A MIMO channel modeling method based on a dual-generator generative adversarial network

Abstract

The application provides a MIMO channel modeling and channel data generation method. The method first preprocesses the measured channel data, combines spatial position information to construct a conditional feature vector as an input constraint of the model; on this basis, a GAN architecture composed of a generator, a generator and a discriminator configured with a reality evaluation branch and a source identification branch is constructed, wherein the source identification branch determines the category of the sample physical source, and in the back propagation process, differential gradient constraints are applied to the two generators to guide the two generators to respectively depict different statistical subspaces in the real channel distribution; through adversarial training, the two generators respectively model the channel distribution characteristics, and the joint output is used to approximate the real channel distribution. The application uses the mixed generation distribution of the double generator to approximate the real channel distribution, which ensures the fidelity of the generated sample and significantly improves the coverage ability of the model to the range of real data distribution.

Description

Technical Field

[0001] This invention relates to the field of wireless communication and channel modeling technology, specifically to a multiple-input multiple-output (MIMO) channel modeling method based on a dual-generator generative adversarial network. Background Technology

[0002] In wireless communication systems, accurate wireless channel modeling is fundamental to system design and performance evaluation. In complex MIMO systems, the statistical characteristics of the channel directly affect system capacity, reliability, and the performance of signal processing algorithms. Traditional modeling methods, such as ray tracing, are computationally complex and dependent on environmental accuracy, while geometric statistical models struggle to accurately characterize the nonlinear properties of the channel. In recent years, data-driven channel modeling methods based on Generative Adversarial Networks (GANs) have been increasingly applied to wireless channel modeling. These methods achieve the fitting and generation of channel statistical characteristics by adversarially learning the distribution of real channel samples. However, most existing GAN-based channel modeling methods employ a single generator structure. In complex channel environments, these models are prone to generating similar samples, making it difficult for the generated models to fully cover the statistical features of the real channel distribution. Taking a communication scenario under a subway tunnel as an example, the confined spatial structure and complex electromagnetic propagation environment combine to cause the wireless channel to exhibit significant multipath effects and complex statistical characteristics. Relying on a single generator is clearly insufficient to effectively characterize both the overall statistical characteristics of the channel and the detailed features of local multipath propagation. Therefore, how to improve the expressive and coverage capabilities of the real channel distribution in generative models has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0003] This invention addresses the problem in existing generative adversarial network (GAN)-based wireless channel modeling methods where single-generator structures tend to generate similar samples, making it difficult to fully cover the statistical variations caused by multipath effects and other factors in complex wireless channels. It proposes a dual-generator GAN-based MIMO channel modeling and channel data generation method. By constructing a dual-generator cooperative adversarial framework, two generators are guided to learn the real wireless channel distribution under the same constraints, achieving joint modeling of the real wireless channel distribution. This results in MIMO channel data that more closely approximates the real wireless channel distribution while also considering the diversity of generated samples, thus improving the model's ability to fit the channel statistical characteristics and coverage in complex wireless propagation environments.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A MIMO channel modeling method based on a dual-generator generative adversarial network includes the following steps:

[0006] Acquire wireless channel data and preprocess the data to construct channel samples for model training. At the same time, extract channel-related conditional information and encode it into a conditional vector c, which, together with the random noise vector z, serves as the input to the model.

[0007] Construct a dual-generator generative adversarial network, consisting of two generators. and and a discriminator Generator and Sharing the same input format and adversarial training framework, channel samples are generated under the same conditional vector c input: , And form different generation distributions. and ;

[0008] based on and The corresponding distributions construct a hybrid generation distribution. :

[0009] ,

[0010] in, Weight coefficients are assigned to the generated samples, and a mixed generation distribution is generated. Approximating the real channel distribution ;

[0011] generator and The mapping from the low-dimensional latent space to the high-dimensional wireless channel space is generated through a deep neural network. The conditional vector c corresponding to the conditional information and the random noise vector z are fused and mapped through a fully connected structure. The deep feature learning and reconstruction of the statistical distribution of the wireless channel is achieved through multi-layer linear feature mapping, feature enhancement and progressive upsampling (including but not limited to at least one of convolutional structure, transposed convolutional structure, residual structure, self-attention structure, etc.).

