Channel prediction method and device, computer equipment, storage medium and computer program product

Abstract

The invention relates to a channel prediction method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring a noisy pilot matrix corresponding to a channel to be analyzed in a communication system; inputting the noise-containing pilot frequency matrix into a trained first channel prediction model to obtain a first prediction channel matrix corresponding to the to-be-analyzed channel, and obtaining a second prediction channel matrix corresponding to the to-be-analyzed channel based on the noise-containing pilot frequency matrix through a second channel prediction model or a channel prediction algorithm; and according to a first weight corresponding to the first prediction channel matrix and a second weight corresponding to the second prediction channel matrix, performing fusion processing on the first prediction channel matrix and the second prediction channel matrix to obtain a target prediction channel matrix corresponding to the to-be-analyzed channel. By adopting the method, the accuracy of channel prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a channel prediction method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] With the development of MIMO (Multiple-Input Multiple-Output) technology, how to perform channel prediction becomes crucial to improve the performance of communication systems.

[0003] In traditional technologies, a single prediction method (such as a single model or algorithm) is generally used during channel prediction. However, the channel characteristics of different scenarios vary greatly, and a single prediction method is difficult to cover the physical propagation mechanisms of all scenarios, resulting in low channel prediction accuracy. Summary of the Invention

[0004] Based on this, it is necessary to provide a channel prediction method, apparatus, computer equipment, computer-readable storage medium and computer program product that can improve the accuracy of channel prediction in order to address the above technical problems.

[0005] In a first aspect, the present application provides a channel prediction method, comprising:

[0006] Obtaining a noisy pilot matrix corresponding to a channel to be analyzed in a communication system;

[0007] Inputting the noisy pilot matrix into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and obtaining a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm;

[0008] The first predicted channel matrix and the second predicted channel matrix are fused according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

[0009] In one embodiment, the trained first channel prediction model includes a plurality of trained denoisers and a trained generator in a generative adversarial network;

[0010] Inputting the noisy pilot matrix into the trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed includes:

[0011] input the noisy pilot matrix into the plurality of trained denoiser to obtain a clean pilot image corresponding to the noisy pilot matrix;

[0012] input the clean pilot image into the generator to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

[0013] In one of the embodiments, the inputting the noisy pilot matrix into the plurality of trained denoiser to obtain a clean pilot image corresponding to the noisy pilot matrix comprises:

[0014] splitting the noisy pilot matrix to obtain a first noisy pilot image corresponding to a real part of the noisy pilot matrix and a second noisy pilot image corresponding to an imaginary part of the noisy pilot matrix;

[0015] inputting the first noisy pilot image into the plurality of trained denoiser to obtain a plurality of first clean pilot images corresponding to the first noisy pilot image and inputting the second noisy pilot image into the plurality of trained denoiser to obtain a plurality of second clean pilot images corresponding to the second noisy pilot image;

[0016] fusing the plurality of first clean pilot images to obtain a first target clean pilot image corresponding to the first noisy pilot image and fusing the plurality of second clean pilot images to obtain a second target clean pilot image corresponding to the second noisy pilot image;

[0017] using the first target clean pilot image and the second target clean pilot image as the clean pilot image corresponding to the noisy pilot matrix.

[0018] In one of the embodiments, the inputting the clean pilot image into the generator to obtain a first predicted channel matrix corresponding to the channel to be analyzed comprises:

[0019] inputting the first target clean pilot image and the second target clean pilot image into the generator respectively to obtain a third predicted channel matrix corresponding to the first target clean pilot image and a fourth predicted channel matrix corresponding to the second target clean pilot image;

[0020] combining the third predicted channel matrix and the fourth predicted channel matrix to obtain the first predicted channel matrix corresponding to the channel to be analyzed.

[0021] In one of the embodiments, each trained denoiser corresponds to a noise type;

[0022] Each trained denoiser is obtained by training in the following manner:

[0023] obtaining a sample pilot matrix, splitting the sample pilot matrix to obtain a sample pilot image, and determining a noise type corresponding to the denoiser to be trained; the sample pilot image includes a clean pilot image corresponding to a real part of the sample pilot matrix and a clean pilot image corresponding to an imaginary part of the sample pilot matrix;

[0024] determining a first sample noise and a second sample noise matched with the noise type, performing noise adding on the sample pilot image according to the first sample noise to obtain a first sample noisy pilot image corresponding to the sample pilot image, and performing noise adding on the sample pilot image according to the second sample noise to obtain a second sample noisy pilot image corresponding to the sample pilot image; the first sample noise and the second sample noise correspond to different signal-to-noise ratios;

[0025] performing iterative training on the denoiser to be trained according to the first sample noisy pilot image and the second sample noisy pilot image to obtain the trained denoiser.

[0026] In one embodiment, the performing iterative training on the denoiser to be trained according to the first sample noisy pilot image and the second sample noisy pilot image to obtain the trained denoiser comprises:

[0027] inputting the first sample noisy pilot image into the denoiser to be trained to obtain a denoised pilot image corresponding to the first sample noisy pilot image;

[0028] obtaining a structure loss value and a feature loss value corresponding to the denoiser to be trained according to a difference between the denoised pilot image and the second sample noisy pilot image;

[0029] performing fusion processing on the structure loss value and the feature loss value to obtain a target loss value;

[0030] performing iterative training on the denoiser to be trained according to the target loss value to obtain the trained denoiser.

[0031] In one embodiment, the obtaining a structure loss value and a feature loss value corresponding to the denoiser to be trained according to a difference between the denoised pilot image and the second sample noisy pilot image comprises:

[0032] determining a brightness similarity, a contrast similarity, and a structure similarity between the denoised pilot image and the second sample noisy pilot image;

[0033] Fusing the luminance similarity, the contrast similarity and the structure similarity to obtain a target similarity between the denoised pilot image and the second sample noisy pilot image;

[0034] According to the target similarity, a structure loss value corresponding to the denoiser to be trained is determined.

[0035] In one of the embodiments, the structure loss value and the feature loss value corresponding to the denoiser to be trained are obtained according to the difference between the denoised pilot image and the second sample noisy pilot image, further comprising:

[0036] Feature extraction is respectively performed on the denoised pilot image and the second sample noisy pilot image to obtain a first feature vector corresponding to the denoised pilot image and a second feature vector corresponding to the second sample noisy pilot image;

[0037] According to the difference between the first feature vector and the second feature vector, a feature loss value corresponding to the denoiser to be trained is obtained.

[0038] In one of the embodiments, before the fusion processing of the first prediction channel matrix and the second prediction channel matrix according to the first weight corresponding to the first prediction channel matrix and the second weight corresponding to the second prediction channel matrix to obtain the target prediction channel matrix corresponding to the channel to be analyzed, further comprising:

[0039] Obtaining a reference index corresponding to the noisy pilot matrix;

[0040] According to the comparison result between the reference index and a preset threshold, the first weight corresponding to the first prediction channel matrix and the second weight corresponding to the second prediction channel matrix are determined.

[0041] In one of the embodiments, after the fusion processing of the first prediction channel matrix and the second prediction channel matrix according to the first weight corresponding to the first prediction channel matrix and the second weight corresponding to the second prediction channel matrix to obtain the target prediction channel matrix corresponding to the channel to be analyzed, further comprising:

[0042] According to the target prediction channel matrix, the channel gain and the phase difference from each transmitting antenna to the corresponding receiving antenna are determined; the channel gain is used to represent the attenuation of the transmitting signal of the each transmitting antenna in the transmission process; the phase difference is used to represent the phase change of the transmitting signal of the each transmitting antenna in the transmission process;

[0043] In the three-dimensional map corresponding to the channel to be analyzed, the channel gain and the phase difference from each transmitting antenna to the corresponding receiving antenna are displayed.

