equivalent to the extrapolation error in the two cases of f(ω)=0. The results show that introducing the cycle-consistent loss function and the fine-grained optimization scheme can improve the extrapolation accuracy to a certain extent.

[0074] To further demonstrate the performance of the high-frequency channel map extrapolated by the proposed scheme in large-scale MIMO transmission, the obtained channel map is applied to a real system and its transmission rate is compared. Considering the l-th user in the target area, its transmission rate can be expressed as...

[0075] Where H l P represents the channel matrix between the base station and the user. l V is the precoding matrix. l The covariance matrix representing the total interference, including additive white Gaussian noise and inter-user interference, is expressed as follows:

[0076] in L represents the variance of Gaussian noise. total This represents the total number of all users. Based on formulas (24) and (25), the transmission and rate can be expressed as...

[0077] Figure 5 illustrates the user and rate performance of the high-frequency channel map extrapolated by the proposed scheme in massive MIMO transmission, with communication scenarios set as "38.901_UMI_LOS" and "38.901_UMI_NLOS". It can be seen that in the LOS scenario, the channel map of the proposed scheme can assist multi-band massive MIMO systems in achieving a sum and rate closest to perfect channel information. In the NLOS scenario, although the scatterer distribution is more complex and the channel map extrapolation accuracy decreases, resulting in lower user and rate performance compared to perfect channel information, the proposed scheme still significantly outperforms the two baseline algorithms.

[0078] An embodiment of the present invention discloses a computer system, including a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor. When the computer program / instructions are executed by the processor, they implement the steps of the cross-frequency channel map extrapolation method.

[0079] An embodiment of the present invention discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the cross-frequency channel map extrapolation method.

[0080] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways without departing from the spirit and scope of this application. The current embodiments are merely exemplary examples and should not be considered limiting, nor should the specific content given limit the purpose of this application. For example, the communication model can be modified, the network structure can be changed, and some features can be ignored or not implemented.

[0081] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for extrapolating cross-frequency channel maps, characterized in that, Includes the following steps: The target region is discretized and statistical channel information is mapped to each grid. Based on grid-to-pixel conversion, the channel map is transformed into a multi-channel image. The problem of cross-frequency channel map extrapolation is modeled as an image style transfer task. A first generative network model and a corresponding, inverse second generative network model are designed to realize the mining of correlations between frequency points, the extraction of prior information features, and the mapping of channel maps. The first / second generative network model takes the first / second frequency point channel spectrum image as input and the second / first frequency point channel spectrum image as output target; Supervised network training is performed on the dataset using the base loss functions of two generative network models, or by introducing cycle consistency loss on the base loss functions, as the joint loss function for training the paired network models, to obtain the optimal weight parameters of the model. Based on two jointly trained generative network models, the complete second / first frequency channel map image is extrapolated using the known first / second frequency channel map image, and the fine-grained optimized second / first frequency channel map is obtained based on pixel-to-grid conversion.

2. The method for cross-frequency channel map extrapolation according to claim 1, characterized in that, The target region is discretized into a grid along the spatial dimension, and a data vector that can represent statistical channel information is filled into each grid. The grid-to-pixel transformation operation is used to subdivide each grid into ξ×ζ subgrids, each subgrid corresponding to a pixel. The pixel value of each pixel is inherited from its parent grid and has the same channel information value. When ξ=ζ=1, the grid does not need to be subdivided and directly corresponds to the pixel.

3. The method for extrapolating cross-frequency channel maps according to claim 1, characterized in that, Pixel-level grayscale conversion is performed on each channel of the multi-channel image, and the pixel values ​​are converted to the range of 0 to 1 using the max-min normalization method.

4. The method for cross-frequency channel map extrapolation according to claim 1, characterized in that, By analogy with the problem of channel map extrapolation, an image style transfer task is designed based on the data structure characteristics of multi-channel images. The first and second generative network models have the same structure, both of which are generative architectures based on encoder-decoder, used to compress the channel map images of the first / second frequency points into the latent space and then map them to the channel map images of the second / first frequency points.

5. The method for extrapolating cross-frequency channel maps according to claim 1, characterized in that, The two paired generative network models mainly include: The input module is used to perform data shaping and preliminary feature extraction on the channel spectrum images of the first / second frequency points. The downsampling module is used to perform further key feature extraction and dimensionality reduction, compressing the channel spectrum image of the first / second frequency point into a low-dimensional data manifold in the latent space Z; The residual module is used to mine the correlation between different frequency points and learn the mapping relationship of key features between different frequency points; The upsampling module is used to restore the data manifold on the latent space Z to the channel spectrum image of the second / first frequency point; The output module is used to perform data shaping on the channel spectrum image of the second / first frequency point to realize cross-frequency channel spectrum extrapolation.

6. The method for cross-frequency channel map extrapolation according to claim 5, characterized in that, The input module consists of convolutional layers, group normalization layers, and ReLU activation function layers; the downsampling module consists of four encoding layers, each including a convolutional layer, a group normalization layer, and a ReLU activation function layer; the residual module consists of three residual blocks, each including a convolutional layer, an instance normalization layer, and a ReLU activation function layer; the upsampling module consists of four decoding layers, each including a deconvolutional layer, a group normalization layer, and a ReLU activation function layer; and the output module consists of a deconvolutional layer, a group normalization layer, and a Tanh activation function layer.

7. The method for extrapolating cross-frequency channel maps according to claim 1, characterized in that, The joint loss function is: L(Θ,Θ')=L basic (Θ)+L basic (Θ')+L wcyc (Θ,Θ'); where: Where i,j represent different frequency points, I i I represents the channel spectrum image at the i-th frequency point. j Represents the channel spectrum image at the j-th frequency point; Ψ i,j Indicates from I i To I j The channel graph extrapolation network, where Θ is the set of weight parameters for this network; Ψ j,i Indicates from I j To I i A channel map extrapolation network is used to fit Ψ. i,j The reciprocal function of Θ′, where Θ′ is the set of weight parameters for the network; f(ω) is the adaptive weight function of the cycle consistency loss, where ω is the normalized iteration number; E{·} represents the expectation function, ||·|| F This represents the Frobenius norm.

8. The method for extrapolating cross-frequency channel maps according to claim 1, characterized in that, The pixel-to-grid conversion operation utilizes the concept of integrated AI. It obtains the second / first frequency channel information value at the pixel level by predicting multiple pixels under the parent grid in parallel. Then, it averages the above pixel values ​​through an average pooling operation. The resulting data is the second / first frequency channel information value of the parent grid. When the grid and pixels are directly one-to-one, fine-grained optimization schemes are omitted.

9. A computer system comprising a memory, a processor, and computer programs / instructions stored in the memory and executable on the processor, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the cross-frequency channel map extrapolation method according to any one of claims 1-8.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the cross-frequency channel map extrapolation method according to any one of claims 1-8.

Images