Wireless channel modeling method and apparatus

Abstract

The application provides a wireless channel modeling method and device, the method comprising: inputting noise data to a generator of a target model, a linear block mapping a latent vector of the noise data to a higher-dimensional feature space, a convolution transpose block up-sampling the abstract high-dimensional features to make them more specific, the specific high-dimensional features being standardized in a batch layer to adjust the mean and variance of the features; the high-dimensional features being compressed by global spatial information, re-labeled by feature channels and re-weighted by a channel attention module to obtain a feature re-weighting result, and finally generating wireless channel data by an activation layer; inputting the generated wireless channel data and real channel data to a discriminator of the target model, expanding the input data in a zero padding block, extracting features by a convolution block and discriminating by a linear block to output a discrimination result, updating the weight parameters of the generator according to the discrimination result, and training the generator to generate wireless channel data similar to the real channel data. The application realizes effective channel modeling.

Description

Technical Field

[0001] The present invention relates to the field of wireless network technology, and in particular to a wireless channel modeling method and device. Background Art

[0002] Wireless channel modeling is the cornerstone of wireless network design and performance analysis. It provides a theoretical foundation for understanding channel behavior, which in turn guides wireless system design and optimization. This modeling effort must consider not only the physical characteristics of signal transmission, such as attenuation, multipath effects, and Doppler shift, but also environmental factors, such as the influence of terrain, buildings, and other obstacles. However, due to the inherent characteristics of wireless channels, including multipath propagation, fading, and shadowing, they exhibit high levels of nonlinearity and time-varying properties, making channel characteristics complex and difficult to accurately describe using traditional mathematical models.

[0003] With the development of deep learning technology, a series of deep learning-based wireless channel modeling methods have emerged in recent years. These methods leverage the powerful function approximation capabilities of neural networks to build channel models by learning from large amounts of channel measurement data, aiming to capture the complex nonlinear relationships in channel characteristics. They typically rely on large-scale field measurement data to ensure that the model accurately reflects the physical and statistical characteristics of the channel. However, this reliance introduces two major problems: first, the field measurement process is often time-consuming and costly, especially in complex or inaccessible environments; second, as the communication environment constantly changes, the amount of data required to continuously update the model is enormous, and traditional methods face exponentially growing data demands, further exacerbating the resource burden. Summary of the Invention

[0004] The present invention provides a wireless channel modeling method and device to overcome the defects of the existing technology that wireless channel modeling relies on large-scale field measurement data and is high in cost. The method combines the attention mechanism with WGAN-GP for channel modeling, which can significantly reduce the amount of channel data required in the modeling process and effectively improve the accuracy and diversity of wireless channel samples.

[0005] The present invention provides a wireless channel modeling method, comprising:

[0006] The noise data is input into the generator of the target model. The linear block maps the latent vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features. The convolutional transposition block upsamples the abstract high-dimensional features to make them more specific. The specific high-dimensional features are normalized in the batch layer to adjust the mean and variance of the features. The high-dimensional features are compressed by global spatial information, recalibrated by feature channels, and reweighted in the channel attention module to obtain the feature reweighted results for each channel. Finally, the wireless channel data output is generated through the activation layer.

[0007] The generated wireless channel data and the real channel data are input as input data to the discriminator of the target model, wherein the input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block to output an identification result;

[0008] The weight parameters of the generator are updated according to the identification result, and the generator is trained to generate wireless channel data whose similarity with the real channel data is less than a preset threshold, thereby realizing effective wireless channel modeling.

[0009] According to a wireless channel modeling method provided by the present invention, the high-dimensional features are subjected to global spatial information compression, feature channel recalibration, and feature reweighting in a channel attention module to obtain a feature reweighted result for each channel, including:

[0010] Compressing the global spatial information of the feature map of the noise data by an adaptive average pooling operation, reducing the spatial dimensions of each channel of the noise data in height and width to 1, and obtaining the global spatial features of the noise data;

[0011] Passing the global spatial features through a network including a fully connected layer to obtain a weight coefficient for each channel;

[0012] After the weight coefficient of each channel is expanded to the spatial dimension of the corresponding feature map, it is multiplied by the feature map of each channel to obtain the feature re-weighted result of each channel.

