Cross-frequency-point channel spectrum extrapolation method and system

Abstract

The invention discloses a cross-frequency-point channel spectrum extrapolation method and a cross-frequency-point channel spectrum extrapolation system. Based on an image style migration technology, channel information correlation among different frequency points is ascertained, and channel maps of other frequency points are estimated by utilizing priori knowledge of a channel map of a certain frequency point, so that cross-frequency-point extrapolation is realized. The method comprises the following steps: firstly, modeling a channel map with statistical channel information into a multi-channel image, and combining a cross-frequency point extrapolation problem with an image style migration task; secondly, mining of correlation between frequency points, extraction of prior information features and mapping of a channel map are achieved by designing a generative neural network; further introducing cyclic consistency loss to carry out reciprocal relation fitting; and finally, carrying out fine-grained optimization on an extrapolation result of the channel map through pixel-lattice point conversion. The method not only can realize high-precision extrapolation between any frequency points, but also has relatively strong generalization ability in different wireless communication scenes, and can assist a wireless communication system to realize high-speed transmission.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology and relates to a cross-frequency channel spectrum extrapolation method and system based on an image style migration method. Background Art

[0002] In B5G / 6G mobile communication systems, multi-band massive MIMO technology can effectively alleviate the shortage of spectrum resources, and is expected to provide personalized services to terminals while improving the utilization of spectrum resources. However, with the continuous increase in the number of antennas on the base station side and the massive growth of access devices, the acquisition of channel information under multi-band MIMO faces severe challenges such as a surge in pilot overhead and huge training overhead. Traditional algorithms are difficult to match the evolving new wireless communication scenarios. To address the above problems, channel graph technology has become a potential solution.

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

[0004] In recent years, the academic community has conducted a lot of research on the construction of channel maps. However, on the one hand, most of the current channel maps only store some simple large-scale information, such as path loss, shadow fading, etc., which cannot fully reflect the channel characteristics corresponding to the current position coordinates and are difficult to apply to actual communication systems. On the other hand, most of the current channel maps are only constructed for single-frequency systems. If these methods are directly used for multi-band large-scale MIMO, that is, a channel map is constructed for each frequency point, the computational overhead will be huge and unacceptable. In fact, since the propagation environment of each frequency point (building distribution and vegetation coverage, etc.) is basically similar, there is a certain correlation between their channel maps. Using the prior information of the low-frequency channel map to assist the construction of the high-frequency channel map will effectively reduce the algorithm complexity and improve the accuracy. At present, some studies have begun to try to explore the channel correlation between each frequency point in the multi-band MIMO system, but the methods used are still limited to traditional numerical methods such as fitting, interpolation, and AR models. Once the physical position of the terminal in the channel map increases, it will lead to a surge in the required algorithm overhead and storage overhead, and the extrapolation error will also deteriorate rapidly. Summary of the invention

[0005] Purpose of the invention: In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method and system for extrapolating channel spectra across frequency points, which can mine the correlation of channel characteristics between frequency points and extrapolate the channel spectra of other frequency points based on the prior knowledge of the channel spectra of a certain frequency point. While ensuring the accuracy of the extrapolation, the time complexity, algorithm complexity and storage complexity can be further reduced compared with the existing methods.

[0006] Technical solution: In order to achieve the above-mentioned invention purpose, the present invention proposes a cross-frequency channel spectrum extrapolation method and system, which uses grid-pixel conversion to model the channel spectrum as a multi-channel image, and converts the cross-frequency channel spectrum extrapolation problem into an image style transfer task, constructs a cycle-consistent generative neural network to mine the correlation between frequency points, extracts prior information features, learns channel spectrum mapping, and designs a channel spectrum fine-grained optimization scheme based on pixel-grid conversion to improve the prediction accuracy, thereby realizing the extrapolation of the cross-frequency channel spectrum. The cross-frequency channel spectrum extrapolation method includes the following steps:

[0007] Discretize the target area and correspond the statistical channel information to each grid. Based on the grid-pixel conversion, convert the channel map into a multi-channel image.

