Channel state information acquisition method based on conditional generative adversarial network

Abstract

The invention belongs to the technical field of wireless communication, and particularly relates to a channel state information acquisition method based on a conditional generative adversarial network. The method is based on CSI fusion and a conditional generative adversarial network (GAN), and serves a large-scale MIMO scene; according to the method, firstly, CSI based on an uplink reference signal (SRS) and CSI of a downlink feedback channel are fused, a generator generates predicted CSI of another frequency band after receiving the fused CSI, and a discriminator optimizes the performance of the generator by comparing the real CSI with the generated predicted CSI, so that the prediction accuracy is improved. Experimental results show that the acquisition precision of the channel state information in the FDD system can be effectively improved, and the fusion of the SRS and the downlink feedback and the conversion of the CSI between the frequency bands are realized. The method is suitable for wireless communication, a 5G / 6G MIMO system and other application scenes needing efficient CSI transmission.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a method for acquiring channel state information. Background Art

[0002] In frequency division duplexing (FDD) large-scale multiple-input multiple-output (MIMO) systems, accurate acquisition of downlink channel state information (CSI) is crucial for achieving efficient beamforming, interference suppression, and resource allocation. However, traditional high-precision CSI acquisition methods often rely on complex calculations and expensive hardware resources, which limits their application in actual wireless communication systems. Therefore, developing a low-computational complexity and high-precision downlink CSI acquisition method has become a research focus for optimizing FDD system performance.

[0003] In FDD systems, accurate CSI acquisition faces challenges due to uplink and downlink frequency differences and channel non-reciprocity. Channel feedback technology can assist the transmitter by feeding back downlink CSI, but it will increase the system load. Although methods such as codebook design and channel quantization have been used to reduce feedback overhead, it is still difficult to meet accuracy requirements under high-dimensional CSI, affecting system capacity.

[0004] In recent years, machine learning, especially GAN, has gradually become an effective method for uplink and downlink channel information mapping. GAN can establish a nonlinear mapping relationship between uplink and downlink channels in FDD systems through the mutual adversarial learning of the generator and the discriminator. The uplink channel information is used to generate the corresponding downlink channel state information (CSI), and the accuracy of CSI acquisition is improved without increasing the feedback bandwidth, thereby improving the beamforming and resource allocation efficiency of FDD massive MIMO systems. Summary of the invention

[0005] The purpose of the present invention is to provide a channel state information acquisition method based on a conditional generative adversarial network to improve the accuracy of FDD downlink CSI acquisition and achieve higher-precision channel estimation, thereby improving the overall performance and capacity of the FDD MIMO system.

[0006] The channel state information acquisition method based on conditional generative adversarial network provided by the present invention serves the large-scale MIMO scenario based on CSI fusion and conditional generative adversarial network (GAN); GAN can establish a nonlinear mapping relationship between uplink and downlink channels in the FDD system through mutual adversarial learning between the generator and the discriminator; the uplink channel information is used to generate the corresponding downlink channel state information (CSI), and the accuracy of CSI acquisition is improved without increasing the feedback bandwidth, thereby improving the beamforming and resource allocation efficiency of the FDD large-scale MIMO system. The specific steps are:

[0007] Step 1: Set environmental parameters such as scene, motion trajectory and antenna layout, select channel model and configure key parameters; generate multipath channel data through experiments or simulations; and generate channel covariance matrix based on channel data;

[0008] Step 2: Map the data in the channel matrix to the three channels (red, green, and blue) of the RGB image, and fuse the uplink reference signal (SRS) channel information and the downlink channel information through the fusion module to generate fused CSI data;

[0009] Step 3: The generator model in the conditional generative adversarial network receives the fused CSI data and generates the predicted CSI of another frequency band;

[0010] Step 4: Use the discriminator model in the conditional generative adversarial network to compare the generated predicted CSI with the real CSI to optimize the discriminator performance, and iteratively optimize the generator based on the feedback of the discriminator to improve the accuracy of CSI conversion between frequency bands;

[0011] Step 5: Performance evaluation: Specifically, the normalized mean square error (NMSE) and cosine similarity are used to evaluate the model-generated images to comprehensively measure its pixel-level accuracy and subspace alignment quality.