[0012] Construct a discriminator The structure includes a feature extraction layer and an output layer. The feature extraction layer uses a convolutional network structure to extract multiple layers of features from the input wireless channel samples to learn deep statistical features that characterize the propagation characteristics of the wireless channel. A realism evaluation head is then set up. Source identification head ,in Used to output the Wasserstein score. Used to classify the source of input samples;

[0013] We construct a joint optimization objective function based on a Conditional Wasserstein Generative Adversarial Network (WCGAN) architecture with gradient penalty. :

[0014]

[0015] Among them, generator and By minimizing this objective function, a hybrid generation distribution can be achieved. For the actual channel distribution Approximation and complete coverage; discriminator By maximizing the objective function to minimize its own loss function, the accuracy of authenticity assessment and source identification can be improved.

[0016] Furthermore, combating losses Including the expected values ​​of real samples and generated samples:

[0017]

[0018] in, This indicates that for probability density... Generate channel data Mapping results from the discriminator Take the expected value , This represents the authenticity evaluation function. Represents the probability function Measured channel data Mapping results from the discriminator Take the expected value.

[0019] Furthermore, the gradient penalty term Represented as:

[0020]

[0021] in, This represents the interpolated sample obtained by random linear interpolation between the real sample and the generated sample. and its distribution Find the expected value of the number. Indicates the interpolated sample Perform partial derivative operations. The discriminator's authenticity assessment head In the condition vector Below, for interpolated samples The gradient vector obtained by taking the derivative, This indicates that the Euclidean norm is taken for the above gradient vector.

[0022] Furthermore, the classification loss function Based on source identification head The branched multi-class loss is used to classify the input samples into three categories, where the categories correspond to the real channel samples, ... and Generate samples:

[0023]

[0024] in, This represents the three-class classification probability output by the discriminator. The probability that the discriminator identifies the data as real channel data. The discriminator determines it as The probability of generating data, The discriminator determines it as The probability of generating data.

[0025] Furthermore, the classification loss is applied to the generator during backpropagation. and Because samples from different sources are subject to differentiated constraints in the source identification branch, and Different optimization directions can be obtained through gradient updates, thereby learning different data distribution characteristics under the same input conditions.

[0026] Furthermore, during the joint optimization process, and Different statistical sub-distributions that approximate the real channel distribution are used to generate a hybrid distribution to achieve joint modeling of the overall real channel distribution, thereby improving the ability of generated samples to cover complex channel statistical characteristics.

[0027] The model is trained using an alternating optimization training strategy, with the discriminator and generator update frequencies set to n:1. This means that after every n discriminator updates, the generator parameters are updated once. When the current training epoch satisfies the discriminator update frequency constraint, the two generators are fixed. and Parameters for the discriminator The parameters are updated using gradients; when the current training round number satisfies the generator update frequency constraint, the discriminator is fixed. The parameters are updated simultaneously for both generators. and The parameters are set; the process is iterated repeatedly until the total number of training rounds E is reached. total This is to achieve stable adversarial training between the generator and the discriminator.

[0028] Using a trained dual-generator generative adversarial network, given a conditional vector... The input is used to jointly generate wireless channel samples.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. To address the issue of traditional single-generator GANs generating only a single type of sample, a source identification branch is introduced to construct classification constraints for samples from different sources. This allows the two generators to form differentiated optimization directions under the same input conditions, thereby learning different statistical subspaces of the real channel distribution and improving the diversity of generated samples.

[0031] 2. By constructing a dual-generator collaborative adversarial modeling framework, the two generators learn the real channel distribution differently under the same adversarial target. By mixing the generated distribution, the output samples of multiple generators are fused to achieve a more comprehensive coverage of the real channel distribution and improve the realism and completeness of the generated channel samples in fitting the real channel samples.