[0044] In a second aspect, the present application further provides a channel prediction device, comprising:

[0045] A pilot matrix acquisition module is used to obtain a noisy pilot matrix corresponding to a channel to be analyzed in a communication system;

[0046] a channel matrix prediction module, configured to input the noisy pilot matrix into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and obtain a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm;

[0047] A channel matrix fusion module is used to fuse the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

[0048] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0049] Obtaining a noisy pilot matrix corresponding to a channel to be analyzed in a communication system;

[0050] Inputting the noisy pilot matrix into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and obtaining a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm;

[0051] The first predicted channel matrix and the second predicted channel matrix are fused according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

[0052] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0053] Obtaining a noisy pilot matrix corresponding to a channel to be analyzed in a communication system;

[0054] Inputting the noisy pilot matrix into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and obtaining a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm;

[0055] The first predicted channel matrix and the second predicted channel matrix are fused according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

[0056] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the following steps:

[0057] Obtaining a noisy pilot matrix corresponding to a channel to be analyzed in a communication system;

[0058] Inputting the noisy pilot matrix into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and obtaining a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm;

[0059] The first predicted channel matrix and the second predicted channel matrix are fused according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

[0060] The channel prediction method, device, computer device, storage medium and computer program product described above first acquire a noisy pilot matrix corresponding to a channel to be analyzed in a communication system, then input the noisy pilot matrix into a first channel prediction model trained, obtain a first predicted channel matrix corresponding to the channel to be analyzed, and based on the noisy pilot matrix, obtain a second predicted channel matrix corresponding to the channel to be analyzed through a second channel prediction model or a channel prediction algorithm, and perform fusion processing on the first predicted channel matrix and the second predicted channel matrix according to a first weight corresponding to the first predicted channel matrix and a second weight corresponding to the second predicted channel matrix, to obtain a target predicted channel matrix corresponding to the channel to be analyzed. In this way, in the process of channel prediction, the noisy pilot matrix corresponding to the channel to be analyzed in the communication system is predicted by using two different channel prediction methods, and fusion processing is performed based on the weights corresponding to the obtained predicted channel matrices, so that the target predicted channel matrix corresponding to the channel to be analyzed can be more accurately obtained, which is beneficial to improving the determination accuracy of the target predicted channel matrix, and thus the accuracy of channel prediction is improved. Moreover, by combining different methods for channel prediction, the defect that a single prediction method is difficult to cover all scene physical propagation mechanisms and leads to low accuracy of channel prediction is avoided, and the accuracy of channel prediction is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other related drawings can be obtained by those skilled in the art without creative labor.

[0062] Figure 1 A flowchart of a channel prediction method in an embodiment;

[0063] Figure 2 A flowchart of a channel prediction method in another embodiment;

[0064] Figure 3 A schematic diagram of model training in an embodiment;

[0065] Figure 4 A schematic diagram of digital twinning in an embodiment;

[0066] Figure 5 A flowchart of a channel prediction method in another embodiment;

[0067] Figure 6 A structural block diagram of a channel prediction device in an embodiment;

[0068] Figure 7 Fig. 1 is a schematic diagram of an internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0069] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0070] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of the related data need to comply with relevant regulations.

[0071] In one exemplary embodiment, as shown in Figure 1 A channel prediction method is provided, and the present embodiment is exemplarily described by taking the method applied to a server; it can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is realized through the interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, notebook computers, smart phones and tablet computers; the server can be realized by an independent server or a server cluster composed of multiple servers. The method in the present embodiment includes the following steps:

[0072] In step S101, a noisy pilot matrix corresponding to a to-be-analyzed channel in a communication system is obtained.

[0073] The communication system refers to a MIMO communication system.

[0074] The to-be-analyzed channel refers to a channel that needs to be predicted.

[0075] The noisy pilot matrix refers to a pilot matrix containing noise.

[0076] Exemplarily, in the MIMO communication system, there are multiple transmitting antennas and multiple receiving antennas, and the relationship between the transmitting antennas and the receiving antennas is a many-to-many channel relationship, and the channel between the transmitting antennas and the receiving antennas is the to-be-analyzed channel; each transmitting antenna sends a pilot signal, and the pilot signal passes through the to-be-analyzed channel and is received by the receiving antenna; the server combines the pilot signals received by the receiving antenna to obtain the noisy pilot matrix corresponding to the to-be-analyzed channel in the communication system.

[0077] In step S102, the noisy pilot matrix is ​​input into the trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and a second predicted channel matrix corresponding to the channel to be analyzed is obtained based on the noisy pilot matrix through a second channel prediction model or a channel prediction algorithm.

[0078] The trained first channel prediction model includes multiple trained denoisers and a trained generator in a generative adversarial network.

[0079] The first predicted channel matrix refers to a predicted channel matrix obtained by performing channel prediction on the noisy pilot matrix based on the trained first channel prediction model.

[0080] The second channel prediction model refers to another channel prediction model different from the first channel prediction model, such as a Transformer model, a CNN (Convolutional Neural Network) model, and the like.

[0081] The channel prediction algorithm refers to an algorithm for predicting a channel matrix, such as a minimum mean square error (MMSE) algorithm.

[0082] The second predicted channel matrix refers to a predicted channel matrix obtained by performing channel prediction on the noisy pilot matrix based on a second channel prediction model or a channel prediction algorithm.

[0083] It should be noted that the first channel prediction model can be used in conjunction with the second channel prediction model or with a channel prediction algorithm. For example, when the first channel prediction model includes multiple trained denoisers and a trained generator from a generative adversarial network, the second channel prediction model can be a Transformer model or a CNN model, and the channel prediction algorithm can be a minimum mean square error algorithm.

[0084] Exemplarily, the server inputs the noisy pilot matrix into the trained first channel prediction model, denoises the noisy pilot matrix through the trained first channel prediction model to obtain a clean pilot matrix, and determines the predicted channel matrix corresponding to the clean pilot matrix as the first predicted channel matrix corresponding to the channel to be analyzed; then, the server performs channel prediction processing on the noisy pilot matrix through the second channel prediction model or channel prediction algorithm to obtain a second predicted channel matrix corresponding to the channel to be analyzed.

[0085] For example, when channel prediction processing is performed on a noisy pilot matrix using a minimum mean square error algorithm, the server inputs the noisy pilot matrix into a trained second channel prediction model to obtain a channel autocorrelation matrix and noise variance value corresponding to the noisy pilot matrix, and extracts a pilot position correlation matrix corresponding to the noisy pilot matrix from the channel autocorrelation matrix; then, based on the channel autocorrelation matrix, the noise variance value and the pilot position correlation matrix, a target weight matrix corresponding to the noisy pilot matrix that minimizes the mean square error is obtained; then, based on the target weight matrix and the noisy pilot matrix, an initial predicted channel matrix corresponding to the channel to be analyzed is determined, and the initial predicted channel matrix is ​​interpolated and expanded to obtain a second predicted channel matrix corresponding to the channel to be analyzed.

[0086] Furthermore, when channel prediction processing is performed on the noisy pilot matrix through the Transformer model, the server inputs the noisy pilot matrix into the trained second channel prediction model, and normalizes the noisy pilot matrix through the trained second channel prediction model to obtain a normalized noisy pilot matrix; then, encoding processing is performed on the normalized noisy pilot matrix (for example, adding sine and cosine position codes to each matrix element in the normalized noisy pilot matrix) to obtain an encoded noisy pilot matrix; then, an initial predicted channel matrix corresponding to the encoded noisy pilot matrix is ​​determined, and the initial predicted channel matrix is ​​denormalized to obtain a second predicted channel matrix corresponding to the channel to be analyzed.

[0087] Furthermore, when the noisy pilot matrix is ​​subjected to channel prediction processing through the CNN model, the server performs dimension adjustment processing on the noisy pilot matrix to obtain an adjusted noisy pilot matrix to adapt to the data processing dimension of the CNN model; then, the adjusted noisy pilot matrix is ​​input into the trained second channel prediction model, and the adjusted noisy pilot matrix is ​​subjected to multiple feature extraction processing through the trained second channel prediction model to obtain the eigenvector corresponding to the adjusted noisy pilot matrix; then, through the trained second channel prediction model, based on the eigenvector corresponding to the adjusted noisy pilot matrix, multiple predicted channel matrices corresponding to the eigenvector and the predicted probability corresponding to each predicted channel matrix are obtained, and finally, the predicted channel matrix with the largest predicted probability is used as the second predicted channel matrix corresponding to the channel to be analyzed.

[0088] Step S103 , fusing the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

[0089] The first weight refers to the weight corresponding to the first predicted channel matrix.