[0012] The present invention is described in further detail below:

[0013] The specific process of step 1 is as follows:

[0014] Consider establishing a MIMO wireless channel between a base station equipped with a ±45° dual-polarized uniform planar array (UPA) antenna and a user equipped with a single-polarized UPA. Assume that the channel between the user and the base station consists of L paths, each path corresponds to a direction, and the channel matrix for:

[0015]

[0016] Among them, NBS and N UE are the number of antennas at the base station side and the user side respectively; L is the number of paths; α l is the complex path gain of path l, including path loss and phase; a BS (θ BS,l ,φ BS,l ) is the direction vector of the UPA array at the base station side; α UE (θ UE,l ,φ UE,l ) is the direction vector of the user-side UPA array, where θ is the elevation angle and φ is the azimuth angle, which are used to describe the propagation direction of the signal in the vertical and horizontal planes, respectively. The specific steps for constructing the array direction vector of the transmitter or receiver are as follows.

[0017] For ±45° dual-polarization UPA, there are N V antennas, there are N in the horizontal direction H The antenna spacing in the vertical and horizontal directions is d V and d H . Its array response vector is defined as

[0018]

[0019] in represents the amplitude component, specifically, Respectively represent the field pattern (FP) in the vertical and horizontal directions of the +45° polarized antenna. Correspondingly, Respectively represent the FP in the vertical and horizontal directions of the -45° polarized antenna, where Represents the phase component and satisfies:

[0020]

[0021] Among them, 1≤i≤N V ,1≤j≤N H f∈{UL,DL}, where UL and DL represent the carrier frequencies of the uplink and downlink, respectively, v and d H The antenna spacing in the vertical and horizontal directions, N V and N H are the number of antennas in the horizontal and vertical directions, respectively. c represents the speed of light, c≈3.00×10 8 m / s.

[0022] The channel matrix H can be modeled as a complex Gaussian matrix, whose covariance matrix R H It is expressed as:

[0023] R H =E[HHH ], (5)

[0024] In practice, R can be approximated by averaging multiple channel realizations. H .

[0025] In step 2, the fusion module is used to fuse the received uplink SRS and downlink PMI feedback channel information to generate more accurate and robust channel state information data. The specific process is:

[0026] (1) The process of the base station obtaining the quantized feedback channel state information is as follows:

[0027] ① The base station sends the CSI-RS reference signal s CSI-RS , the received signal model on the user side is:

[0028] y UE =H d Vs CSI-RS +w, (6)

[0029] in, represents the receiving vector, represents the downlink channel matrix, represents the outer weight matrix, is the receiving end noise.

[0030] ② The user side estimates the downlink equivalent channel H e , the corresponding equivalent channel matrix is:

[0031]

[0032] ③ And calculate the corresponding covariance matrix R e ;

[0033]

[0034] Wherein, T and J represent the number of frequency domain subcarriers and the number of OFDM symbols respectively, and P is the number of ports.

[0035] ④ The user side quantizes the downlink CSI based on the protocol to obtain the quantization codeword And feed back to the base station.

[0036] ⑤According to the feedback vector, the PMI downlink feedback covariance information is obtained

[0037] (2) Uplink SRS and downlink PMI feedback channel state information fusion

[0038] ① Mapping: such as Figure 1 , in order to implement the conditional generative adversarial network, we first get the uplink SRS channel covariance matrix and the channel covariance matrix of the downlink PMI quantization feedback Generate a pseudo-color image. This process involves separating the real and imaginary parts of the matrix and mapping them to the RGB channels of the image, with dimensions N×N×3, where N is the number of antennas at the base station. In this representation, the real value corresponds to the red channel, the imaginary value corresponds to the green channel, and the blue channel is filled with a constant value.

[0039] ② Normalization: In order to facilitate the convergence of the subsequent training neural network, all samples are normalized to between [-1, 1]. Due to the particularity of the channel matrix, the normalization method of image processing cannot be used. The present invention chooses to scale the channel matrix by uniformly dividing it by a constant (determined according to the absolute value of the matrix).

[0040] ③ Fusion: The image generated based on the uplink channel covariance matrix of SRS and the downlink feedback covariance matrix is ​​then fused in the third dimension to obtain the fused sample like Figure 2 As shown in the figure, the blue channel does not provide any information and is removed in actual training.

[0041] In step 3, the fused sample data is received by the generator model in the conditional generative adversarial network, and a predicted channel covariance sample of another frequency band is generated.

[0042] (1) Building a conditional generative adversarial network

[0043] like Figure 2 As shown in Figure 1, the conditional generative adversarial network is a generative model that introduces conditional constraints to guide the generation of specific attributes or features of samples. The generator generates samples based on random noise and conditional variables, and the discriminator determines the authenticity of the generated samples and whether they meet the conditions. Through this adversarial training, the generated samples are not only realistic, but also consistent with the input conditions.