[0013] According to a wireless channel modeling method provided by the present invention, the global spatial features are passed through a network including a fully connected layer to obtain a weight coefficient for each channel, including:

[0014] Performing dimensionality reduction on the global spatial features through a first fully connected layer;

[0015] Processing the reduced-dimensional global spatial features through a ReLU activation function to increase the expressive power of the target model;

[0016] Performing dimensionality restoration on the processed global spatial features through a second fully connected layer;

[0017] The global spatial features after dimension restoration are passed through the Sigmoid activation function to obtain the weight coefficient of each channel.

[0018] According to a wireless channel modeling method provided by the present invention, the linear block maps the potential vector of the noise data to a higher-dimensional feature space to obtain a high-dimensional feature, including:

[0019] Inputting the noise data into the fully connected layer, batch normalization layer and Leaky ReLU activation function in the linear block in sequence;

[0020] The fully connected layer is used to map the potential vector in the noise data to a feature space of a specific dimension;

[0021] The batch normalization layer is used to perform batch normalization processing on the feature vector of the noise data to accelerate the training process of the target model and stabilize the target model training;

[0022] The Leaky ReLU activation function is used to introduce nonlinear transformation and enhance the expressive power of the target model.

[0023] According to a wireless channel modeling method provided by the present invention, the convolution transposition block upsamples the abstract high-dimensional features to make them more specific, and the specific high-dimensional features are normalized in the batch processing layer to adjust the mean and variance of the features, including:

[0024] The feature map output by the linear block is sequentially convolved and transposed through multiple convolutional transposition layers in the convolutional transposition block;

[0025] Each of the convolutional transpose layers except the last convolutional transpose layer in the multiple convolutional transpose layers is sequentially connected to a batch normalization layer, a channel attention module, and a Leaky ReLU activation function;

[0026] The last convolutional transpose layer upsamples the feature map to one channel, and finally outputs the generated wireless channel data through a Tanh activation function.

[0027] According to a wireless channel modeling method provided by the present invention, the generated wireless channel data and the real channel data are input as input data to the discriminator of the target model, the input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block to output an identification result, including:

[0028] Passing the real wireless channel data and the generated wireless channel data sequentially through the zero-padding module, the convolution block, and the linear block of the discriminator to obtain the identification result;

[0029] The zero padding module preprocesses the real wireless channel data and the generated wireless channel data through a Zero Padding function;

[0030] The convolution block is used to perform feature extraction on the real wireless channel data and the generated wireless channel data;

[0031] The linear block is used to identify the generated wireless channel data and output the identification result.

[0032] The present invention also provides a wireless channel modeling device, comprising:

[0033] A generation module is used to input noise data into the generator of the target model. The linear block maps the latent vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features. The convolution transposition block upsamples the abstract high-dimensional features to make them more specific. The specific high-dimensional features are normalized in the batch processing layer to adjust the mean and variance of the features. The high-dimensional features are compressed by global spatial information, recalibrated by feature channels, and reweighted in the channel attention module to obtain the feature reweighted results of each channel. Finally, the activation layer generates wireless channel data output.

[0034] an identification module, configured to input the generated wireless channel data and the real channel data as input data into the discriminator of the target model, wherein the input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block to output an identification result;

[0035] A training module is used to update the weight parameters of the generator according to the identification result, train the generator to generate wireless channel data whose similarity with the real channel data is less than a preset threshold, and realize effective wireless channel modeling.

[0036] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any one of the wireless channel modeling methods described above is implemented.

[0037] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which implements any of the wireless channel modeling methods described above when executed by a processor.

[0038] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any one of the wireless channel modeling methods described above.

[0039] The wireless channel modeling method and device provided by the present invention, by introducing a channel attention module (SEL) into the WGAN-GP generator, can adaptively recalibrate the importance of each feature channel, thereby focusing more on features that are critical to channel modeling. This not only enhances the model's ability to capture channel characteristics, but also improves the quality of generated channel samples, thereby enhancing the accuracy and efficiency of wireless channel modeling. Furthermore, the introduction of the attention mechanism helps improve the training stability and efficiency of the model, enabling the model to perform channel modeling more flexibly and accurately in complex and changing wireless environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 Schematic diagram of the structure of GAN in the wireless channel modeling method provided by the present invention;

[0042] Figure 2 It is a flow chart of the wireless channel modeling method provided by the present invention;

[0043] Figure 3 This is one of the structural diagrams of the SE-GAN model in the wireless channel modeling method provided by the present invention;

[0044] Figure 4 This is the second structural diagram of the SE-GAN model in the wireless channel modeling method provided by the present invention;

[0045] Figure 5 Schematic diagram of the structure of the channel attention module in the wireless channel modeling method provided by the present invention;