[0090] The second weight refers to the weight corresponding to the second predicted channel matrix.

[0091] The target predicted channel matrix refers to a predicted channel matrix obtained by fusing the first predicted channel matrix and the second predicted channel matrix.

[0092] Exemplarily, the server inputs the first predicted channel matrix and the second predicted channel matrix into the trained importance prediction model respectively to obtain the first importance corresponding to the first predicted channel matrix and the second importance corresponding to the second predicted channel matrix; then, the server determines the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix based on the first importance and the second importance; then, the server sums the first predicted channel matrix and the second predicted channel matrix according to the first weight and the second weight to obtain the target predicted channel matrix corresponding to the channel to be analyzed.

[0093] In the above-mentioned channel prediction method, a noisy pilot matrix corresponding to a channel to be analyzed in a communication system is first obtained. The noisy pilot matrix is ​​then input into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed. A second predicted channel matrix corresponding to the channel to be analyzed is obtained based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm. The first predicted channel matrix and the second predicted channel matrix are then fused according to first weights corresponding to the first predicted channel matrix and second weights corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed. Thus, during the channel prediction process, channel prediction is performed on the noisy pilot matrix corresponding to the channel to be analyzed in the communication system using two different channel prediction methods. The fusion process is then performed based on the weights corresponding to the obtained predicted channel matrices, thereby more accurately obtaining the target predicted channel matrix corresponding to the channel to be analyzed. This is beneficial for improving the accuracy of determining the target predicted channel matrix, thereby improving the accuracy of channel prediction. Furthermore, by combining different channel prediction methods, the drawback of a single prediction method's difficulty in covering all physical propagation mechanisms in all scenarios, which results in low channel prediction accuracy, is avoided, further improving the accuracy of channel prediction.

[0094] In an exemplary embodiment, the trained first channel prediction model includes a plurality of trained denoisers and a trained generator in a generative adversarial network.

[0095] Then, the above-mentioned step S102, inputting the noisy pilot matrix into the trained first channel prediction model to obtain the first predicted channel matrix corresponding to the channel to be analyzed, specifically includes the following contents: inputting the noisy pilot matrix into multiple trained denoisers to obtain clean pilot images corresponding to the noisy pilot matrix; inputting the clean pilot images into the generator to obtain the first predicted channel matrix corresponding to the channel to be analyzed.

[0096] The denoiser refers to a network model used to denoise a noisy pilot matrix, such as a convolutional neural network model.

[0097] Among them, the Generative Adversarial Network (GAN) refers to a deep learning model that learns through an adversarial game mechanism.

[0098] The generator refers to a network structure that maps a clean pilot image to a predicted channel matrix.

[0099] The clean pilot image refers to the noisy pilot matrix after denoising.

[0100] Exemplarily, the server inputs the noisy pilot matrix into multiple trained denoisers, and each trained denoiser performs denoising on the noisy pilot matrix to obtain a clean pilot image corresponding to the noisy pilot matrix; then, the server inputs the clean pilot image into the generator to obtain a predicted channel matrix corresponding to the clean pilot image, which serves as the first predicted channel matrix corresponding to the channel to be analyzed.

[0101] For example, refer to Figure 2 The server inputs the actually received noisy pilot matrix into the channel prediction model constructed by multiple denoisers and a generative adversarial neural network to obtain a restored channel matrix as the predicted channel matrix.

[0102] In this embodiment, by inputting the noisy pilot matrix into multiple trained denoisers for collaborative denoising, compared with a single denoiser, it can more comprehensively handle complex noise scenarios, which is beneficial to improving the restoration accuracy of the clean pilot image; moreover, by using a generator to perform channel feature mapping on the clean pilot image, it is possible to generate a predicted channel matrix that is closer to the actual channel characteristics, which is beneficial to improving the determination accuracy of the first predicted channel matrix.

[0103] In an exemplary embodiment, a noisy pilot matrix is ​​input into multiple trained denoisers to obtain a clean pilot image corresponding to the noisy pilot matrix. The method specifically includes the following steps: splitting the noisy pilot matrix to obtain a first noisy pilot image corresponding to the real part of the noisy pilot matrix and a second noisy pilot image corresponding to the imaginary part of the noisy pilot matrix; inputting the first noisy pilot image into multiple trained denoisers to obtain multiple first clean pilot images corresponding to the first noisy pilot image; and inputting the second noisy pilot image into multiple trained denoisers to obtain multiple second clean pilot images corresponding to the second noisy pilot image; fusing the multiple first clean pilot images to obtain a first target clean pilot image corresponding to the first noisy pilot image, and fusing the multiple second clean pilot images to obtain a second target clean pilot image corresponding to the second noisy pilot image; and using both the first target clean pilot image and the second target clean pilot image as the clean pilot image corresponding to the noisy pilot matrix.

[0104] Each element in the noisy pilot image is in complex form and includes a real part and an imaginary part.

[0105] The first noisy pilot image refers to a two-dimensional image corresponding to the real element of the noisy pilot matrix.

[0106] The second noisy pilot image refers to a two-dimensional image corresponding to the imaginary element of the noisy pilot matrix.

[0107] The first clean pilot image refers to a real two-dimensional image processed by a denoiser that has been trained individually.

[0108] The second clean pilot image refers to an imaginary two-dimensional image processed by a denoiser that has been trained individually.

[0109] The first target clean pilot image refers to a final real part clean image obtained by fusing multiple first clean pilot images.

[0110] The second target clean pilot image refers to a final real part clean image obtained by fusing multiple second clean pilot images.

[0111] Exemplarily, the server splits the noisy pilot matrix to obtain a first noisy pilot image corresponding to the real part of the noisy pilot matrix and a second noisy pilot image corresponding to the imaginary part of the noisy pilot matrix; then, the server normalizes the first noisy pilot image and the second noisy pilot image respectively to obtain a normalized first noisy pilot image and a normalized second noisy pilot image; then, the server inputs the normalized first noisy pilot image into a plurality of trained denoisers to obtain a plurality of first clean pilot images corresponding to the first noisy pilot image, and inputs the normalized second noisy pilot image into a plurality of trained denoisers to obtain a plurality of second clean pilot images corresponding to the second noisy pilot image; then, the server performs a normalization process on the first noisy pilot image and the second noisy pilot image according to the noise matrix. The noise type corresponding to the noisy pilot matrix is ​​used to determine the corresponding weight of each trained denoiser (for example, if the noise type corresponding to the noisy pilot matrix is ​​impulse noise, the corresponding weight of the denoiser trained using impulse noise is increased); then, according to the corresponding weight of each trained denoiser, a plurality of first clean pilot images are fused to obtain a first target clean pilot image corresponding to the first noisy pilot image; and according to the corresponding weight of each trained denoiser, a plurality of second clean pilot images are fused to obtain a second target clean pilot image corresponding to the second noisy pilot image; finally, the server uses the first target clean pilot image and the second target clean pilot image as the clean pilot images corresponding to the noisy pilot matrix.

[0112] In this embodiment, by splitting the noisy pilot matrix into real and imaginary parts for independent processing, and using multiple trained denoisers for parallel denoising, the advantages of different denoisers in suppressing different noise types can be fully utilized, reducing the limitations of a single denoiser in complex noise scenarios, and facilitating improved restoration accuracy of clean pilot images.

[0113] In an exemplary embodiment, a clean pilot image is input into a generator to obtain a first predicted channel matrix corresponding to a channel to be analyzed, which specifically includes the following steps: inputting a first target clean pilot image and a second target clean pilot image into the generator respectively to obtain a third predicted channel matrix corresponding to the first target clean pilot image and a fourth predicted channel matrix corresponding to the second target clean pilot image; and combining the third predicted channel matrix and the fourth predicted channel matrix to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

[0114] The third predicted channel matrix refers to a predicted channel matrix obtained based on the first target clean pilot image.

[0115] The fourth predicted channel matrix refers to a predicted channel matrix obtained based on the second target clean pilot image.