[0008] The cross-frequency channel map extrapolation problem is modeled as an image style transfer task, and a first generative network model and a corresponding and inverse second generative network model are designed to mine the correlation between frequencies, extract prior information features, and map the channel map; the first / second generative network model takes the first / second frequency channel map image as input and the second / first frequency channel map image as output target;

[0009] Using the basic loss function of the two generative network models, or introducing the cycle consistency loss on the basic loss function as the joint loss function for the training of the paired network model, supervised network training is performed on the data set to obtain the optimal weight parameters of the model;

[0010] Based on two paired generative network models that have been jointly trained, the known first / second frequency channel map images are used to extrapolate the complete second / first frequency channel map images, and the fine-grained optimized second / first frequency channel maps are obtained based on pixel-grid conversion.

[0011] Preferably, the target area is discretized into a grid along the spatial dimension, and a data vector that can characterize the statistical channel information is filled in each grid. Each grid is subdivided into ξ×ξ subgrids, i.e., pixels, through the grid-pixel conversion operation, wherein the pixel value of each pixel point is inherited from its parent grid and has the same channel information value. In particular, in some embodiments, when ξ=ζ=1, the grid does not need to be subdivided and directly corresponds to the pixel. Based on this, the channel map is modeled as a multi-channel image.

[0012] Preferably, a grayscale conversion at the pixel level is performed on each channel of the multi-channel image, and the pixel value is converted to a range of 0 to 1 using a maximum-minimum normalization method.

[0013] Preferably, the channel map extrapolation problem is compared with the image style transfer task, and a generative network is designed according to the data structure characteristics of the multi-channel image. The first generative network model and the second generative network model have the same structure, and both are generative architectures based on encoding-decoding, which are used to compress the first / second frequency point channel map image into a latent space, and then map it to the second / first frequency point channel map image.

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

[0015] An input module, used for performing data shaping and preliminary feature extraction on the input channel spectrum image of the first / second frequency point;

[0016] Downsampling module, used for further key feature extraction and dimensionality reduction, compressing the channel spectrum image of the first / second frequency point to 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] An 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 is composed of a convolution layer, a group normalization layer and a Relu activation function layer; the downsampling module is composed of four encoding layers, each encoding layer includes a convolution layer, a group normalization layer and a Relu activation function layer; the residual module is composed of three residual blocks, each residual block includes a convolution layer, an instance normalization layer and a Relu activation function layer; the upsampling module is composed of four decoding layers, each decoding layer includes a deconvolution layer, a group normalization layer and a Relu activation function layer; the output module is composed of a deconvolution layer, a group normalization layer and a Tanh activation function layer.

[0021] As a preference, on the basis of the two basic loss functions, a cycle consistency loss whose weights change adaptively with the number of iterations is introduced to fit a reciprocal relationship between the cross-frequency channel spectra, and the joint loss function of the paired network model training is: L(Θ,Θ')=L basic (Θ)+L basic(Θ')+L wcyc (Θ,Θ'). The first and second terms represent the basic loss functions of the generative network model, which can be expressed as The last item L wcyc (Θ,Θ') represents the cycle consistency loss function that introduces weights that change adaptively with the number of iterations, which can be expressed as Among them, i,j represent different frequency points, I i represents the channel spectrum image at the i-th frequency point, I j Represents the channel spectrum image at the jth frequency point; Ψ i,j Indicates that from I i To I j The channel spectrum extrapolation network, Θ is the network weight parameter set; ψ j,i Indicates that from I j To I i The channel spectrum extrapolation network is used to fit ψ i,j The reciprocal function of , Θ' is the network weight parameter set; f(ω) is the adaptive weight function of the cycle consistency loss, ω is the normalized number of iterations; E{·} represents the expected function, ||·|| F represents the Frobenius norm.

[0022] Preferably, the pixel-grid conversion operation utilizes the integrated AI concept, obtains the second / first frequency channel information value at the pixel level by predicting multiple pixels under the parent grid in parallel, and then averages the above pixel values ​​through an average pooling operation, and the obtained data is the second / first frequency channel information value of the parent grid. In particular, in some embodiments, when ζ=ζ=1, the 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 / instruction stored in the memory and executable on the processor, wherein the computer program / instruction, when executed by the processor, implements the steps of the cross-frequency channel spectrum extrapolation method.

[0024] The present invention also provides a computer program product, including a computer program / instruction, which implements the steps of the cross-frequency channel spectrum extrapolation method when executed by a processor.