[0046] Figure 6 2. It is a schematic diagram comparing the probability distribution of data generated by different models in the wireless channel modeling method provided by the present invention;

[0047] Figure 7 It is a structural diagram of the wireless channel modeling device provided by the present invention;

[0048] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0050] Rapid advances in wireless communication technologies are driving the development of modern society, making them indispensable in daily life and industrial applications. The scope of wireless communication technologies is rapidly expanding, with widespread adoption ranging from personal mobile communications to the Internet of Things (IoT), 5G networks, and even the upcoming 6G technology. The successful implementation of these technologies not only facilitates the instant exchange of information but also provides the foundation for innovative applications such as smart cities, telemedicine, and autonomous vehicles. In all of this, the wireless channel, as a core component of communication systems, plays a crucial role. The characteristics of the wireless channel and its accurate modeling have a decisive impact on the design of efficient, reliable, and forward-looking communication systems.

[0051] The introduction of Generative Adversarial Networks (GAN) provides a new perspective for wireless channel modeling. GAN provides a new approach to solving two problems in existing wireless channel modeling methods. The structure of GAN is as follows: Figure 1 As shown, GAN can train a generative model by learning from a limited channel measurement dataset. This model can learn the data distribution of the channel dataset, obtain the channel distribution, and generate a large amount of channel data consistent with the statistical characteristics of the real channel environment. Using GAN for channel modeling can not only significantly reduce the amount of channel data required in the modeling process, but also improve the adaptability and flexibility of the model, providing an efficient and cost-effective modeling method for the design and optimization of wireless communication systems.

[0052] Building on this foundation, the Wasserstein GAN (WGAN) effectively mitigates the mode collapse problem in traditional GAN ​​training by adopting the Wasserstein distance as a loss function, improving the quality of generated samples. In particular, an improved version of WGAN, WGAN-GP (WGAN with Gradient Penalty), further stabilizes the training process by introducing a gradient penalty term, enabling the model to generate more realistic and diverse data samples.

[0053] The following combination Figure 2 A wireless channel modeling method of the present invention is described, comprising:

[0054] In step 201, the noise data is input into the generator of the target model. The linear block maps the potential vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features. The convolution transpose block upsamples the abstract high-dimensional features to make them more specific. The specific high-dimensional features are standardized in the batch layer to adjust the mean and variance of the features. The high-dimensional features are compressed by global spatial information, recalibrated by feature channels, and reweighted by features in the channel attention module to obtain the feature reweighted results of each channel. Finally, the wireless channel data output is generated through the activation layer.

[0055] The noise data is in complex form.

[0056] Step 202: The generated wireless channel data and the real channel data are input as input data to the discriminator of the target model. The input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block to output a discrimination result.

[0057] Step 203 : updating the weight parameters of the generator according to the identification result, training the generator to generate wireless channel data whose similarity with the real channel data is less than a preset threshold, thereby achieving effective wireless channel modeling.

[0058] The target model is the WGAN-GP network, an improvement on the GAN. The WGAN-GP network consists of a generator (G) and a discriminator (D). The generator's goal is to generate data that closely resembles the real data distribution, while the discriminator's goal is to distinguish between real and generated data. Ultimately, the generator generates data that closely resembles real data. The discriminator outputs a scalar value, representing the probability that the generated wireless channel data is real channel data.

[0059] This embodiment combines the attention mechanism with the WGAN-GP network and applies it to the channel modeling field to enhance the model's ability to learn channel features. Figure 3 As shown in the figure, the SE-GAN model is formed by introducing a channel attention mechanism (Squeeze-and-Excitation Layer, SEL) based on the WGAN-GP network. This mechanism explicitly models inter-channel dependencies, enhancing the model's ability to express and recalibrate channel features. This significantly improves the model's ability to learn wireless channel characteristics and the authenticity of generated samples. It also adaptively recalibrates the channel's characteristic responses, strengthening the model's capture of important channel features. This effectively improves the accuracy and diversity of wireless channel samples without increasing computational complexity.

[0060] The SE-GAN structure proposed in this embodiment is as follows Figure 4 As shown in Figure 2, the generator consists of linear blocks and convolutional transpose blocks, and introduces a SEL layer with a channel attention mechanism. The linear block primarily maps the latent vector to a high-dimensional feature space. The convolutional transpose block performs convolution and transposition on its input feature map to generate wireless channel data.