[0116] Exemplarily, the server performs feature extraction processing on the first target clean pilot image and the second target clean pilot image, respectively, to obtain a feature vector corresponding to the first target clean pilot image and a feature vector corresponding to the second target clean pilot image; then, the server inputs the feature vector corresponding to the first target clean pilot image and the feature vector corresponding to the second target clean pilot image, respectively, into the generator, to obtain a third predicted channel matrix corresponding to the first target clean pilot image and a fourth predicted channel matrix corresponding to the second target clean pilot image; then, the server performs regularization processing on the third predicted channel matrix and the fourth predicted channel matrix, respectively, to obtain a regularized third predicted channel matrix and a regularized fourth predicted channel matrix; then, the server combines the regularized third predicted channel matrix and the regularized fourth predicted channel matrix to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

[0117] In this embodiment, the denoised real and imaginary images are respectively input into the generator to obtain the corresponding predicted channel matrix, and the predicted channel matrix is ​​combined to reconstruct the complete complex channel matrix, which ensures the phase consistency and physical rationality of the channel estimation and is conducive to improving the determination accuracy of the first predicted channel matrix.

[0118] In an exemplary embodiment, each trained denoiser corresponds to a noise type.

[0119] Then, each trained denoiser is trained in the following manner: obtaining a sample pilot matrix, splitting the sample pilot matrix to obtain a sample pilot image, and determining the noise type corresponding to the denoiser to be trained; the sample pilot image includes a clean pilot image corresponding to the real part of the sample pilot matrix, and a clean pilot image corresponding to the imaginary part of the sample pilot matrix; determining a first sample noise and a second sample noise matching the noise type, performing noise processing on the sample pilot image according to the first sample noise, to obtain a first sample noisy pilot image corresponding to the sample pilot image, and performing noise processing on the sample pilot image according to the second sample noise, to obtain a second sample noisy pilot image corresponding to the sample pilot image; the first sample noise and the second sample noise correspond to different signal-to-noise ratios; performing iterative training on the denoiser to be trained according to the first sample noisy pilot image and the second sample noisy pilot image to obtain a trained denoiser.

[0120] Among them, each trained denoiser corresponds to a noise type. Figure 3 As shown in FIG, during the training process of the denoisers that have been trained differently, different types of noise are added to the denoisers.

[0121] The noise type refers to different types of noise, such as Gaussian white noise, impulse noise, Rayleigh fading noise, etc.Figure 3 as shown.

[0122] wherein the sample pilot matrix refers to a pilot matrix used for training the denoiser.

[0123] wherein the sample pilot image refers to a pilot image obtained by splitting the sample pilot matrix.

[0124] wherein the first sample noise and the second sample noise are noises generated for the same noise type, corresponding to different signal-to-noise ratios. For example, the SNR (Signal-to-Noise Ratio) of the first sample noise is 5dB, and the SNR of the second sample noise is 10dB.

[0125] wherein the first sample noisy pilot image is used to represent the result of adding the first sample noise to the sample pilot image.

[0126] wherein the second sample noisy pilot image is used to represent the result of adding the second sample noise to the sample pilot image.

[0127] Exemplarily, the server generates channel responses of different scenarios (such as city, countryside, and indoor) by using a statistical channel model, and generates a sample pilot matrix in combination with a preset pilot sequence; then, the server splits the sample pilot matrix to obtain a pilot image corresponding to the real part of the sample pilot matrix and a pilot image corresponding to the imaginary part of the sample pilot matrix, and takes both the pilot image corresponding to the real part and the pilot image corresponding to the imaginary part as sample pilot images; then, the server determines a noise type corresponding to the denoiser to be trained; then, the server generates a first sample noise and a second sample noise matched with the noise type according to different signal-to-noise ratios; then, the server performs scaling processing on the sample pilot images to obtain scaled sample pilot images; then, the server performs noise adding processing on the scaled sample pilot images according to the first sample noise to obtain a first sample noisy pilot image corresponding to the sample pilot image, and performs noise adding processing on the scaled sample pilot images according to the second sample noise to obtain a second sample noisy pilot image corresponding to the sample pilot image; then, the server performs iterative training on the denoiser to be trained according to the first sample noisy pilot image and the second sample noisy pilot image, to obtain a trained denoiser.

[0128] For example, refer to Figure 3, the sample pilot image is denoised using noises of different SNRs (e.g., Gaussian white noise 1, Gaussian white noise 2, impulse noise 1, impulse noise 2, Rayleigh fading noise 1, Rayleigh fading noise 2) to obtain a first sample noisy pilot image and a second sample noisy pilot image; then, corresponding denoising devices are used (e.g., Gaussian white noise corresponds to denoiser 1, impulse noise corresponds to denoiser 2, and Rayleigh fading noise corresponds to denoiser n) to obtain corresponding denoised pilot images; then, based on the difference between the denoised pilot image and the second sample noisy pilot image, the loss value corresponding to the denoiser is obtained (e.g., the loss value of denoiser 1, the loss value of denoiser 2, and the loss value of denoiser n).

[0129] Figure 3 The method also includes a training process of a generative adversarial network, which is specifically described as follows: the denoised pilot image is input into the generator to obtain a predicted channel, and then the predicted channel and the corresponding actual channel are input into the discriminator to obtain the discrimination results corresponding to the predicted channel and the actual channel respectively, and then a first loss value is obtained according to the discrimination results corresponding to the predicted channel and the actual channel respectively, and a second loss value is obtained according to the difference between the predicted channel and the corresponding actual channel, and the parameters of the generative adversarial network are adjusted through the joint judgment of the first loss value and the second loss value.

[0130] In this embodiment, by training different denoisers for different noise types and introducing multiple groups of noisy samples with different signal-to-noise ratios, the denoiser can learn the statistical characteristics of noise and adapt to interference of different intensities, thereby achieving specialisation and adaptive optimization of the denoiser.

[0131] In an exemplary embodiment, a denoiser to be trained is iteratively trained based on a first sample noisy pilot image and a second sample noisy pilot image to obtain a trained denoiser, specifically comprising the following steps: inputting the first sample noisy pilot image into the denoiser to be trained to obtain a denoised pilot image corresponding to the first sample noisy pilot image; obtaining a structural loss value and a feature loss value corresponding to the denoiser to be trained based on a difference between the denoised pilot image and the second sample noisy pilot image; fusing the structural loss value and the feature loss value to obtain a target loss value; and iteratively training the denoiser to be trained based on the target loss value to obtain a trained denoiser.

[0132] The denoised pilot image refers to the first sample noisy pilot image after denoising.

[0133] The structural loss value refers to the SSIM (Structural Similarity Image Measurement) loss value.

[0134] Among them, the feature loss value refers to the feature-level MSE (Mean Square Error) loss value.

[0135] Among them, the target loss value refers to the loss value obtained by fusing the structural loss value and the feature loss value.

[0136] Exemplarily, the server inputs the first sample noisy pilot image into the denoiser to be trained, and performs denoising on the first sample noisy pilot image through the denoiser to be trained to obtain a denoised pilot image corresponding to the first sample noisy pilot image; then, the server obtains the structural loss value and the feature loss value corresponding to the denoiser to be trained based on the difference between the denoised pilot image and the second sample noisy pilot image; then, the server sums the structural loss value and the feature loss value according to the weight corresponding to the structural loss value and the weight corresponding to the feature loss value to obtain a target loss value; then, the server adjusts the model parameters of the denoiser to be trained based on the target loss value; then, the server retrains the denoiser after the model parameters are adjusted until the target loss value obtained by the trained denoiser is less than the loss value threshold, then stops training, and uses the trained denoiser as the trained denoiser.

[0137] In this embodiment, the denoiser to be trained is trained using the first sample noisy pilot image, and the structural loss value is used to measure the pixel-level differences to ensure that the denoising result is close to the real signal in basic form. At the same time, the feature loss value is used to capture the changes in high-level semantic features to avoid the loss of key signal information. By fusing the two loss values ​​to form a target loss value, accurate optimization of the denoiser can be achieved.

[0138] In an exemplary embodiment, based on the difference between the denoised pilot image and the second sample noisy pilot image, a structural loss value and a feature loss value corresponding to the denoiser to be trained are obtained, specifically including the following contents: determining the brightness similarity, contrast similarity and structural similarity between the denoised pilot image and the second sample noisy pilot image; fusing the brightness similarity, contrast similarity and structural similarity to obtain the target similarity between the denoised pilot image and the second sample noisy pilot image; and determining the structural loss value corresponding to the denoiser to be trained based on the target similarity.