[0025] Beneficial effects: Compared with the prior art, the present invention transforms the cross-frequency channel map extrapolation problem into an image style transfer problem, and completes the mining of correlations between frequency points, extraction of prior information features and mapping of channel maps based on the encoding-decoding generative neural network architecture, and introduces cycle consistency loss to fit the reciprocal relationship between cross-frequency channel maps. Finally, through a fine-grained optimization scheme based on pixels and grids, the extrapolation of the channel map of another frequency point assisted by the prior knowledge of one frequency point is completed. Furthermore, the present invention also designs a weight for the cycle consistency loss that can adaptively change with the number of iteration steps to balance its ratio with the basic loss function. The proposed cross-frequency channel map extrapolation method can not only achieve high-precision extrapolation between arbitrary frequency points, but also has strong generalization capabilities in different wireless communication scenarios, can assist wireless systems in achieving high-speed transmission, avoid complex and repeated channel information estimation, reduce pilot overhead, and improve spectrum utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of channel spectrum 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 Schematic diagram of cycle-consistent characteristics in an embodiment of the present invention;

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

[0030] Figure 5 A schematic diagram of the sum rate of channel spectrum assisted massive MIMO transmission in an embodiment of the present invention. DETAILED DESCRIPTION

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

[0032] like Figure 1 As shown, a cross-frequency channel spectrum extrapolation method disclosed in an embodiment of the present invention mainly includes the following steps:

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

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

[0035] In the target area, there are N frequency points F = {f1, f2, ..., f N}Cooperative transmission, each frequency point is equipped with M antennas i , (i=1,2,...,N) uniform array. The user side is equipped with a single antenna at each frequency point, and its physical position coordinate is marked as q. Then, f i The channel H between the base station and the user at q on the frequency point q,i It can be written as

[0036]

[0037] in is the complex channel gain, is the rudder vector, θ and are the elevation and azimuth of the departure angle respectively. Then θ and Discretize into and And ensure that θ(k)∈[θ min ,θ max ], make

[0038]

[0039] Then (1) can be written as H q,i =A i β q,i Assume that the channel gains at different angles are independent of each other, that is

[0040]

[0041] 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

[0042]

[0043] in represents the channel beam domain covariance matrix. When the number of massive MIMO antennas tends to infinity, A i will become a 2D-DFT matrix, so formula (5) can be regarded as the i The eigenvalue decomposition of (q) only needs to obtain its eigenvalue sequence The channel covariance matrix can be calculated.

[0044] The above eigenvalue sequence is the data vector characterizing the statistical channel information derived in this embodiment. It is understandable that the scheme of the present invention is not limited to this specific channel spectrum representation form. In other embodiments, any data vector that can fully characterize the statistical channel information can be selected or derived, such as channel gain, path loss, angle information, delay information, channel covariance matrix, etc.

[0045] Secondly, based on the above discussion, a multi-frequency channel spectrum model is established and the cross-frequency channel spectrum extrapolation problem is characterized:

[0046] The actual environment map of size X×Y is gridded, where each grid is Δx long and Δy wide, and the target area is discretized into L x Column, L y For each grid, the first frequency point (set to f in this example) at the corresponding position is i The channel covariance matrix eigenvalue sequence of the i-th frequency point is arranged in descending order and stored in the grid to form the basic model of the channel spectrum of the i-th frequency point, which is recorded as Then, the grid-pixel conversion operation is introduced to subdivide each grid into ξ×ζ subgrids, i.e., pixels, where the pixel values ​​of each pixel are inherited from its parent grid and have the same channel information value. In this embodiment, the grayscale conversion of each channel of the multi-channel image is performed at the pixel level, and the pixel value is converted to the range of 0 to 1 using the maximum-minimum normalization method. In this way, the original channel map It is modeled as a multi-channel image Among them I i The first (h,w,m i ) elements are expressed as

[0047]