[0032] 3. By introducing a conditional Wasserstein generative adversarial network (WCGAN) and combining it with gradient penalty constraints, the discriminator satisfies 1-Lipschitz continuity, which improves training stability. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall process of the MIMO channel modeling method based on dual generator generative adversarial networks in an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the channel data preprocessing process in an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of a dual-generator generative adversarial network model architecture in an embodiment of the present invention, including a generator. and And a discriminator with authenticity evaluation branch and source identification branch;

[0036] Figure 4 This is a schematic diagram comparing the statistical parameters of the generated channel sample and the measured channel sample in an embodiment of the present invention. The statistical parameters include, but are not limited to, the root mean square delay spread.

[0037] Figure 5 This is a schematic diagram comparing the cumulative distribution function of statistical parameters of the generated channel sample and the measured channel sample in an embodiment of the present invention. The statistical parameters include, but are not limited to, the root mean square delay spread. Detailed Implementation

[0038] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0039] This embodiment provides a detailed description of the MIMO channel modeling method based on a dual-generator generative adversarial network provided by the present invention. For example... Figure 1 As shown, this invention provides a MIMO channel modeling method based on a dual-generator generative adversarial network. The overall process includes channel data construction, condition information encoding, dual-generator adversarial modeling, and channel sample generation.

[0040] This invention provides a MIMO channel modeling method based on a dual-generator generative adversarial network, comprising the following steps:

[0041] S1: Acquire wireless channel data and preprocess the wireless channel data to construct channel sample data for model training and conditional information related to the wireless channel propagation scenario; wherein, the wireless channel data in S1 is multiple-input multiple-output (MIMO) channel data, represented by complex channel impulse response, as follows:

[0042]

[0043] in, Indicates the number of multipath components. and Indicates the first Path gain and delay of each path Indicates the first Phase changes along the path.

[0044] Preprocessing of wireless channel data includes: solving the real and imaginary parts of the complex channel data, constructing a three-dimensional real-valued channel matrix representation, and normalizing the channel data before inputting it into the generation model.

[0045] Specifically, such as Figure 2 As shown, this embodiment preprocesses the wireless channel data. First, it acquires MIMO channel data at different spatial locations. When the signal travels from the first... The root transmitting antenna transmits to the first During the process of receiving the antenna, the first The received signal can be modeled as a superposition of multipath components with different amplitudes and phases, and this channel data can be represented by the complex impulse response (CIR):

[0046]

[0047] in, This indicates the number of consecutive measurements taken at this measurement point. One cyclic sample, Indicates the number of multipath components. , Indicates the first Gain and delay of each path Indicates the first The phase change of each path. To adapt to the input requirements of deep learning networks, it is split into real and imaginary parts, and the channel matrix collected at each measurement point is... Reconstructed into a three-dimensional real-valued channel matrix The first dimension, 2, represents the real and imaginary parts of the original complex number CIR. This represents the total number of transmit and receive antenna pairs, i.e., the number of spatial sub-channels. Min-max normalization is applied to the data at each location, mapping the matrix elements to [-1, 1].

[0048]

[0049] in, and These represent the maximum and minimum values ​​of the channel matrix at the current measurement location, respectively. After data preprocessing, the spatial location information corresponding to the channel is one-hot encoded into a conditional vector c, and then combined with the random noise vector... Concatenate the data to create a generator. and Input .

[0050] In this embodiment, the condition information is used to characterize the propagation scenario state corresponding to the wireless channel, including but not limited to at least one of the following: spatial location environment, relative distance between the transceiver and receiver, scenario category information, antenna array parameters, etc. The condition information includes the spatial location index corresponding to the wireless channel measurement location, and a condition vector c is constructed using a One-Hot encoding method.

[0051] S2: Construct a dual-generator generative adversarial network, including two generators. and and a discriminator The generator and Sharing the same input format and adversarial training framework, the conditional information is encoded into a conditional vector c, which is then concatenated with a random noise vector z and used as the generator input to generate channel samples respectively. , Different generation distributions are formed. and , , express and Function to generate channel data.

[0052] Generator in this step and The mapping from a low-dimensional latent space to a high-dimensional wireless channel space is achieved through a deep neural network. This includes an input mapping layer, a feature extraction layer, and a feature recovery generation layer. The input mapping layer uses a fully connected structure to fuse and map the conditional vector c corresponding to the conditional information with the random noise vector z. The feature extraction and feature recovery generation layers include, but are not limited to, at least one of the following: convolutional structure, transposed convolutional structure, residual structure, self-attention structure, etc. Through multi-layer sublinear feature mapping, feature enhancement, and stepwise upsampling, deep feature learning and reconstruction of the statistical distribution of the wireless channel are achieved, thereby generating high-dimensional MIMO channel samples that match the statistical characteristics of the real wireless channel.