[0139] The brightness similarity is used to indicate the similarity between the average brightness of the denoised pilot image and the second sample noisy pilot image.

[0140] The contrast similarity is used to indicate the degree of similarity in contrast between the denoised pilot image and the second sample noisy pilot image.

[0141] The structural similarity is used to represent the correlation between the image pixels of the denoised pilot image and the second sample noisy pilot image.

[0142] The target similarity refers to the similarity obtained by fusing brightness similarity, contrast similarity and structure similarity.

[0143] Exemplarily, the server determines a first image mean and a first image standard deviation of the denoised pilot image, a second image mean and a second image standard deviation of the second sample noisy pilot image, and a standard deviation between the denoised pilot image and the second sample noisy pilot image; then, the server determines a brightness similarity between the denoised pilot image and the second sample noisy pilot image based on the first image mean and the second image mean; then, the server determines a contrast similarity between the denoised pilot image and the second sample noisy pilot image based on the first image standard deviation and the second image standard deviation; then, the server determines a contrast similarity between the denoised pilot image and the second sample noisy pilot image based on the first image standard deviation and the second image standard deviation; The server determines the structural similarity between the denoised pilot image and the second sample noisy pilot image based on the standard deviation of the second image, the standard deviation of the second image, and the standard deviation between the denoised pilot image and the second sample noisy pilot image. Then, the server sums the brightness similarity, contrast similarity, and structural similarity according to their respective preset weights to obtain the target similarity between the denoised pilot image and the second sample noisy pilot image. Then, the server determines the structural loss value corresponding to the target similarity based on the target similarity, as the structural loss value corresponding to the denoiser to be trained.

[0144] For example, the brightness similarity, contrast similarity, structural similarity, and target similarity between the denoised pilot image and the second sample noisy pilot image, as well as the structural loss value corresponding to the denoiser to be trained, can be calculated using the following formula:

[0145]

[0146] Among them, l(x, y) is the brightness similarity, μ x is the first image mean, μ y is the second image mean, and C1 is a constant.

[0147]

[0148] Among them, c(x, y) is the contrast similarity, σ x is the first image standard deviation, σ y is the second image standard deviation, and C2 is a constant.

[0149]

[0150] Where s(x, y) is the structural similarity, and C′2=C2 / 2.

[0151] SSIM(x, y)=l(x, y)·c(x, y)·s(x, y), equation (4)

[0152] Among them, SSIM(x, y) is the target similarity.

[0153]

[0154] in, Refers to the structural loss value.

[0155] In this embodiment, by accurately analyzing the similarity between the denoised pilot image and the second sample noisy pilot image in three key dimensions: brightness, contrast, and structure, a fine characterization of the multi-dimensional feature matching of the image is achieved, and the target similarity is obtained through fusion processing to construct a structural loss value, thereby being able to more comprehensively and accurately measure the difference between the denoising result and the noisy image.

[0156] In an exemplary embodiment, a structural loss value and a feature loss value corresponding to the denoiser to be trained are obtained based on the difference between the denoised pilot image and the second sample noisy pilot image, specifically including the following contents: feature extraction processing is performed on the denoised pilot image and the second sample noisy pilot image respectively to obtain a first eigenvector corresponding to the denoised pilot image and a second eigenvector corresponding to the second sample noisy pilot image; and a feature loss value corresponding to the denoiser to be trained is obtained based on the difference between the first eigenvector and the second eigenvector.

[0157] The first eigenvector refers to the eigenvector corresponding to the denoised pilot image.

[0158] The second eigenvector refers to the eigenvector corresponding to the second sample noisy pilot image.

[0159] Exemplarily, the server inputs the denoised pilot image and the second sample noisy pilot image into the feature extraction model respectively, performs feature extraction processing on the denoised pilot image and the second sample noisy pilot image through the feature extraction model, and obtains a first eigenvector corresponding to the denoised pilot image and a second eigenvector corresponding to the second sample noisy pilot image; then, the server obtains the feature loss value corresponding to the denoiser to be trained based on the difference between each element in the first eigenvector and each element in the second eigenvector.

[0160] For example, the feature loss value can be calculated by the following formula:

[0161]

[0162] Among them, Loss refers to the feature loss value, n refers to the total number of elements involved in the calculation (such as the number of pixels), f 1i Refers to each element in the first eigenvector, f 2i refers to each element in the second eigenvector.

[0163] In this embodiment, features are extracted from the denoised pilot image and the second sample noisy pilot image respectively with the help of a feature extraction model to obtain corresponding feature vectors, and then the feature loss value is determined by calculating the difference between the two vector elements, thereby capturing the differences in abstract features such as texture and structure of the image, thereby improving the training effect of the denoiser.

[0164] In an exemplary embodiment, the above-mentioned step S103, before fusing the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain the target predicted channel matrix corresponding to the channel to be analyzed, specifically includes the following contents: obtaining a reference index corresponding to the noisy pilot matrix; determining the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix based on the comparison result between the reference index and the preset threshold.

[0165] The reference indicator refers to an indicator used to determine the first weight and the second weight, and may be a signal-to-noise ratio, coherence time, local correlation coefficient, etc. It should be noted that when channel prediction is performed on a noisy pilot matrix by combining the trained first channel prediction model with the minimum mean square error algorithm, the reference indicator refers to the signal-to-noise ratio; when channel prediction is performed on a noisy pilot matrix by combining the trained first channel prediction model with the Transformer model, the reference indicator refers to the coherence time; and when channel prediction is performed on a noisy pilot matrix by combining the trained first channel prediction model with the CNN model, the reference indicator refers to the local correlation coefficient.

[0166] The signal-to-noise ratio (SNR) represents the ratio of the pilot signal power to the noise power of the noisy pilot matrix.

[0167] The coherence time is used to represent the time interval during which the noisy channel matrix remains approximately unchanged.

[0168] Among them, the local correlation coefficient is used to represent the local similarity of the noisy channel matrix in the spatial dimension.

[0169] The term "preset threshold" refers to a pre-set threshold used for comparison with a reference indicator. It should be noted that the preset threshold varies depending on the circumstances. If the reference indicator is the signal-to-noise ratio, the preset threshold refers to the preset signal-to-noise ratio; if the reference indicator is the coherence time, the preset threshold refers to the preset coherence time; and if the reference indicator is the local correlation coefficient, the preset threshold refers to the preset local correlation coefficient.

[0170] Exemplarily, when channel prediction is performed on a noisy pilot matrix by combining a trained first channel prediction model with a minimum mean square error algorithm, the server obtains the pilot signal power and noise power of the noisy pilot matrix, and uses the ratio of the pilot signal power to the noise power of the noisy pilot matrix as the signal-to-noise ratio corresponding to the noisy pilot matrix; then, the server determines the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix based on a comparison result between the signal-to-noise ratio and a preset signal-to-noise ratio; for example, when the comparison result indicates that the signal-to-noise ratio is less than the preset signal-to-noise ratio, the first weight corresponding to the first predicted channel matrix will decrease (for example, 0.2), and the second weight corresponding to the second predicted channel matrix will increase (for example, 0.8); when the comparison result indicates that the signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio, the first weight corresponding to the first predicted channel matrix will increase (for example, 0.7), and the second weight corresponding to the second predicted channel matrix will decrease (for example, 0.3).

[0171] Furthermore, when the noisy pilot matrix is ​​predicted by combining the trained first channel prediction model and the Transformer model, the server obtains the coherence time of the noisy pilot matrix; then, the server determines the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix based on the comparison result between the coherence time and the preset coherence time; for example, when the comparison result indicates that the coherence time is less than the preset coherence time, the first weight corresponding to the first predicted channel matrix will decrease (for example, 0.1), and the second weight corresponding to the second predicted channel matrix will increase (for example, 0.9); when the comparison result indicates that the coherence time is greater than or equal to the preset coherence time, the first weight corresponding to the first predicted channel matrix will increase (for example, 0.8), and the second weight corresponding to the second predicted channel matrix will decrease (for example, 0.2).