[0048] in represents the floor function. Note that when ξ=ζ=1, the grid does not need to be subdivided and directly corresponds to the pixel. Since the signal propagation environment (building distribution, vegetation coverage, etc.) is similar, the statistical channel information between the frequency points is correlated, and there is a corresponding mapping relationship between the channel spectra. The prior knowledge of the first frequency point channel spectrum can be used to extrapolate the second frequency point (set to f in this example). j Therefore, the channel spectrum extrapolation problem is transformed into an image style migration problem, and the extrapolation is achieved by using the correlation between the input domain and the output domain, that is, finding I i with I j The mapping function Ψ between i,j , so that it satisfies

[0049] Ψ i,j :I i→I j (7)

[0050] Among them Ψ i,j It contains a reciprocal relationship between i and j, because i and j can take any value. i,j It is difficult to express it in analytical form. Here, a deep learning method is used to fit it. Therefore, the channel spectrum extrapolation problem can be characterized as

[0051]

[0052] Among them Ψ j,i To fit Ψ i,j The reciprocal function of is used to ensure that the channel spectra across frequency points satisfy a one-to-one correspondence.

[0053] Step S2: Model the cross-frequency channel map extrapolation problem as an image style transfer task, design a first generative network model and a corresponding and inverse second generative network model to mine the correlation between frequencies, extract prior information features, and map the channel map.

[0054] Specifically, the network Ψ is designed based on the characteristics of the channel spectrum extrapolation problem. i,j With j,i :

[0055] According to the characteristics of the image style transfer problem, we design Figure 2 The generative network based on encoding and decoding is shown in Figure 1. This network is similar to an "image converter" that compresses the original data into a 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 a 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 expressed as

[0056] X1=Conv2(X)(9)where Conv2(·) represents the convolution operation. The convolution layer can be configured with convolution kernels of different sizes, and use different sizes of receptive fields to achieve different degrees of feature extraction. Then a group normalization layer is connected to reshape the data. It should be noted that the group normalization operation GN(·) normalizes the data by group, which does not depend on the batch size and is conducive to the stability of model training. Let the output be X2, then the g-th group of data in X2 can be expressed as

[0057]

[0058] where μ g and σ g Respectively represent [X1] g The mean and standard deviation of gand b g represents the learnable parameter, and ε is a small constant used to enhance the stability of model training. Finally, an activation function layer is connected to enhance the generalization ability of the model by selecting a nonlinear activation function. Based on the above feature extraction components, the network Ψ i,j As an example, the network is divided into five modules: input module, downsampling module, residual module, upsampling module and output module.

[0059] 1) Input module: This module consists of a feature extraction component, in which the convolution kernel size of the convolution layer is large, and data shaping and preliminary feature extraction are performed through a large receptive field of view. The activation function is ReLU, and the negative value is set to zero. For the input f i Channel spectrum image of frequency point I i , the output of the input module can be expressed as:

[0060] I i =ReLU(GN(Conv2(I i ))) (11)

[0061] 2) Downsampling module: This module consists of several feature extraction components, in which the convolution kernel size of the convolution layer is smaller than that used in the input module, which is used to further extract more detailed key features of the data. At the same time, this module will reduce the dimension of the data and compress it to the latent space Z, thereby obtaining a low-dimensional data manifold representation in the latent space Z. This operation is similar to the encoding process. The activation function ReLU is selected, and its output is expressed as

[0062] I i = [ReLU(GN(Conv2(I i ')))] n (12)

[0063] in[·] n Indicates loop n times.

[0064] 3) Residual module: This module consists of several standard residual blocks, with the convolution kernel size fixed at 3×3 and the step size fixed at 1. By stacking multiple residual blocks, the network can be further deepened, the efficiency of feature extraction and the ability of correlation mining can be improved, while preventing the gradient vanishing and gradient exploding problems, and avoiding the degradation of the network. Its output is expressed as

[0065] I i ”'=[Res(I i ”)] n' (13)

[0066] Where Res(·) represents the residual block operation, [·] n' Indicates loop n' times.

[0067] 4) Upsampling module: This module consists of several feature extraction components, and its structure is completely matched with the downsampling module. The convolution layer uses deconvolution operation and the convolution kernel is the same size as that used in the downsampling module. It is used to gradually restore the low-dimensional data manifold in the latent space Z and map the data to f j This operation is similar to the decoding process. The activation function ReLU is selected, and its output is expressed as

[0068] I j = [ReLU(GN(TransConv2(I i ”')))] n (14)

[0069] Where TransConv2(·) represents the deconvolution operation.