[0061] This embodiment introduces the channel attention module SEL into the generator of WGAN-GP. The model can adaptively recalibrate the importance of each feature channel, thereby focusing more on features that are critical to channel modeling. This not only enhances the model's ability to capture channel characteristics, but also improves the quality of the generated channel samples, thereby improving the accuracy and efficiency of wireless channel modeling. In addition, the introduction of the attention mechanism also helps to improve the training stability and efficiency of the model, so that the model can perform channel modeling more flexibly and accurately when facing complex and changing wireless environments. In general, this method of combining the attention mechanism with WGAN-GP provides a new perspective and tool for wireless channel modeling, showing significant improvements and advantages.

[0062] Based on the above embodiment, the high-dimensional features in this embodiment are subjected to global spatial information compression, feature channel recalibration, and feature reweighting in the channel attention module to obtain feature reweighted results for each channel, including:

[0063] Compressing the global spatial information of the feature map of the noise data by an adaptive average pooling operation, reducing the spatial dimensions of each channel of the noise data in height and width to 1, and obtaining the global spatial features of the noise data;

[0064] Passing the global spatial features through a network including a fully connected layer to obtain a weight coefficient for each channel;

[0065] After the weight coefficient of each channel is expanded to the spatial dimension of the corresponding feature map, it is multiplied by the feature map of each channel to obtain the feature re-weighted result of each channel.

[0066] The design philosophy of the Channel Attention Module (SEL) is that different feature channels contribute differently to the model's output, and some channels may contain more valuable information for the current task. Therefore, by introducing a mechanism that adaptively adjusts the importance of each channel, the model can focus more on the more critical feature channels, thereby improving overall model performance. Figure 5 The structure of SEL is shown in Figure 2. SEL can be divided into three parts: global spatial information compression, feature channel recalibration, and feature reweighting.

[0067] Global spatial information compression: SEL first compresses the global spatial information of the input feature map through the AdaptiveAvgPool 2D operation, reducing the spatial dimensions (height and width) of each channel to 1, thereby obtaining a global spatial feature. This step helps capture the global statistical characteristics of each channel and provides a basis for subsequent importance evaluation.

[0068] Feature channel recalibration: The compressed features are passed through a network containing two fully connected layers to obtain the weight coefficient of each channel.

[0069] Feature reweighting: The calculated weight coefficient for each channel is multiplied by the original feature map of each channel to achieve dynamic recalibration of the channel features. The Expand step expands the weight coefficient to the spatial dimensions of the original feature map so that the features of each channel are multiplied by the corresponding weight coefficient.

[0070] This embodiment introduces a channel attention mechanism to explicitly model inter-channel dependencies, enhancing the model's ability to express and recalibrate channel features. This significantly improves the model's ability to learn wireless channel characteristics and the authenticity of generated samples. Furthermore, it adaptively recalibrates the channel's characteristic responses, strengthening the model's capture of important channel features. This effectively improves the accuracy and diversity of wireless channel samples without increasing computational complexity.

[0071] Based on the above embodiment, in this embodiment, the global spatial features are passed through a network including a fully connected layer to obtain a weight coefficient for each channel, including:

[0072] Performing dimensionality reduction on the global spatial features through a first fully connected layer;

[0073] Processing the reduced-dimensional global spatial features through a ReLU activation function to increase the expressive power of the target model;

[0074] Performing dimensionality restoration on the processed global spatial features through a second fully connected layer;

[0075] The global spatial features after dimension restoration are passed through the Sigmoid activation function to obtain the weight coefficient of each channel.

[0076] When recalibrating feature channels, the compressed features are passed through a network consisting of two fully connected layers. The first Linear layer implements feature dimensionality reduction to reduce the number of parameters and computational complexity, while increasing the expressiveness of the model through the ReLU activation function; the second Linear layer is responsible for restoring the feature dimension, and finally outputs the weight coefficient of each channel through the Sigmoid activation function.

[0077] Based on the above embodiment, the linear block in this embodiment maps the latent vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features, including:

[0078] Inputting the noise data into the fully connected layer, batch normalization layer and Leaky ReLU activation function in the linear block in sequence;

[0079] The fully connected layer is used to map the potential vector in the noise data to a feature space of a specific dimension;

[0080] The batch normalization layer is used to perform batch normalization processing on the feature vector of the noise data to accelerate the training process of the target model and stabilize the target model network training;

[0081] The Leaky ReLU activation function is used to introduce nonlinear transformation and enhance the expressive power of the target model.