[0172] Furthermore, when channel prediction is performed on the noisy pilot matrix by combining the trained first channel prediction model and the CNN model, the server obtains the local correlation coefficient of the noisy pilot matrix; then, the server determines the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix based on the comparison result between the local correlation coefficient and the preset local correlation coefficient; for example, when the comparison result indicates that the local correlation coefficient is greater than or equal to the preset local correlation coefficient, the first weight corresponding to the first predicted channel matrix will decrease (for example, 0.2), and the second weight corresponding to the second predicted channel matrix will increase (for example, 0.8); when the comparison result indicates that the local correlation coefficient is less than the preset local correlation coefficient, the first weight corresponding to the first predicted channel matrix will increase (for example, 0.7), and the second weight corresponding to the second predicted channel matrix will decrease (for example, 0.3).

[0173] In this embodiment, by obtaining a reference index of the noisy pilot matrix and comparing it with a preset threshold, the weights of the first predicted channel matrix and the second predicted channel matrix are dynamically determined, thereby achieving adaptive optimization of channel estimation, breaking through the limitations of traditional fixed weight fusion, and enabling channel estimation to achieve both accuracy and robustness in different noise scenarios, and being suitable for different wireless communication environments.

[0174] In an exemplary embodiment, the above-mentioned step S103, after the first predicted channel matrix and the second predicted channel matrix are fused according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain the target predicted channel matrix corresponding to the channel to be analyzed, specifically includes the following contents: according to the target predicted channel matrix, the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna are determined; the channel gain is used to represent the attenuation of the transmitted signal of each transmitting antenna during the transmission process; the phase difference is used to represent the phase change of the transmitted signal of each transmitting antenna during the propagation process; in the three-dimensional map corresponding to the channel to be analyzed, the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna are displayed.

[0175] Among them, the transmitting antenna refers to a device in a communication system used to convert electrical signals into electromagnetic waves and radiate them into space.

[0176] Among them, the receiving antenna refers to a device that captures electromagnetic waves in space and converts them into electrical signals.

[0177] Among them, the transmitted signal refers to the electromagnetic wave generated by the transmitting end and radiated through the transmitting antenna after modulation. It is the carrier of information transmission.

[0178] The three-dimensional map refers to a digital map that performs three-dimensional modeling of the physical space corresponding to the channel to be analyzed.

[0179] Exemplarily, the server determines the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna based on the value of each element in the target predicted channel matrix; then, the server determines the display method of the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna based on the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna; then, the server displays the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna in the three-dimensional map corresponding to the channel to be analyzed according to the display method of the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna.

[0180] It should be noted that the channel matrix describes the relationship between each transmitting antenna and receiving antenna during signal transmission in a multiple-input, multiple-output (MIMO) communication system. It reflects the signal propagation conditions along different channels, including factors such as signal attenuation, phase variation, and interference. Each element of the channel matrix represents the channel gain and phase difference from one transmitting antenna to one receiving antenna. The starting position of the transmitted signal is plotted on the X, Y, and Z axes, representing space. After passing through the channel and noise, the signal undergoes phase and intensity variations with distance. This can be visualized by mapping the received signal along the X, Y, and Z axes. Signal strength is represented by light and dark colors along the X, Y, and Z axes, along with directional distance. The channel gain indicates signal attenuation during transmission, while the phase difference describes the phase variation during signal propagation.

[0181] For example, refer to Figure 4 The signal sent by each transmitting antenna is gn, and the signal received by the corresponding receiving antenna is Gn. Through digital twin modeling, the channel gain and phase difference from multiple transmitting antennas to the corresponding receiving antenna are presented in 3D (Three-Dimensional) form in the digital twin system.

[0182] In this embodiment, by converting the channel parameters into an intuitive spatial distribution form, it is possible to accurately understand the attenuation law and phase evolution of the signal during propagation in three-dimensional space, which is conducive to quickly locating the weak links in the channel, and provides an intuitive and accurate basis for antenna layout optimization and signal transmission strategy adjustment, thereby improving the efficiency of channel analysis and the targeted tuning of communication system performance.

[0183] In an exemplary embodiment, Figure 5 As shown, another channel prediction method is provided, which is described by taking the application of this method to a server as an example, and includes the following steps:

[0184] Step S501: Obtain a noisy pilot matrix corresponding to a channel to be analyzed in a communication system.

[0185] Step S502: Split the noisy pilot matrix to obtain a first noisy pilot image corresponding to the real part of the noisy pilot matrix and a second noisy pilot image corresponding to the imaginary part of the noisy pilot matrix.

[0186] Step S503: Input the first noisy pilot image to multiple trained denoisers to obtain multiple first clean pilot images corresponding to the first noisy pilot image, and input the second noisy pilot image to multiple trained denoisers to obtain multiple second clean pilot images corresponding to the second noisy pilot image.

[0187] Step S504 : Fusing multiple first clean pilot images to obtain first target clean pilot images corresponding to the first noisy pilot images, and fusing multiple second clean pilot images to obtain second target clean pilot images corresponding to the second noisy pilot images.

[0188] Step S505: The first target clean pilot image and the second target clean pilot image are both used as clean pilot images corresponding to the noisy pilot matrix.

[0189] In step S506, the first target clean pilot image and the second target clean pilot image are respectively input into the generator of the trained generative adversarial network to obtain a third predicted channel matrix corresponding to the first target clean pilot image and a fourth predicted channel matrix corresponding to the second target clean pilot image.

[0190] Step S507: Combine the third predicted channel matrix and the fourth predicted channel matrix to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

[0191] Step S508: Obtain a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm.

[0192] Step S509: Obtain a signal-to-noise ratio corresponding to the noisy pilot matrix; and determine a first weight corresponding to the first predicted channel matrix and a second weight corresponding to the second predicted channel matrix based on a comparison result between the signal-to-noise ratio and a preset threshold.

[0193] Step S510 , fusing the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

[0194] In the above-mentioned channel prediction method, during the channel prediction process, two different channel prediction methods are used to perform channel prediction on the noisy pilot matrix corresponding to the channel to be analyzed in the communication system, and fusion processing is performed based on the weights corresponding to the obtained predicted channel matrix, so that the target predicted channel matrix corresponding to the channel to be analyzed can be obtained more accurately, which is conducive to improving the determination accuracy of the target predicted channel matrix, thereby improving the accuracy of channel prediction; moreover, by combining different methods to perform channel prediction, the defect that a single prediction method is difficult to cover the physical propagation mechanism of all scenarios, resulting in low channel prediction accuracy, is avoided, and the accuracy of channel prediction is further improved.

[0195] In an exemplary embodiment, in order to more clearly illustrate the channel prediction method provided in the embodiment of the present application, the channel prediction method is specifically described below with a specific embodiment. In one embodiment, the present application also provides a wireless channel estimation method for digital twins. Specifically including the following contents:

[0196] Assume there are N users and set the pilot length to L (L is less than the number of transmit antennas A). The pilot matrix is ​​processed by channel + noise to obtain a noisy pilot matrix.

[0197] The real and imaginary parts of the noisy pilot matrix are separated to form two images of the same size. From then on, they can be used as training materials in the next step of the N2N (Node to Node) denoising method and the generative adversarial network just like images.

[0198] The process of training the denoising process and the generative adversarial neural network, namely N2N denoising + generative adversarial neural network, includes the following two stages:

[0199] 1. N2N denoising training process: The pilot signal is sent multiple times within the coherence time (at this time the channel remains unchanged), and two Gaussian white noises of the same size as the clean pilot image are independently sampled (the noise power can come from different SNR situations) and added to the pilot image to form two noisy pilot images Z and Z'. The received pilots can be regarded as multiple independent noisy versions from the same clean pilot. After having two noisy images, Z is input into the denoiser to obtain Cε(Y), where Cε(Y) represents the denoiser with parameter ε. Using the 2-norm as the loss function to train the denoiser, it can be expressed as:

[0200] Loss1=E[(Z'-Cε(Z))^2], formula (7)

[0201] After the training is completed, the noisy pilot image in the test data set is input into the denoiser, and the denoiser will output the corresponding denoising result.