[0053] Discriminator This tool is used to determine the authenticity and learn features of input wireless channel samples. Its input includes real wireless channel samples or generated wireless channel samples and corresponding condition vectors. Its structure includes a feature extraction layer and an output layer. The feature extraction layer uses a convolutional network structure to extract multiple layers of features from the input wireless channel samples in order to learn deep statistical features that characterize the propagation characteristics of wireless channels, and completes the distribution determination of the input wireless channel samples based on the extracted features.

[0054] S3: Based on the above and The corresponding distributions construct a hybrid generation distribution. :

[0055] ,

[0056] in, Assign weight coefficients to the generated samples; the hybrid generation distribution Used to approximate the real channel distribution .

[0057] S4: The discriminator Set up an authenticity assessment header Source identification head ,in Used to output Wasserstein scores to assess the realism of the samples. Used to classify the source of input samples into three categories, corresponding to real data samples, and The generated samples are then used to construct a classification loss function based on the source identification branch, thereby constraining the differentiation of samples from different sources.

[0058] S5: During adversarial training, a classification loss function is constructed using the source identification branch, and adversarial optimization is combined to make... and Differentiated feature representations are formed under a unified training framework, thereby achieving joint modeling of the overall channel statistical characteristics through the hybrid generation distribution;

[0059] S6: Construct an adversarial optimization target based on the hybrid distribution and the real distribution, and perform alternating optimization training on the discriminator and generator, and use the trained GAN model to jointly generate wireless channel samples in the target scenario.

[0060] This embodiment constructs a generator... and With discriminator The dual-generator generative adversarial network structure is composed of the following model architecture: Figure 3 As shown. Two generators are used under the same input conditions. The following are the generation channel samples: , This results in two generating distributions. and .

[0061] The discriminator performs both authenticity assessment and source identification functions, and its output includes an authenticity assessment branch. and source identification branch .in Used to output the Wasserstein score. Used to classify the source of input samples;

[0062] We construct a joint optimization objective function based on a Conditional Wasserstein Generative Adversarial Network (WCGAN) architecture with gradient penalty. :

[0063]

[0064] Among them, generator and By minimizing this objective function, a hybrid generation distribution can be achieved. For the actual channel distribution Approximation and complete coverage; discriminator By maximizing the objective function to minimize its own loss function, the accuracy of authenticity assessment and source identification can be improved.

[0065] In this embodiment, the source identification head The Softmax function is used to calculate the samples from different labels. Corresponding to real channel data, Generate samples, The probability of generating a sample :

[0066]

[0067] in, Constructing vectors Used for the corresponding sample source and meets the following requirements , This indicates that the source identification branch of the discriminator is under the condition Below is an example of input samples Corresponding to the Linear prediction output value for the class.

[0068] Based on the above structure, the two generators form a cooperative adversarial mechanism under the source identification constraint of the discriminator. Specifically, when the distributions of the two generators tend to be consistent, the discriminator can more easily distinguish their sources, leading to an increase in classification loss and imposing constraints on generator parameter updates during backpropagation; while when the two generators learn the real channel distribution separately... When different features are distributed in the data, the classification loss decreases, thus driving it to converge to different distribution regions. Therefore, the two generators will adaptively adjust to the real channel distribution. Decomposition modeling is performed, and its mixed generation distribution is obtained. Gradually approximating the true data distribution :

[0069] ,

[0070] in, Weighting coefficients are assigned to the generated samples. Through this mechanism, the two generators respectively cover the real channel distribution, thereby ensuring that the generated samples are close to the real samples while improving the diversity of the samples.

[0071] In terms of loss function construction, a conditional WGAN-based adversarial loss function is introduced to provide adversarial loss. Including the expected values ​​of real samples and generated samples:

[0072]

[0073] in, Indicating resistance to loss, This indicates that for probability density... Generate channel data Mapping results from the discriminator Take the expected value , This represents the authenticity evaluation function. Represents the probability function Measured channel data Mapping results from the discriminator Take the expected value of the number. This is the encoded conditional vector.