[0082] like Figure 4 As shown in the figure, the fully connected layer Linear in the linear block maps the input latent vector to a feature vector of a specific dimension. The BatchNorm 1D layer batch normalizes the feature vector to accelerate training and stabilize model training. The Leaky ReLU activation function then introduces nonlinear transformations to enhance the model's expressiveness.

[0083] Based on the above embodiment, the convolution transposition block in this embodiment upsamples the abstract high-dimensional features to make them more specific. The specific high-dimensional features are normalized in the batch layer to adjust the mean and variance of the features, including:

[0084] The feature map output by the linear block is sequentially convolved and transposed through multiple convolutional transposition layers in the convolutional transposition block;

[0085] Each of the convolutional transpose layers except the last convolutional transpose layer in the multiple convolutional transpose layers is sequentially connected to a batch normalization layer, a channel attention module, and a Leaky ReLU activation function;

[0086] The last convolutional transpose layer upsamples the feature map to one channel, and finally outputs the generated wireless channel data through a Tanh activation function.

[0087] like Figure 4As shown in the figure, the convolutional transpose block consists of five convolutional transpose layers. The first four convolutional transpose layers (ConvTranspose2D) are followed by a batch normalization layer (BatchNorm 2D), a SEL layer, and a Leaky ReLU activation function. The fifth convolutional transpose layer upsamples the feature map to one channel and finally outputs the resulting channel data through the Tanh activation function.

[0088] Based on the above embodiment, in this embodiment, the generated wireless channel data and the real channel data are input as input data to the discriminator of the target model. The input data is first expanded in the zero-padding block, then subjected to convolution block feature extraction, and finally discriminated by the linear block to output the identification result, including:

[0089] Passing the real wireless channel data and the generated wireless channel data sequentially through the zero-padding module, the convolution block, and the linear block of the discriminator to obtain the identification result;

[0090] The zero padding module preprocesses the real wireless channel data and the generated wireless channel data through a Zero Padding function;

[0091] The convolution block is used to perform feature extraction on the real wireless channel data and the generated wireless channel data;

[0092] The linear block is used to identify the generated wireless channel data and output the identification result.

[0093] like Figure 4 As shown in Figure 1, the discriminator consists of three main modules: a zero-padding module, a convolutional block, and a linear block. The zero-padding module preprocesses the input signal using a zero-padding function. The height of the input signal is padded and expanded, while the width remains unchanged. This operation aims to preserve the spatial characteristics of the signal and prevent information loss in subsequent convolution operations.

[0094] The convolutional block consists of four convolutional layers and four Leaky ReLU activation functions. The output of the convolutional block is flattened by a view function to form a high-dimensional vector. This vector is then input to the Linear block. The Linear block consists of a fully connected layer that maps the high-dimensional input vector to a one-dimensional output.

[0095] During the forward propagation process, the input signal first passes through the zero-padding module, then passes through the convolution block for feature extraction, and finally passes through the linear block for discrimination. The output discrimination result is a scalar value. This scalar value indicates the probability that the input signal belongs to real channel data or generated channel data.

[0096] Extensive simulation experiments validate the effectiveness of the proposed algorithm. First, similarity and diversity metrics are used to numerically verify the consistency and discreteness of real and generated channel data. The relevant simulation parameter settings are shown in Table 1. The data used comes from Wireless Intelligence, a public channel dataset. The dataset used is a full channel information dataset in a MIMO channel environment with 32 transmit antennas and 4 receive antennas. The channel data consists of 4000 different samples. The samples are in complex form, each with a size of 4096, and the dimensionality is arranged as 4 receive antennas, 32 transmit antennas, and 32 delay spreads.

[0097] Table 1 Simulation parameter settings

[0098] parameter Numerical Carrier frequency 3.5GHz System bandwidth 10MHz Subcarrier spacing 15kHz Number of transmitting antennas 32 Number of receiving antennas 4 Delay spread 32 Number of subcarriers 12

[0099] In order to evaluate the similarity between the channel data distribution generated by the network and the real channel data distribution, the similarity is expressed as:

[0100]

[0101] Where N fake is the number of generated channel samples, N real is the number of real channel samples, and H i are the vectored channels of the j-th generated channel sample and the i-th real channel sample respectively.

[0102] In addition, the diversity of the generated channel data is calculated using variance, and the diversity is expressed as:

[0103]

[0104] Where Var(·) represents the variance calculation of vector l, and l=[l1,l2,……l real Each element l in ] i0 Expressed as:

[0105]

[0106] where arg Max represents the real channel sample index corresponding to the maximum value of the above equation, and N(·) represents the value for all N fake The generated channel sample is taken the number of times the real channel sample index corresponds to i0.