[0044] ①Build a generator: Figure 3 As shown in the figure, the generator adopts the U-Net network structure. U-Net is a classic generator structure that adopts a symmetrical encoder-decoder architecture and shares feature information between encoding and decoding through jump connections. The encoder is responsible for extracting gradually abstract features, while the decoder restores the image resolution by upsampling and combines the high-resolution features of the encoding stage to retain details. The final output layer generates the target image through convolution operations. In the conditional generative adversarial network, using U-Net as a generator can efficiently extract and restore features and generate more accurate downlink channel covariance images.

[0045] ②Build the discriminator: Figure 4As shown in the figure, the discriminator adopts the structure of Convolutional Neural Network (CNN), which is designed to distinguish generated images from real images. It is mainly composed of multiple convolutional layers and activation functions, which gradually extract the high-dimensional features of the image. Through downsampling operations, the discriminator can capture spatial local information and finally output a scalar value through the fully connected layer, indicating the probability that the input image is real.

[0046] (2) The conditional variable (the fused sample covariance matrix R fusion ) is input into the generator together with the random noise vector z, and the conditional variable is used to guide the generator to generate samples with specific characteristics.

[0047] The generator uses the input random noise and conditional variables to generate pseudo samples G(z|R fusion ), the purpose of generating samples is to be as close to the real data distribution as possible.

[0048] The discriminator receives the generated sample G(z|R fusion ) and the real sample R DL , combined with the conditional variable R fusion , judging whether the input is from the true downlink covariance R DL Or a generator.

[0049] The discriminator outputs a probability value D(G(z|R fusion )|R fusion ) or D(R DL |R fusion ), indicating that the network judges the input samples (z|R fusion ) and R DL Under the given condition R fusion The probability value of whether it is a true sample.

[0050] In step 4, the network parameters are optimized and iteratively optimized, using the adversarial loss function to alternately optimize the generator and discriminator through back propagation:

[0051]

[0052] The discriminator D estimates the probability of whether the data is a real downlink channel based on the input. The last term in the objective function represents the pixel-by-pixel difference between the generated image and the corresponding "real" image, and imposes a penalty term based on the size of this difference on the generator.

[0053] The goal of the generator is to minimize the objective function, while the discriminator strives to maximize the objective function, forming a minimum-maximum optimization framework as shown below:

[0054] G * = arg minG max D L CGAN (G,D), (10)

[0055] Through this iterative optimization method, the generator and discriminator gradually improve their performance in the confrontation, and the final generated sample distribution is close to the real data distribution.

[0056] The batch size is selected based on the size of the dataset and hardware conditions, and the training cycle is set based on the convergence requirements. To improve training stability, the training balance between the generator and the discriminator can be dynamically adjusted. For example, the discriminator is updated after every three rounds of training. This approach can prevent the discriminator from overfitting the training data, thereby ensuring that the generator has enough time to improve the quality of the generated images. The learning rate can also be adjusted dynamically, and ultimately the optimal model performance can be achieved through experimental tuning.

[0057] In step 5, the performance evaluation is different from the performance evaluation of traditional image processing training. The present invention uses the normalized mean square error (NMSE) and the cosine of the one-dimensional subspace angle of the matrix to measure the similarity between the downlink channel covariance matrix generated by the generator and the true downlink channel covariance matrix.

[0058] The present invention first fuses the CSI based on the uplink reference signal (SRS) and the downlink feedback channel CSI. The generator receives the fused CSI and generates the predicted CSI of another frequency band. The discriminator optimizes the generator performance by comparing the real CSI with the generated predicted CSI, thereby improving the accuracy of the prediction. Experimental results show that the present invention can effectively improve the accuracy of channel state information acquisition in FDD systems, realize the fusion of SRS and downlink feedback, and the conversion of CSI between frequency bands. The present invention is suitable for wireless communications, 5G / 6G MIMO systems, and other application scenarios that require efficient transmission of CSI.