[0074] Authenticity assessment head Satisfying the 1-Lipschitz continuity constraint, and achieving this through a gradient penalty term:

[0075]

[0076] in, This is a gradient penalty term. This represents the interpolated sample obtained by random linear interpolation between the real sample and the generated sample. and its distribution Find the expected value of the number. Indicates the interpolated sample Perform partial derivative operations. The discriminator's authenticity assessment head Under the condition vector c, for the interpolated samples The gradient vector obtained by taking the derivative, This indicates that the Euclidean norm is taken with respect to the above gradient vector. Where, the interpolation sample... , .in, Let the random factor follow a uniform distribution in [0,1]. To obtain real channel data distribution The samples obtained from sampling Indicates from and Mixed generation distribution The sample obtained from sampling.

[0077] Meanwhile, the source identification and classification loss function is:

[0078]

[0079] in, The probability that the discriminator identifies the data as real channel data. The discriminator determines it as The probability of generating data, The discriminator determines it as The probability of generating data.

[0080] Therefore, the total loss function of the discriminator is:

[0081]

[0082] in, and These are the weight coefficients for gradient loss and classification loss, respectively. Adversarial loss. Used to approximate the true distribution; gradient loss Used to ensure training stability; classification loss This loss is used to classify the physical origin of samples, and it is applied to the corresponding generators during backpropagation, making... and Learn different data distribution characteristics under the same input conditions.

[0083] The generator optimization objective is to optimize the generator based on a given condition vector. Under the constraints, by minimizing the total loss function of the generator. Bootstrap generator They jointly approximate the real channel distribution, the The expression is:

[0084]

[0085] Minimizing the first term is used to ensure that the generated sample is in the discriminator's authenticity evaluation head. To obtain a higher authenticity score; minimizing the second term is equivalent to maximizing the source identification head. This makes the generated samples more likely to be classified as the true category. The classification loss's constraint on the distinction between samples from different sources makes... and Differentiated distributions are formed in the feature space, thereby improving the coverage of the real channel distribution by the hybrid generated distribution.

[0086] In step S6, the discriminator and generator undergo alternating optimization training, including:

[0087] S6.1: Set the total number of training rounds E total and initialize the generator. , With discriminator Network parameters;

[0088] S6.2: When the current training round number satisfies the discriminator update frequency constraint, fix the two generators. and Parameters, using real channel samples, generated samples, and corresponding condition vectors. Based on the total loss function For discriminator The parameters are updated using gradients;

[0089] S6.3: When the current training round number satisfies the generator update frequency constraint, fix the discriminator. The parameters are updated simultaneously for both generators. and The parameters are minimized by optimizing the objective. Guided model training;

[0090] S6.4: Iterate through the loop, repeating S63 and S64 above, until the total number of training rounds E is reached. total .

[0091] In the alternating optimization training, the discriminator and generator update frequencies are set to n:1. Within each training cycle, n discriminator parameter updates (S6.2) are performed first, followed by one generator parameter update (S6.3) after every n discriminator updates. Using the trained dual-generator generative adversarial network, given a conditional vector... The input is used to jointly generate wireless channel samples.

[0092] To verify the effectiveness of the dual-generator adversarial structure, in this example... and A heterogeneous convolutional neural network structure was designed. A self-attention mechanism is introduced during feature extraction, and attention is paid to the following: and A self-attention module is constructed at the feature scale to enhance the model's ability to model the global statistical features of the wireless channel. This self-attention mechanism learns the long-range spatial correlation characteristics between different propagation path components in the wireless channel by establishing correlations between different locations in the feature map. Its calculation form is as follows:

[0093]

[0094] in, , , These represent the query matrix, key matrix, and value matrix obtained from the input feature mapping, respectively. This is used to normalize the correlation weights between different locations. By adaptively modeling the correlation of global features, the generator can more effectively learn the spatial propagation characteristics and statistical distribution properties in complex wireless propagation environments.