[0202] 2. Generate adversarial neural network training process:

[0203] The generator is responsible for estimating the channel image based on the clean pilot image, and the discriminator is responsible for distinguishing whether the input is a real channel image or a channel image generated by the generator. In each round of training, a clean pilot image X that has passed through the channel is randomly selected from the training data set and input into the generator, and the generator outputs J θ (X), where J θ (X) represents the generator with parameters θ. Then the channels K and J that generate X are θ (X) The discriminator Dφ(Y) with input parameter φ, the output of the discriminator indicates whether the input is a real channel image.

[0204] 1-norm loss is used. Therefore, the loss function of the generated adversarial network contains Loss2 and Loss3:

[0205]

[0206] Combining Loss2 and Loss3 to obtain the objective function of stage 2 training is:

[0207] min θ max φ Loss2+γ·Loss3, Formula (10)

[0208] A large number of different pilot matrices are used for training.

[0209] After the training of stage 1 and stage 2 is completed, the denoised pilot signal is input into the generator of stage 2, and the generator outputs the estimation result of the corresponding channel.

[0210] The received noisy signal (i.e., the noisy channel matrix) is input into the N2N denoising + generative adversarial neural network to obtain the predicted channel matrix for digital twin modeling.

[0211] It should be noted that generative adversarial networks have significant advantages in generating high-quality samples and adapting to a variety of tasks. However, they have disadvantages such as instability in the training process (adversarial training between the generator and the discriminator may cause the model to not converge, and the training process is often unstable and prone to oscillation), mode collapse (the generator may only learn to generate a limited number of samples, resulting in a lack of diversity in the generated data and an inability to cover the entire data distribution), and sensitivity to hyperparameters (generative adversarial networks are very sensitive to the choice of hyperparameters such as learning rate, batch size, and network architecture; improper selection may lead to training failure). Therefore, in addition to using N2N+ generative adversarial networks for channel estimation, the minimum mean square error (MMSE) method should also be used for channel estimation, and then the results of the two methods should be compromised. If the signal-to-noise ratio is low, below a certain threshold, and there is accurate statistical information (such as the second-order statistical characteristics of the channel, including the autocorrelation function and cross-correlation function of the channel. These statistical information helps the estimator to compensate and recover the signal at the receiving end, thereby minimizing the estimation error. In addition, MMSE channel estimation also needs to consider the statistical characteristics of noise, such as the power spectral density of noise, etc.), the proportion of the results produced by the minimum mean square error (MMSE) method for channel estimation should be increased, otherwise, the proportion should be reduced.

[0212] In the above embodiment, during the channel prediction process, two different channel prediction methods are used to perform channel prediction on the noisy pilot matrix corresponding to the channel to be analyzed in the communication system, and fusion processing is performed based on the weights corresponding to the obtained predicted channel matrix, so that the target predicted channel matrix corresponding to the channel to be analyzed can be obtained more accurately, which is beneficial to improving the determination accuracy of the target predicted channel matrix, and thus improving the accuracy of channel prediction; moreover, by combining different methods to perform channel prediction, the defect that a single prediction method is difficult to cover the physical propagation mechanism of all scenarios, resulting in low channel prediction accuracy, is avoided, and the accuracy of channel prediction is further improved. At the same time, combining the characteristics of wireless channels and image processing, the noisy signal is converted into an image for use in the N2N denoising method. Channel characteristics are used to normalize the sampling of N2N denoising training samples, making them more suitable for the N2N denoising method. In the training of the generative adversarial neural network, the 1-norm of the channel and the generator output is incorporated into the objective function, so that the generated image and the real image correspond at the pixel level. Traditional channel estimation methods require the use of pilot signals with a length greater than the number of antennas. For large-scale MIMO systems with hundreds of antennas, the pilot sequences used consume wireless resources. This solution enables the use of pilot signals with a relatively small length (less than the number of antennas) for model training to obtain an accurate model, saving wireless resources while obtaining an accurate channel prediction model. This solution can be applied in the field of wireless network digital twin simulation technology. Wireless network digital twin simulation technology combines AI (artificial intelligence) technology to accurately simulate wireless network status. It can be used for wireless signal coverage adjustment and optimization, service dialing and drive test simulation, and provides theoretical and practical guidance for wireless network optimization.

[0213] It should be understood that, although the steps in the flowcharts of the above embodiments are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the flowcharts of the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0214] Based on the same inventive concept, embodiments of the present application also provide a channel prediction device for implementing the aforementioned channel prediction method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations in one or more of the channel prediction device embodiments provided below can be found in the aforementioned limitations on the channel prediction method and will not be further elaborated here.

[0215] In an exemplary embodiment, Figure 6 As shown, a channel prediction device is provided, including: a pilot matrix acquisition module 601, a channel matrix prediction module 602 and a channel matrix fusion module 603, wherein:

[0216] The pilot matrix acquisition module 601 is used to acquire a noisy pilot matrix corresponding to a channel to be analyzed in a communication system.

[0217] The channel matrix prediction module 602 is used to input the noisy pilot matrix into the trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and to obtain a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix through a second channel prediction model or a channel prediction algorithm.

[0218] The channel matrix fusion module 603 is used to fuse the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain the target predicted channel matrix corresponding to the channel to be analyzed.

[0219] In an exemplary embodiment, the channel matrix prediction module 602 is further used to input the noisy pilot matrix into multiple trained denoisers to obtain a clean pilot image corresponding to the noisy pilot matrix; and input the clean pilot image into a generator to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

[0220] In an exemplary embodiment, the channel matrix prediction module 602 is further configured to split the noisy pilot matrix to obtain a first noisy pilot image corresponding to the real part of the noisy pilot matrix and a second noisy pilot image corresponding to the imaginary part of the noisy pilot matrix; input the first noisy pilot image into multiple trained denoisers to obtain multiple first clean pilot images corresponding to the first noisy pilot image; and input the second noisy pilot image into multiple trained denoisers to obtain multiple second clean pilot images corresponding to the second noisy pilot image; fuse the multiple first clean pilot images to obtain a first target clean pilot image corresponding to the first noisy pilot image; and fuse the multiple second clean pilot images to obtain a second target clean pilot image corresponding to the second noisy pilot image; and use the first target clean pilot image and the second target clean pilot image as the clean pilot images corresponding to the noisy pilot matrix.

[0221] In an exemplary embodiment, the channel matrix prediction module 602 is further used to input the first target clean pilot image and the second target clean pilot image into the generator respectively to obtain a third predicted channel matrix corresponding to the first target clean pilot image and a fourth predicted channel matrix corresponding to the second target clean pilot image; and combine the third predicted channel matrix and the fourth predicted channel matrix to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

[0222] In an exemplary embodiment, the channel prediction apparatus further includes a denoiser training module, configured to obtain a sample pilot matrix, split the sample pilot matrix to obtain a sample pilot image, and determine a noise type corresponding to a denoiser to be trained; the sample pilot image includes a clean pilot image corresponding to the real part of the sample pilot matrix and a clean pilot image corresponding to the imaginary part of the sample pilot matrix; determine a first sample noise and a second sample noise matching the noise type, perform noise processing on the sample pilot image based on the first sample noise to obtain a first sample noisy pilot image corresponding to the sample pilot image, and perform noise processing on the sample pilot image based on the second sample noise to obtain a second sample noisy pilot image corresponding to the sample pilot image; the first sample noise and the second sample noise correspond to different signal-to-noise ratios; and iteratively train the denoiser to be trained based on the first sample noisy pilot image and the second sample noisy pilot image to obtain a trained denoiser.

[0223] In an exemplary embodiment, the denoiser training module is further used to input a first sample noisy pilot image into the denoiser to be trained to obtain a denoised pilot image corresponding to the first sample noisy pilot image; obtain a structural loss value and a feature loss value corresponding to the denoiser to be trained based on the difference between the denoised pilot image and the second sample noisy pilot image; fuse the structural loss value and the feature loss value to obtain a target loss value; and iteratively train the denoiser to be trained based on the target loss value to obtain a trained denoiser.

[0224] In an exemplary embodiment, the denoiser training module is further configured to determine the brightness similarity, contrast similarity and structural similarity between the denoised pilot image and the second sample noisy pilot image; fuse the brightness similarity, contrast similarity and structural similarity to obtain a target similarity between the denoised pilot image and the second sample noisy pilot image; and determine a structural loss value corresponding to the denoiser to be trained based on the target similarity.