[0070] 5) Output module: This module consists of a feature extraction component, whose structure completely matches the input module. The convolution layer uses deconvolution operation and the convolution kernel size is consistent with that used in the input module to restore the data structure and extrapolate to obtain the final f j Channel spectrum data of the frequency point. It should be noted that the activation function here is Tanh, which can map the data to the range of -1 to 1 to achieve data shaping. Based on this, the result of the output module can be expressed as

[0071]

[0072] in is the final output of the entire generative network, i.e. f j Coarse-grained extrapolation of frequency channel map images.

[0073] Network j,i With i,j So far, the extrapolation of the cross-frequency channel spectrum is preliminarily completed. i,j and the reciprocal function fitting network Ψ j,i , and its basic loss function is expressed as

[0074]

[0075] Step S3: Based on the loss function of paired network model training, supervised network training is performed on the data set to obtain the optimal weight parameters of the model.

[0076] In this embodiment, based on the one-to-one mapping property requirement of the cross-frequency channel spectrum, a cycle consistency loss is introduced in which the weight changes adaptively with the number of iteration steps:

[0077] Althoughi,j and j,i Can be realized separately and , but and I j The equality of and I i , because what the network learns is essentially a data distribution that satisfies additive closure in the data domain. Therefore, we introduce Figure 3 The cycle consistency property shown here forces and I i Keep the same, that is

[0078]

[0079] Based on formula (18), the cycle consistency loss function L is designed cyc (Θ,Θ')

[0080]

[0081] It should be noted that after introducing the cycle consistency loss, Ψ i,j and j,i Will satisfy the following properties

[0082]

[0083] Where ο represents the composite operation of the function, I i,i and I j,j are two identical transformations. This means that Ψ i,j and j,i The combination will form an autoencoder in structure, but the ordinary autoencoder is a transformation from i domain → latent space → i domain, while this composite structure is a transformation from i domain → j domain → i domain. Furthermore, a weight that can adaptively change with the number of iterations is designed for the above cycle consistency loss to balance its ratio with the basic loss function. In the present invention, a feasible scheme is to design the weight as a function f(ω) related to the number of iterations, where ω is the normalized number of iterations. Note that f(ω) here can be 0. When it is 0, it is equivalent to not introducing a cycle consistency loss function. In summary, the cycle consistency loss function in which the weights adaptively change with the number of iterations can be expressed as

[0084] L wcyc (Θ,Θ')=f(ω)L cyc (Θ,Θ')(21)

[0085] Step S4: Based on the two paired generative network models that have been jointly trained, the known first / second frequency channel map images are used to extrapolate the complete second / first frequency channel map images, and the fine-grained optimized second / first frequency channel maps are obtained based on pixel-grid conversion.

[0086] In this embodiment, the pixel-grid optimization scheme is designed as follows:

[0087] The input data is normalized to the maximum and minimum and fed into the trained model to obtain f j Coarse-grained extrapolation of frequency channel spectrogram images It should be noted that at this time Each element in represents a pixel rather than a grid point, which cannot correspond to the channel map model, so a pixel-grid conversion is required. Since each grid point has been subdivided into ξ×ζ pixels in the grid-pixel conversion operation, a simpler method, namely, the average pooling operation, is adopted in the present invention to integrate and restore the pixels. The convolution kernel size of the average pooling operation is defined as ξ×ζ, and the restored channel map is but The first (l y ,l x ,m j ) elements can be represented as

[0088]

[0089] That is, the channel information value of each grid point is the average of the pixel values ​​in its sub-pixels. In fact, the essence of the grid-pixel conversion is to copy a single grid into ξ×ζ copies, so that the network can make parallel predictions on multiple copies at the same time, and then integrate the results using the pixel-grid conversion. Although extrapolation errors in neural networks are inevitable, this strategy can effectively mitigate the impact of extreme errors, reduce the variance of the model, and improve the generalization ability of the model. This concept is similar to the idea of ​​neural network integration. In particular, when ξ=ζ=1, executing the pixel-grid optimization scheme is the same as not executing it. If the fine-grained optimization scheme is omitted, the results will remain consistent.