[0095] Residual connections are introduced into the network structure. The input features first undergo two nonlinear transformations consisting of convolution, batch normalization, and LeakyReLU activation functions, and then are added element-wise to the original input. The mathematical expression is as follows:

[0096]

[0097] This identity mapping design allows gradients to be directly transmitted to shallow layers through short-circuit connections during network training, effectively alleviating the gradient vanishing problem and enabling the network to focus more on characterizing detailed multipath perturbation features.

[0098] Through the aforementioned structural differences, the two generators adopt different emphases during feature modeling, and their differences are further enhanced under the source identification constraint of the discriminator. This facilitates the two generators in learning different feature patterns in the real channel distribution. It should be noted that the above generator structure is only one specific embodiment of the present invention, and the present invention does not limit the specific network structure of the generator.

[0099] During model training, an alternating optimization strategy is used to update the parameters of the discriminator and the two generators. In each training iteration, the generator parameters are first fixed, and the discriminator parameters are updated n times using gradients. Subsequently, the discriminator parameters are fixed, and the parameters are updated simultaneously. and Parameters are set to achieve stable adversarial training until the training ends, using two generators. and Channel samples are generated separately and then combined to obtain the final generated channel sample.

[0100] To verify the effectiveness of the generated channel, such as Figure 4 As shown, a comparative analysis of the root mean square delay spread of the generated channel samples and the measured channel samples is performed, which is defined as:

[0101]

[0102] in, Indicates origin from different tags The probability of channel data. Furthermore, a comparative analysis is performed on the cumulative distribution functions of the above statistical parameters, such as... Figure 5 As shown in the figure. The results show that the generated channel samples and the measured channel samples have good consistency in statistical distribution, thus verifying the effectiveness of the method of the present invention.

[0103] This invention effectively solves the technical problems of insufficient diversity of channel samples generated by traditional single-generator models in complex wireless environments and easy mode collapse by constructing a collaborative adversarial architecture of dual generators and a discriminator with a realism evaluation branch and a source identification branch, and by utilizing the differentiated gradient signal generated by the source identification branch.

[0104] The specific combination of self-attention modules and residual block structures described in this specification is only a preferred embodiment for implementing the heterogeneous generation architecture of the present invention, and is intended to illustrate the technical principles and beneficial effects of the present invention. It does not constitute a limitation on the specific network structure of the generator.

[0105] Those skilled in the art should understand that any equivalent substitutions or improvements made to the generator's internal network layers, activation function types, or parameter configurations without departing from the core technical concept of this invention should be included within the scope of protection of this invention. The scope of protection of this invention is defined by the claims.

Claims

1. A method for MIMO channel modeling and channel data generation, characterized in that, include: S1: Acquire wireless channel data and preprocess the wireless channel data to construct channel sample data for model training and condition information related to the wireless channel propagation scenario. S2: Construct a dual-generator generative adversarial network, including two generators. and and a discriminator The generator and Sharing the same input format and adversarial training framework, the conditional information is encoded into a conditional vector c, which is then concatenated with a random noise vector z and used as generator input to generate channel samples respectively. , Different generation distributions are formed. and , and express and Function to generate channel data; S3: Based on the above and The corresponding distributions construct a hybrid generation distribution. : ， ， in, Assign weight coefficients to the generated samples; the hybrid generation distribution Used to approximate the real channel distribution ; S4: The discriminator Set up an authenticity assessment header Source identification head ,in Used to output Wasserstein scores to assess the realism of the samples. Used to classify the source of input samples into three categories, corresponding to real data samples, and The generated samples are then used to construct a classification loss function based on the source identification branch, which distinguishes samples from different sources. S5: During adversarial training, a classification loss function is constructed using the source identification branch, and adversarial optimization is combined to make... and Differentiated feature representations are formed under a unified training framework, thereby achieving joint modeling of the overall channel statistical characteristics through the hybrid generation distribution; S6: Construct an adversarial optimization target based on the hybrid distribution and the real distribution, and perform alternating optimization training on the discriminator and generator, and use the trained GAN model to jointly generate wireless channel samples in the target scenario.