[0059] Based on the traditional FDD downlink CSI acquisition method, the present invention combines the advantages of deep learning algorithms and proposes an innovative channel state information (CSI) fusion method based on conditional generative adversarial network (CGAN), which can efficiently integrate and model uplink (UL) and downlink (DL) information, and realize accurate modeling of complex channel distribution. By combining uplink SRS and downlink PMI feedback information, the mapping error problem caused by the uplink and downlink frequency difference in the traditional method is overcome, and the performance degradation caused by the quantization accuracy error of the feedback method in a weak channel environment is avoided. In terms of performance, the present invention significantly reduces the normalized mean square error (NMSE), improves the cosine similarity, and shows higher channel information reconstruction accuracy and adaptability, especially in complex channels and large-scale MIMO scenarios. The present invention has wide applicability and robustness, provides a new solution for improving the efficiency of UL and DL information interaction, and has important value in both theory and practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a diagram illustrating channel fusion in the method of the present invention.

[0061] Figure 2 This is a flowchart of the neural network prediction in the method of the present invention.

[0062] Figure 3 , 4 This is a structural diagram of the conditional generative adversarial network generator and discriminator in the method of the present invention.

[0063] Figure 5 , 6 It is a simulation comparison diagram of the present invention and the existing design in the method of the present invention. DETAILED DESCRIPTION

[0064] The present invention is further described below through specific implementation examples.

[0065] As an example, this experiment simulates a wireless channel scenario based on a multiple-input multiple-output communication system. The base station side uses a 2×8 dual-polarization uniform planar array antenna, a total of 32 antenna units, an antenna spacing of 0.5 wavelengths, and is fixedly installed at a height of 10 meters in the center of the scene. The user equipment is equipped with 4 omnidirectional antennas. 80% of the users are located indoors, randomly distributed in multiple rooms with a size of 10 meters × 10 meters and a height of 3 meters in a building, and 20% of the users are located in the outdoor area within a radius of 500 meters from the base station. The channel model is based on the 3GPP TR 38.901 standard. The outdoor user adopts the UMa scenario, taking into account the line of sight (Line of Sight, LoS) and non-line of sight (Non-Line of Sight, NLoS) conditions. The user equipment moves randomly in the scene at a speed of 1 meter / second, and the path length is 10 meters. Channel multipath characteristics including randomly generated angle of arrival (AoA) and angle of departure (AoD), delay spread, Doppler shift and path loss are modeled to fully evaluate the performance of MIMO communication channels in dynamic environments. Channel data is generated by the QuaDRiGa channel simulation generator.

[0066] To simulate the feedback acquisition process, a PMI feedback mechanism based on the 3GPP Release 16 protocol is introduced. In this process, the user calculates the optimal downlink beamforming matrix based on the uplink channel state information and feeds back the relevant PMI to the base station. The base station further generates the downlink channel estimation covariance based on the PMI and uplink channel information fed back by the user, which is used to enrich the input features of the generator.

[0067] The uplink channel covariance generated by QuaDRiGa and the downlink feedback channel covariance generated based on R16 quantization are extracted as complex channel matrices respectively. The generated image size is 32×32×3, which is fused by the method in step 2 as the input and target output data sets of the generator.

[0068] During the training process, a comprehensive objective function based on adversarial loss and pixel-level L1 loss is used to optimize the performance of the generator and discriminator. The adversarial loss aims to make the downlink channel image generated by the generator as close to the real data as possible, while the L1 loss further improves the accuracy of the generated results by penalizing pixel-level errors. In training, the Adam optimizer is used, with parameters set to β1=0.5, β2=0.999, the initial learning rate is set to 0.0002, and the batch size is 64. The update strategy of the generator and the discriminator is to update the discriminator every three rounds to maintain a dynamic balance between the generator and the discriminator.

[0069] The experiment verifies the model by generating multiple sets of channel data of different sizes to evaluate its generalization ability and prediction accuracy in different scenarios. Figure 5and Figure 6 The figure shows the performance of the proposed CF-CGAN framework on the test set, including the results under the NMSE and cosine similarity indicators. Under different channel conditions, CF-CGAN significantly reduces NMSE and improves cosine similarity, which fully demonstrates its ability to effectively integrate and model uplink and downlink information, and has obvious advantages over the feedback method and SRS-Hilbert method. Specifically, since the non-negativity of the angular power spectrum is not considered, SRS-Hilbert may cause mapping errors when reconstructing the downlink correlation channel matrix; while the feedback method is prone to performance degradation in weak channel environments due to quantization accuracy errors caused by a large number of antennas.

[0070] Experimental results show that the CSI fusion in the CGAN model designed in the present invention overcomes the inherent limitations of traditional methods through an effective mapping method, exhibits stronger robustness and adaptability in complex channel environments, and provides an efficient solution to the downlink channel estimation problem in frequency division duplex (FDD) systems.