[0090] In order to enable those skilled in the art to better understand the solution of the present invention, performance results of the cross-frequency channel spectrum extrapolation method of this embodiment under several specific system configurations are given below.

[0091] Set the target area size to 32×32m 2, the grid length and width are Δx=Δy=2, and ξ and ζ in the grid-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 on the base station side for simulation verification, and the number of antennas is M1=8×8 and M2=10×10, respectively. For the adaptive change weight of the cycle-consistent loss function, f(ω)=ω is selected. In order to evaluate the accuracy of channel spectrum extrapolation, the normalized minimum mean square error (NMSE) is defined as the measurement indicator, and its mathematical expression is

[0092]

[0093] where i∈{1,2}, [G i ] m and Respectively represent G i and The m-th channel grayscale image. Figure 4 The cross-frequency channel spectrum extrapolation performance of the scheme proposed in the present invention (named CF-CGN) and the other two baseline schemes at different frequencies is demonstrated, including the two directions from f1→f2 and f2→f1. Among them, baseline algorithm 1 is linear interpolation extrapolation, namely LinearExtrapolation, and baseline algorithm 2 is cyclic GAN extrapolation, namely CycleGAN. Note that the selection of f2 includes general Sub-6GHz frequencies, Wi-Fi frequencies and millimeter wave frequencies, which fully demonstrates the adaptability and generalization ability of the scheme of the present invention. It can be seen that the NMSE of the cross-frequency channel spectrum extrapolation result increases as f2 increases from 3.6GHz to 24GHz. In fact, the greater the difference between the two frequencies, the less similar their propagation characteristics are, which weakens the correlation of the channel spectra at the corresponding frequencies, thereby increasing the extrapolation error. However, regardless of the difference between the frequencies, the extrapolation performance of the scheme proposed in the present invention is better than the baseline schemes based on interpolation and CycleGAN, which confirms the excellent performance of the scheme in the cross-frequency channel spectrum extrapolation problem. In addition, Figure 4 The results of ablation experiments are also shown, comparing the results of introducing a cycle-consistent loss function without introducing a fine-grained optimization scheme (i.e., CF-CGN without rrefining scheme, equivalent to ξ=ζ=1) and introducing a fine-grained optimization scheme without introducing a cycle-consistent loss function (i.e., CF-CGN without L cyc , which is equivalent to the extrapolation error in both cases of f(ω)=0). The results show that the introduction of cycle-consistent loss function and fine-grained optimization scheme can improve the extrapolation accuracy to a certain extent.

[0094] To further demonstrate the performance of the high-frequency channel map obtained by the proposed solution in large-scale MIMO transmission, the obtained channel map is applied to the actual system and the sum rate is compared. Considering the lth user in the target area, its transmission rate can be expressed as

[0095]

[0096] Among them, H l represents the channel matrix between the base station and the user, P l is the precoding matrix. l The covariance matrix of the total interference, including additive Gaussian white noise and inter-user interference, is written as

[0097]

[0098] in represents the variance of Gaussian noise, L total represents the total number of all users. Based on formula (24) and formula (25), the transmission sum rate can be expressed as

[0099]

[0100] Figure 5 The user and rate performance of the high-frequency channel map obtained by extrapolating the scheme proposed in the present invention in large-scale MIMO transmission is demonstrated, and the communication scenarios are set to "38.901_UMI_LOS" and "38.901_UMI_NLOS". It can be seen that in the LOS scenario, the channel map of the scheme proposed in the present invention can assist the multi-band large-scale MIMO system to achieve the sum rate closest to the perfect channel information. In the NLOS scenario, although the distribution of scatterers is more complex, the accuracy of channel map extrapolation is reduced, and the user sum rate assisted by the channel map is lower than that of the perfect channel information, the performance of the scheme proposed in the present invention can still significantly exceed the two baseline algorithms.

[0101] A computer system disclosed in an embodiment of the present invention includes a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor. When the computer program / instruction is executed by the processor, the steps of the cross-frequency channel spectrum extrapolation method are implemented.

[0102] A computer program product disclosed in an embodiment of the present invention includes a computer program / instruction, and when the computer program / instruction is executed by a processor, the steps of the cross-frequency channel spectrum extrapolation method are implemented.