2. The MIMO channel modeling and channel data generation method according to claim 1, characterized in that, The wireless channel data in S1 is multiple-input multiple-output (MIMO) channel data, represented by the complex channel impulse response, as follows: ， in, Indicates the number of multipath components. and Indicates the first Path gain and delay of each path Indicates the first Phase changes along the path.

3. The MIMO channel modeling and channel data generation method according to claim 1, characterized in that, S2 generator and The mapping from a low-dimensional latent space to a high-dimensional wireless channel space is achieved through a deep neural network. This includes an input mapping layer, a feature extraction layer, and a feature recovery generation layer. The input mapping layer uses a fully connected structure to fuse and map the conditional vector c corresponding to the conditional information with the random noise vector z. The feature extraction and feature recovery generation layers include, but are not limited to, at least one of the following: convolutional structure, transposed convolutional structure, residual structure, self-attention structure, etc. Through multi-layer sublinear feature mapping, feature enhancement, and stepwise upsampling, deep feature learning and reconstruction of the statistical distribution of the wireless channel are achieved, thereby generating high-dimensional MIMO channel samples that match the statistical characteristics of the real wireless channel.

4. The MIMO channel modeling and channel data generation method according to claim 1, characterized in that, The discriminator described in S2 This tool is used to determine the authenticity and learn features of input wireless channel samples. Its input includes real wireless channel samples or generated wireless channel samples and corresponding condition vectors. Its structure includes a feature extraction layer and an output layer. The feature extraction layer uses a convolutional network structure to extract multiple layers of features from the input wireless channel samples in order to learn deep statistical features that characterize the propagation characteristics of wireless channels, and completes the distribution determination of the input wireless channel samples based on the extracted features.

5. The MIMO channel modeling and channel data generation method according to claim 1, characterized in that, In S4, the dual-generator generative adversarial network adopts an adversarial learning approach based on the Conditional Wasserstein Generative Adversarial Network (WCGAN) and incorporates gradient penalty. The discriminator includes a realism evaluation head. Source identification head Its countermeasure loss is expressed as: ， in, Indicating resistance to loss, This indicates that for probability density... Generate channel data Mapping results from the discriminator Take the expected value , This represents the authenticity evaluation function. Represents the probability function Measured channel data Mapping results from the discriminator Take the expected value of the number. This is the encoded conditional vector.

6. The MIMO channel modeling and channel data generation method according to claim 5, characterized in that, The authenticity assessment head Satisfying the 1-Lipschitz continuity constraint, and achieving this through a gradient penalty term: ， in, This is a gradient penalty term. This represents the interpolated sample obtained by random linear interpolation between the real sample and the generated sample. and its distribution Find the expected value of the number. Indicates the interpolated sample Perform partial derivative operations. The discriminator's authenticity assessment head Under the condition vector c, for the interpolated samples The gradient vector obtained by taking the derivative, This indicates that the Euclidean norm is taken for the above gradient vector.

7. A MIMO channel modeling and channel data generation method according to claim 6, characterized in that, The interpolation sample for: ， in, Let the random factor follow a uniform distribution in [0,1]. To obtain real channel data distribution The samples obtained from sampling Indicates from and Mixed generation distribution The sample obtained from sampling.

8. A MIMO channel modeling and channel data generation method according to claim 1, characterized in that, In S4, the source identification head The Softmax function is used to calculate the samples from different labels. probability , Corresponding to real channel data, Generate samples, The formula for generating samples is as follows: ， in, Constructing vectors Used for the corresponding sample source and meets the following requirements , This indicates that the source identification branch of the discriminator is under the condition Below is an example of input samples Corresponding to the Linear prediction output value for the class.

9. The MIMO channel modeling and channel data generation method according to claim 5, characterized in that, The source identification head Based on the probability of the output Construct a three-class cross-entropy classification loss function: ， in, This represents the cross-entropy loss for three-class classification. This represents the three-class classification probability output by the discriminator. The probability that the discriminator identifies the data as real channel data. The discriminator determines it as The probability of generating data, The discriminator determines it as The probability of generating data.

10. A method for MIMO channel modeling and channel data generation according to claim 1, characterized in that, The total loss function of the discriminator described in S6 is: ， in, and These are the weight coefficients for gradient loss and classification loss, respectively.

Images