Claims

1. A method for acquiring channel state information based on a conditional generative adversarial network, characterized in that: Based on CSI fusion and conditional generative adversarial network (GAN), it serves large-scale MIMO scenarios; GAN establishes a nonlinear mapping relationship between uplink and downlink channels in the FDD system through mutual adversarial learning between the generator and the discriminator. It uses the uplink channel information to generate the corresponding downlink channel state information (CSI), and improves the accuracy of CSI acquisition without increasing the feedback bandwidth, thereby improving the beamforming and resource allocation efficiency of the FDD massive MIMO system. The specific steps are: Step 1: Set environmental parameters such as scene, motion trajectory and antenna layout, select channel model and configure key parameters; generate multipath channel data through experiments or simulations; and generate channel covariance matrix based on channel data; Step 2: Map the data in the channel matrix to the red, green, and blue channels of the RGB image, and fuse the uplink SRS channel information and the downlink channel information to generate fused CSI data; Step 3: The generator model in the conditional generative adversarial network (GAN) receives the fused CSI data and generates the predicted CSI of another frequency band; Step 4: Use the discriminator model in the conditional generative adversarial network (GAN) to compare the generated predicted CSI with the real CSI to optimize the discriminator performance, and iteratively optimize the generator based on the feedback of the discriminator to improve the accuracy of CSI conversion between frequency bands; Step 5: Performance evaluation, specifically using normalized mean square error (NMSE) and cosine similarity to evaluate the model-generated images to comprehensively measure its pixel-level accuracy and subspace alignment quality.

2. The method for acquiring channel state information according to claim 1, characterized in that: The specific process of step 1 is as follows: Consider establishing a MIMO wireless channel between a base station equipped with ±45° dual-polarized uniform planar array (UPA) antennas and a user equipped with single-polarized UPA. Assume that the channel between the user and the base station consists of L paths, each path corresponds to a direction, and the channel matrix is for: Among them, N BS and N UE are the number of antennas at the base station side and the user side respectively; L is the number of paths; α l is the complex path gain of path l, including path loss and phase; a BS (θ BS,l ,φ BS,l ) is the direction vector of the UPA array at the base station side; a UE (θ UE,l ,φ UE,l ) is the direction vector of the UPA array on the user side, θ is the elevation angle, and φ is the azimuth angle, which are used to describe the propagation direction of the signal in the vertical and horizontal planes respectively; the specific steps for constructing the array direction vector of the transmitting end or the receiving end are as follows; For ±45° dual-polarization UPA, there are N V antennas, there are N in the horizontal direction H The antenna spacing in the vertical and horizontal directions is d V and d H , whose array response vector is defined as represents the amplitude component, specifically, Respectively represent the vertical and horizontal field patterns (FP) of the +45° polarized antenna. Correspondingly, They represent the FP in the vertical and horizontal directions of the -45° polarized antenna, respectively; Represents the phase component and satisfies: Among them, 1≤i≤N V ,1≤j≤N H , f∈{UL,DL}, where UL and DL represent the carrier frequencies of the uplink and downlink, respectively, v and d H The antenna spacing in the vertical and horizontal directions, N V and N H are the number of antennas in the horizontal and vertical directions, respectively, and c represents the speed of light; The channel matrix H is modeled as a complex Gaussian matrix, and its covariance matrix R H It is expressed as: R H =E[HH H ], (5) R is approximated by averaging multiple channel realizations H .