[0103] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways without exceeding the spirit and scope of this application. The current embodiment is only an illustrative example and should not be used as a limitation. The specific content given should not 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 executed.

[0104] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A cross-frequency channel spectrum extrapolation method, characterized in that: The following steps are involved: Discretize the target area and correspond the statistical channel information to each grid. Based on the grid-pixel conversion, convert the channel map into a multi-channel image. The cross-frequency channel map extrapolation problem is modeled as an image style transfer task. The first generative network model and the corresponding and inverse second generative network model are designed to mine the correlation between frequency points, extract prior information features and map the channel map. The first / second generative network model takes the first / second frequency point channel spectrum image as input and takes the second / first frequency point channel spectrum image as output target; Using the basic loss function of the two generative network models, or introducing the cycle consistency loss on the basic loss function as the joint loss function for the training of the paired network model, supervised network training is performed on the data set to obtain the optimal weight parameters of the model; Based on two paired generative network models that have been jointly trained, the known first / second frequency channel map images are used to extrapolate the complete second / first frequency channel map images, and the fine-grained optimized second / first frequency channel maps are obtained based on pixel-grid conversion.

2. A cross-frequency channel spectrum extrapolation method according to claim 1, characterized in that: The target area is discretized into grids along the spatial dimension, and a data vector capable of representing the statistical channel information is filled into each grid; each grid is subdivided into ξ×ζ subgrids through the grid-pixel conversion operation, each subgrid corresponds to a pixel, wherein the pixel value of each pixel point 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 cross-frequency channel spectrum extrapolation method according to claim 1, characterized in that: A pixel-level grayscale conversion is performed on each channel of the multi-channel image, and the pixel value is converted to the range of 0 to 1 using the maximum-minimum normalization method.

4. The cross-frequency channel spectrum extrapolation method according to claim 1, characterized in that: The channel map extrapolation problem is compared with the image style transfer task, and a generative network is designed according to the data structure characteristics of the multi-channel image. The first generative network model and the second generative network model have the same structure, both of which are generative architectures based on encoding-decoding, and are used to compress the first / second frequency point channel map image into the latent space, and then map it to the second / first frequency point channel map image.

5. The cross-frequency channel spectrum extrapolation method according to claim 1, characterized in that: The two paired generative network models mainly include: An input module, used for performing data shaping and preliminary feature extraction on the input channel spectrum image of the first / second frequency point; Downsampling module, used for further key feature extraction and dimensionality reduction, compressing the channel spectrum image of the first / second frequency point to 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; An 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. A cross-frequency channel spectrum extrapolation method according to claim 5, characterized in that: The input module is composed of a convolution layer, a group normalization layer and a Relu activation function layer; the downsampling module is composed of four encoding layers, each encoding layer includes a convolution layer, a group normalization layer and a Relu activation function layer; the residual module is composed of three residual blocks, each residual block includes a convolution layer, an instance normalization layer and a Relu activation function layer; the upsampling module is composed of four decoding layers, each decoding layer includes a deconvolution layer, a group normalization layer and a Relu activation function layer; the output module is composed of a deconvolution layer, a group normalization layer and a Tanh activation function layer.

7. The cross-frequency channel spectrum extrapolation method 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: Among them, i,j represent different frequency points, I i represents the channel spectrum image at the i-th frequency point, I j Represents the channel spectrum image at the jth frequency point; Ψ i,j Indicates that from I i To I j The channel spectrum extrapolation network is Θ, which is the network weight parameter set; j,i Indicates that from I j To I i The channel spectrum extrapolation network is used to fit Ψ i,j The reciprocal function of , Θ' is the network weight parameter set; f(ω) is the adaptive weight function of the cycle consistency loss, ω is the normalized number of iterations; E{·} represents the expected function, ||·|| F represents the Frobenius norm.

8. The cross-frequency channel spectrum extrapolation method according to claim 1, characterized in that: The pixel-grid conversion operation utilizes the integrated AI concept to obtain the second / first frequency channel information value at the pixel level by parallel prediction of multiple pixels under the parent grid, and then averages the above pixel values ​​through the average pooling operation. The obtained data is the second / first frequency channel information value of the parent grid; when the grid corresponds directly to the pixel one by one, the fine-grained optimization scheme is omitted.

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

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

Images