[0225] In an exemplary embodiment, the denoiser training module is further used to perform feature extraction processing on the denoised pilot image and the second sample noisy pilot image, respectively, to obtain a first eigenvector corresponding to the denoised pilot image and a second eigenvector corresponding to the second sample noisy pilot image; and based on the difference between the first eigenvector and the second eigenvector, to obtain a feature loss value corresponding to the denoiser to be trained.

[0226] In an exemplary embodiment, the channel prediction device also includes a weight determination module for obtaining a reference index corresponding to the noisy pilot matrix; and determining a first weight corresponding to the first predicted channel matrix and a second weight corresponding to the second predicted channel matrix based on a comparison result between the reference index and a preset threshold.

[0227] In an exemplary embodiment, the channel prediction device also includes an information display module, which is used to determine the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna based on the target predicted channel matrix; the channel gain is used to indicate the attenuation of the transmitted signal of each transmitting antenna during the transmission process; the phase difference is used to indicate the phase change of the transmitted signal of each transmitting antenna during the propagation process; in the three-dimensional map corresponding to the channel to be analyzed, the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna are displayed.

[0228] Each module in the channel prediction device described above can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0229] In an exemplary embodiment, a computer device is provided, which can be a server, and an internal structure diagram of the computer device can be as shown in FIG. 1. Figure 7 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store data such as a noisy pilot matrix and a predicted channel matrix. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with terminals outside through a network connection. The computer program is executed by the processor to implement a channel prediction method.

[0230] Those skilled in the art can understand that Figure 7 The structure shown in FIG. 1 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0231] In an exemplary embodiment, a computer device is provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0232] In an exemplary embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0233] In an exemplary embodiment, a computer program product is provided, which includes a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0234] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0235] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0236] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A channel prediction method, characterized in that: The method comprises: Obtaining a noisy pilot matrix corresponding to a channel to be analyzed in a communication system; Inputting the noisy pilot matrix into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and obtaining a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm; The first predicted channel matrix and the second predicted channel matrix are fused according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

2. The method according to claim 1, characterized in that The trained first channel prediction model includes a plurality of trained denoisers and a trained generator in a generative adversarial network; Inputting the noisy pilot matrix into the trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed includes: Inputting the noisy pilot matrix into the multiple trained denoisers to obtain a clean pilot image corresponding to the noisy pilot matrix; The clean pilot image is input into the generator to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

3. The method according to claim 2, characterized in that Inputting the noisy pilot matrix into the plurality of trained denoisers to obtain a clean pilot image corresponding to the noisy pilot matrix includes: Splitting the noisy pilot matrix to obtain a first noisy pilot image corresponding to a real part of the noisy pilot matrix and a second noisy pilot image corresponding to an imaginary part of the noisy pilot matrix; Inputting the first noisy pilot image to the multiple trained denoisers to obtain multiple first clean pilot images corresponding to the first noisy pilot image, and inputting the second noisy pilot image to the multiple trained denoisers to obtain multiple second clean pilot images corresponding to the second noisy pilot image; performing a fusion process on the multiple first clean pilot images to obtain a first target clean pilot image corresponding to the first noisy pilot image, and performing a fusion process on the multiple second clean pilot images to obtain a second target clean pilot image corresponding to the second noisy pilot image; The first target clean pilot image and the second target clean pilot image are both used as clean pilot images corresponding to the noisy pilot matrix.

4. The method according to claim 3, characterized in that Inputting the clean pilot image into the generator to obtain a first predicted channel matrix corresponding to the channel to be analyzed includes: Inputting the first target clean pilot image and the second target clean pilot image into the generator respectively to obtain a third predicted channel matrix corresponding to the first target clean pilot image and a fourth predicted channel matrix corresponding to the second target clean pilot image; The third predicted channel matrix and the fourth predicted channel matrix are combined to obtain a first predicted channel matrix corresponding to the channel to be analyzed.

5. The method according to claim 2, characterized in that Each trained denoiser corresponds to a noise type; Each trained denoiser is trained in the following manner: Obtaining a sample pilot matrix, splitting the sample pilot matrix to obtain a sample pilot image, and determining a noise type corresponding to a denoiser to be trained; The sample pilot image includes a clean pilot image corresponding to the real part of the sample pilot matrix and a clean pilot image corresponding to the imaginary part of the sample pilot matrix; determining a first noise sample and a second noise sample that match the noise type, performing noise processing on the sample pilot image based on the first noise sample to obtain a first noisy pilot image corresponding to the sample pilot image, and performing noise processing on the sample pilot image based on the second noise sample to obtain a second noisy pilot image corresponding to the sample pilot image; the first noise sample and the second noise sample correspond to different signal-to-noise ratios; The denoiser to be trained is iteratively trained according to the first sample noisy pilot image and the second sample noisy pilot image to obtain the trained denoiser.

6. The method according to claim 5, characterized in that The iterative training of the denoiser to be trained based on the first sample noisy pilot image and the second sample noisy pilot image to obtain the trained denoiser includes: Inputting the first sample noisy pilot image into the denoiser to be trained to obtain a denoised pilot image corresponding to the first sample noisy pilot image; Obtaining a structural loss value and a feature loss value corresponding to the denoiser to be trained according to a difference between the denoised pilot image and the second sample noisy pilot image; Fusing the structural loss value and the characteristic loss value to obtain a target loss value; The denoiser to be trained is iteratively trained according to the target loss value to obtain the trained denoiser.

7. The method according to claim 6, characterized in that The obtaining, according to the difference between the denoised pilot image and the second sample noisy pilot image, a structural loss value and a feature loss value corresponding to the denoiser to be trained, comprises: Determining brightness similarity, contrast similarity, and structural similarity between the denoised pilot image and the second sample noisy pilot image; fusing the brightness similarity, the contrast similarity, and the structural similarity to obtain a target similarity between the denoised pilot image and the second sample noisy pilot image; According to the target similarity, a structural loss value corresponding to the denoiser to be trained is determined.

8. The method according to claim 6, characterized in that The step of obtaining a structural loss value and a feature loss value corresponding to the denoiser to be trained according to a difference between the denoised pilot image and the second sample noisy pilot image further includes: performing feature extraction processing on the denoised pilot image and the second sample noisy pilot image respectively to obtain a first feature vector corresponding to the denoised pilot image and a second feature vector corresponding to the second sample noisy pilot image; A feature loss value corresponding to the denoiser to be trained is obtained according to a difference between the first feature vector and the second feature vector.

9. The method according to claim 1, characterized in that Before fusing the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed, the method further includes: Obtaining a reference indicator corresponding to the noisy pilot matrix; According to a comparison result between the reference indicator and a preset threshold, a first weight corresponding to the first predicted channel matrix and a second weight corresponding to the second predicted channel matrix are determined.

10. The method according to claim 1, characterized in that After fusing the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed, the method further includes: Determining, based on the target predicted channel matrix, a channel gain and a phase difference from each transmitting antenna to a corresponding receiving antenna; the channel gain being used to represent attenuation of a transmitted signal from each transmitting antenna during transmission; and the phase difference being used to represent a phase change of a transmitted signal from each transmitting antenna during propagation; In the three-dimensional map corresponding to the channel to be analyzed, the channel gain and phase difference from each transmitting antenna to the corresponding receiving antenna are displayed.

11. A channel prediction device, characterized in that: The device comprises: A pilot matrix acquisition module is used to obtain a noisy pilot matrix corresponding to a channel to be analyzed in a communication system; a channel matrix prediction module, configured to input the noisy pilot matrix into a trained first channel prediction model to obtain a first predicted channel matrix corresponding to the channel to be analyzed, and obtain a second predicted channel matrix corresponding to the channel to be analyzed based on the noisy pilot matrix using a second channel prediction model or a channel prediction algorithm; A channel matrix fusion module is used to fuse the first predicted channel matrix and the second predicted channel matrix according to the first weight corresponding to the first predicted channel matrix and the second weight corresponding to the second predicted channel matrix to obtain a target predicted channel matrix corresponding to the channel to be analyzed.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 10 are implemented.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images