3. The method for acquiring channel state information according to claim 2, characterized in that: In step 2, the uplink SRS channel information and the downlink channel information are fused by the CSI fusion module to generate fused CSI data. The specific process is as follows: ⑴ The base station obtains the quantized feedback channel state information, and the process is as follows: ① The base station sends the CSI-RS reference signal s CSI-RS , the received signal model on the user side is: y UE =H d Vs CSI-RS +w, (6) in, represents the receiving vector, represents the downlink channel matrix, represents the outer weight matrix, is the noise at the receiving end; ② The user side estimates the downlink equivalent channel H e , the corresponding equivalent channel matrix is: ③ And calculate the corresponding covariance matrix R e ; Where T and J represent the number of frequency domain subcarriers and OFDM symbols, respectively, and P is the number of ports; ④ The user side quantizes the downlink CSI based on the protocol to obtain the quantization codeword And feedback to the base station; ⑤According to the feedback vector, the PMI downlink feedback covariance information is obtained (2) Uplink SRS and downlink PMI feedback channel state information fusion, the process is: ① Mapping: In order to realize the conditional generative adversarial network, firstly, the uplink SRS channel covariance matrix and the channel covariance matrix of the downlink PMI quantization feedback Generate a pseudo-color image; this process involves separating the real and imaginary parts of the matrix and mapping them to the RGB channels of the image, where the dimensions of the image are N×N×3, where N is the number of antennas on the base station side; where the real value corresponds to the red channel, the imaginary value corresponds to the green channel, and the blue channel is filled with a constant value; ② Normalization: In order to facilitate the convergence of the subsequent training neural network, all samples are normalized to between [-1, 1]. Specifically, the channel matrix is ​​uniformly divided by a constant to scale it. ③ Fusion: The images generated by the uplink channel covariance matrix and the downlink feedback covariance matrix based on SRS are fused in the third dimension to obtain the fused samples.

4. The method for acquiring channel state information according to claim 3, characterized in that: In step 3, the generator model in the conditional generative adversarial network (GAN) receives the fused CSI data and generates the predicted CSI of another frequency band. The specific process is as follows: ⑴Build a conditional generative adversarial network The conditional generative adversarial network is a generative model that guides the generation of specific attributes or features of samples by introducing conditional constraints. The generator generates samples based on random noise and conditional variables, and the discriminator determines the authenticity of the generated samples and whether they meet the conditions. Through adversarial training, the generated samples are not only realistic, but also consistent with the input conditions. The specific process is: ① Build the generator: The generator adopts the U-Net network structure, specifically a symmetrical encoder-decoder architecture, and shares feature information between encoding and decoding through jump connections; the encoder is responsible for extracting gradually abstract features, and the decoder restores the image resolution through upsampling, while combining the high-resolution features of the encoding stage to retain details; the final output layer generates the target image through convolution operations; in the conditional generative adversarial network, using U-Net as a generator can efficiently extract and restore features and generate more accurate downlink channel covariance images; ②Build the discriminator: The discriminator adopts the structure of convolutional neural network (CNN), which aims to distinguish generated images from real images. It consists of multiple layers of convolutional layers and activation functions to gradually extract high-dimensional features of the image. Through downsampling operations, the discriminator captures local spatial information and finally outputs a scalar value through the fully connected layer, indicating the probability that the input image is real. ⑵ The conditional variable, that is, the fused sample covariance matrix R fusion Together with the random noise vector z, the conditional variable is input into the generator, and is used to guide the generator to generate samples with specific characteristics; The generator uses the input random noise and conditional variables to generate pseudo samples G(z|R fusion ), the purpose of generating samples is to be as close to the real data distribution as possible; The discriminator receives the generated sample G(z|R fusion ) and the real sample R DL , combined with the conditional variable R fusion , determine whether the input comes from real data or the generator; The discriminator outputs a probability value D(G(z|R fusion )|R fusion ) or D(R DL |R fusion ), indicating that the network judges the input samples (z|R fusion ) and R DL Under the given condition R fusion The probability value of whether it is a true sample.

5. The method for acquiring channel state information according to claim 4, characterized in that: In step 4, the discriminator model in the conditional generative adversarial network (GAN) is used to compare the generated predicted CSI with the actual CSI to optimize the discriminator performance. The generator is iteratively optimized based on the feedback of the discriminator, thereby improving the accuracy of CSI conversion between frequency bands. The specific process is as follows: Using the adversarial loss function, the generator and discriminator are optimized alternately through back propagation, that is, the objective function is: The discriminator D estimates the probability of whether the data is a real downlink channel based on the input; the last term in the objective function represents the pixel-by-pixel difference between the generated image and the corresponding "real" image, and imposes a penalty term based on the size of the difference on the generator; The goal of the generator is to minimize the objective function, while the discriminator strives to maximize the objective function, forming the following minimum-maximum optimization framework: G * =arg min G max D L CGAN (G,D), (10) Through this iterative optimization method, the generator and discriminator gradually improve their performance in the confrontation, and the sample distribution finally generated is close to the real data distribution.

6. The method for acquiring channel state information according to claim 5, characterized in that: The performance evaluation described in step 5 uses the normalized mean square error (NMSE) and the cosine of the one-dimensional subspace angle of the matrix to measure the similarity between the downlink channel covariance matrix generated by the generator and the true downlink channel covariance matrix.

Images