[0107] The numerical results of the similarity and diversity between the channel data generated by SE-GAN and ChannelGAN proposed in this embodiment and the real channel data are shown in Table 2. The greater the similarity between the generated channel data and the real channel data, the smaller the diversity, that is, the smaller the multi / sim, the closer the distribution of the generated false channel data is to the distribution of the real channel data. From the numerical results in Table 2, SE-GAN is superior to ChannelGAN in terms of similarity, and its diversity index is significantly lower than that of ChannelGAN. The comprehensive evaluation index Multi / Sim is much lower than that of ChannelGAN. Therefore, the experimental results show that the SE-GAN model proposed in this embodiment shows higher similarity and lower diversity in generating channel data, and the distribution of the channel data it generates is closer to the distribution of the real channel data, verifying the superiority of SE-GAN in the channel modeling task.

[0108] Table 2 Similarities and differences between different channel modeling and generation schemes

[0109]

[0110] In order to verify the authenticity and validity of the complex channel data generated by the generator, the probability distribution of the amplitudes of the SE-GAN generated data proposed in this embodiment, the generated data of the existing model ChannelGAN, and the real data on the first transmitting antenna, the first receiving antenna, and the first delay spread are compared.

[0111] Figure 6 Shown are the probability distributions of data generated by the SE-GAN model (yellow line), the ChannelGAN model (green line), and the real data (blue line). As can be seen from the figure, the probability distributions of the data generated by the SE-GAN model (yellow line) and the real data (blue line) are relatively similar in overall shape and peak position. The data generated by the SE-GAN has a significant peak near 0, which is consistent with the main peak position of the real data. In addition, the tail decay trend of the SE-GAN-generated data on both the positive and negative sides is also similar to that of the real data, demonstrating that SE-GAN is able to effectively capture the key characteristics of the channel when simulating channel data.

[0112] In contrast, the probability distribution of the data generated by ChannelGAN (green line) also has a peak near 0, but its peak is significantly higher and narrower than the real data and the data generated by SE-GAN. This shows that ChannelGAN has certain deviations in capturing the main characteristics of the channel data. The generated data is too concentrated in certain amplitude values ​​and lacks an accurate simulation of the amplitude distribution in the real data. In addition, the probability distribution of the data generated by ChannelGAN has very obvious two peaks, which is seriously inconsistent with the probability distribution of the real data. Moreover, the tail attenuation trend of the data generated by ChannelGAN on both the positive and negative sides is also significantly different from the real data, further demonstrating its limitations in generating channel data.

[0113] In summary, SE-GAN, by introducing the channel attention module, effectively improves the accuracy of wireless channel data generation. The probability distribution of the generated data is highly consistent with the real data, demonstrating superior performance compared to the existing model, ChannelGAN. These experimental results not only demonstrate the application value of the SE-GAN model in the field of wireless channel modeling and generation, but also provide new directions and ideas for subsequent research.

[0114] The SE-GAN proposed in this embodiment optimizes the network structure, thereby achieving higher efficiency in the process of loading and generating wireless channel data. In the experimental setting, the SE-GAN and ChannelGAN models are required to generate 4,000 data samples respectively under the same conditions to evaluate their time efficiency in the data loading and generation process. The experimental results show that, as shown in Table 3, in the data loading phase, the SE-GAN model only takes 0.746142 seconds, while the ChannelGAN model takes 1.375484 seconds. In addition, in the data generation phase, the SE-GAN model is also more efficient, taking only 0.502677 seconds to complete the generation task, compared to 0.572778 seconds for the ChannelGAN model. These results clearly show that SE-GAN has a significant advantage in data processing speed.

[0115] This significant performance improvement is primarily due to the introduction of the SE module in SE-GAN. By dynamically adjusting feature channels within the network, the SE module effectively increases the network's focus on important information, improving model learning efficiency while also reducing data processing time. Furthermore, SE-GAN's optimized network structure further reduces computational complexity, achieving even greater efficiency in data loading and generation.

[0116] Table 3 Data generation time for different channel modeling and generation schemes

[0117]

[0118] The wireless channel modeling device provided by the present invention is described below. The wireless channel modeling device described below and the wireless channel modeling method described above can be referenced to each other.

[0119] like Figure 7 As shown, the apparatus includes a generation module 701, an identification module 702 and a training module 703, wherein:

[0120] The generation module 701 is used to input the noise data into the generator of the target model. The linear block maps the potential vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features. The convolution transposition block upsamples the abstract high-dimensional features to make them more specific. The specific high-dimensional features are normalized in the batch processing layer to adjust the mean and variance of the features. The high-dimensional features are compressed by global spatial information, recalibrated by feature channels, and reweighted in the channel attention module to obtain the feature reweighted results of each channel. Finally, the activation layer generates wireless channel data output.

[0121] The identification module 702 is used to input the generated wireless channel data and the real channel data as input data to the discriminator of the target model, wherein the input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block to output an identification result;

[0122] The training module 703 is used to update the weight parameters of the generator according to the identification result, and train the generator to generate wireless channel data whose similarity with the real channel data is less than a preset threshold, thereby realizing effective wireless channel modeling.

[0123] This embodiment introduces a channel attention module (SEL) into the WGAN-GP generator. This allows the model to adaptively recalibrate the importance of each feature channel, thereby focusing more on features that are critical for channel modeling. This not only enhances the model's ability to capture channel characteristics but also improves the quality of generated channel samples, thereby enhancing the accuracy and efficiency of wireless channel modeling.

[0124] Figure 8 An example of a physical structure diagram of an electronic device is shown below. Figure 8As shown, the electronic device may include: a processor (processor) 810, a communication interface (Communications Interface) 820, a memory (memory) 830 and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call the logic instructions in the memory 830 to execute the wireless channel modeling method, which includes: inputting noise data into the generator of the target model, the linear block maps the potential vector of the noise data to a higher-dimensional feature space, the convolution transpose block upsamples the abstract high-dimensional features to make them more specific, and the specific high-dimensional features are standardized in the batch layer to adjust the mean and variance of the features; the high-dimensional features are compressed by global spatial information, recalibrated by feature channels, and reweighted by features in the channel attention module to obtain feature reweighted results, and finally the activation layer generates wireless channel data; the generated wireless channel data and the real channel data are input into the discriminator of the target model, the input data is expanded in the zero-padding block, the convolution block feature is extracted, and the linear block is discriminated, and the identification result is output, the weight parameters of the generator are updated according to the identification result, and the generator is trained to generate wireless channel data similar to the real channel data to achieve effective channel modeling.

[0125] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0126] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the wireless channel modeling method provided by the above methods, which includes: inputting noise data into the generator of the target model, the linear block maps the potential vector of the noise data to a higher-dimensional feature space, the convolution transpose block upsamples the abstract high-dimensional features to make them more specific, the specific high-dimensional features are standardized in the batch layer, and the mean and variance of the features are adjusted; the high-dimensional features are compressed by global spatial information, recalibrated by feature channels, and reweighted by features in the channel attention module to obtain feature reweighted results, and finally the activation layer generates wireless channel data; the generated wireless channel data and the real channel data are input into the discriminator of the target model, the input data is expanded in the zero-padding block, the convolution block feature is extracted, and the linear block is discriminated, and the identification result is output, the weight parameters of the generator are updated according to the identification result, and the generator is trained to generate wireless channel data similar to the real channel data to achieve effective channel modeling.

[0127] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the wireless channel modeling method provided by the above-mentioned methods, the method comprising: inputting noise data into the generator of the target model, the linear block mapping the potential vector of the noise data to a higher-dimensional feature space, the convolution transpose block upsampling the abstract high-dimensional features to make them more specific, the specific high-dimensional features being standardized in the batch layer, and the mean and variance of the features being adjusted; the high-dimensional features are subjected to global spatial information compression, feature channel recalibration and feature reweighting in the channel attention module to obtain feature reweighted results, and finally the activation layer generates wireless channel data; the generated wireless channel data and the real channel data are input into the discriminator of the target model, the input data is expanded in the zero-padding block, feature extracted in the convolution block and discriminated in the linear block, and the identification result is output, the weight parameters of the generator are updated according to the identification result, and the generator is trained to generate wireless channel data similar to the real channel data, thereby realizing effective channel modeling.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0129] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A wireless channel modeling method, characterized in that: include: The noise data is input into the generator of the target model. The linear block in the generator maps the latent vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features. The convolution transpose block upsamples the abstract high-dimensional features to make them more specific. The upsampled high-dimensional features are normalized in the batch layer to adjust the mean and variance of the features. The high-dimensional features are compressed by global spatial information, recalibrated by feature channels, and reweighted in the channel attention module to obtain the reweighted feature results of each channel. Finally, the wireless channel data output is generated through the activation layer. The generated wireless channel data and the real channel data are input as input data to the discriminator of the target model, the input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block in the discriminator to output an identification result; The weight parameters of the generator are updated according to the identification result, and the generator is trained to generate wireless channel data whose similarity with the real channel data is less than a preset threshold, thereby realizing effective wireless channel modeling.

2. The wireless channel modeling method according to claim 1, characterized in that The high-dimensional features are subjected to global spatial information compression, feature channel recalibration, and feature reweighting in the channel attention module to obtain feature reweighted results for each channel, including: Compressing the global spatial information of the feature map of the noise data by an adaptive average pooling operation, reducing the spatial dimensions of each channel of the noise data in height and width to 1, and obtaining the global spatial features of the noise data; Passing the global spatial features through a network including a fully connected layer to obtain a weight coefficient for each channel; After the weight coefficient of each channel is expanded to the spatial dimension of the corresponding feature map, it is multiplied by the feature map of each channel to obtain the feature re-weighted result of each channel.

3. The wireless channel modeling method according to claim 2, characterized in that The global spatial features are passed through a network containing a fully connected layer to obtain the weight coefficients of each channel, including: Performing dimensionality reduction on the global spatial features through a first fully connected layer; Processing the reduced-dimensional global spatial features through a ReLU activation function to increase the expressive power of the target model; Performing dimensionality restoration on the processed global spatial features through a second fully connected layer; The global spatial features after dimension restoration are passed through the Sigmoid activation function to obtain the weight coefficient of each channel.

4. The wireless channel modeling method according to claim 1, wherein: The linear block in the generator maps the latent vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features, including: Input the noise data into the fully connected layer, batch normalization layer and LeakyReLU activation function in the linear block in the generator in sequence; The fully connected layer is used to map the potential vector in the noise data to a feature space of a specific dimension; The batch normalization layer is used to perform batch normalization processing on the feature vector of the noise data to accelerate the training process of the target model and stabilize the target model training; The Leaky ReLU activation function is used to introduce nonlinear transformation and enhance the expressive power of the target model.

5. The wireless channel modeling method according to claim 1, wherein: The convolution transposition block upsamples the abstract high-dimensional features to make them more specific. The upsampled high-dimensional features are normalized in the batch layer to adjust the mean and variance of the features, including: Convolutionally and transpose the feature maps output by the linear block in the generator through multiple convolutional transposition layers in the convolutional transposition block; Each of the convolutional transpose layers except the last convolutional transpose layer in the multiple convolutional transpose layers is sequentially connected to a batch normalization layer, a channel attention module, and a Leaky ReLU activation function; The last convolutional transpose layer upsamples the feature map to one channel, and finally outputs the generated wireless channel data through a Tanh activation function.

6. The wireless channel modeling method according to claim 1, wherein: The generated wireless channel data and the real channel data are input as input data to the discriminator of the target model, the input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block in the discriminator to output a discrimination result, including: Passing the real channel data and the generated wireless channel data sequentially through the zero-filling module, the convolution block, and the linear block of the discriminator to obtain the identification result; The zero padding module pre-processes the real channel data and the generated wireless channel data through a Zero Padding function; The convolution block is used to perform feature extraction on the real channel data and the generated wireless channel data; The linear block in the discriminator is used to discriminate the generated wireless channel data and output the discrimination result.

7. A wireless channel modeling device, characterized in that: include: A generation module is configured to input noise data into a generator of a target model. A linear block in the generator maps the latent vector of the noise data to a higher-dimensional feature space to obtain high-dimensional features. A convolutional transposition block upsamples the abstract high-dimensional features to make them more specific. The upsampled high-dimensional features are normalized in a batch layer to adjust the mean and variance of the features. The high-dimensional features are compressed by global spatial information, recalibrated in feature channels, and reweighted in a channel attention module to obtain a reweighted feature result for each channel. Finally, the activation layer generates wireless channel data output. an identification module, configured to input the generated wireless channel data and the real channel data as input data into the discriminator of the target model, wherein the input data is first expanded in a zero-padding block, then subjected to feature extraction by a convolution block, and finally discriminated by a linear block in the discriminator to output an identification result; A training module is used to update the weight parameters of the generator according to the identification result, train the generator to generate wireless channel data whose similarity with the real channel data is less than a preset threshold, and realize effective wireless channel modeling.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the wireless channel modeling method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the wireless channel modeling method according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the wireless channel modeling method according to any one of claims 1 to 6 is implemented.

Images