Small sample channel state information data expansion method based on generative adversarial network

Abstract

The invention provides a small sample channel state information data expansion method based on a generative adversarial network, and the method comprises the steps: generating a vivid time-frequency domain channel sample through a Wasserstein generative adversarial network based on physical constraints under the condition of rare or missing drive test data; and a time-frequency joint discriminator is combined to effectively distinguish real and forged signals. Statistical authenticity of a generated sample is improved through joint time-frequency domain feature discrimination, a physical constraint module is embedded in a generator, and channel physical characteristics are maintained through forced power normalization and a spectrum matching mechanism. A GAN framework with gradient penalty is adopted in the whole training process, and the training stability is ensured. And data sample generation when the nodes are quickly deployed to a new environment is facilitated.

Description

Technical Field

[0001] This invention belongs to the field of physical layer security, and in particular relates to a method for amplifying small sample channel data when nodes are rapidly deployed to new environments. Background Technology

[0002] With the large-scale deployment of the Internet of Things (IoT) and Industrial Internet of Things (IIoT), network nodes are often rapidly introduced into unknown or unmeasured channel environments to support applications such as smart manufacturing, smart cities, and environmental monitoring. In these scenarios, due to short deployment cycles and high operation and maintenance costs, sufficient measured channel state information (CSI) is often unavailable for model training, leading to the failure of physical layer authentication and performance prediction models based on measured data.

[0003] To address the problem of small sample data, academia and industry have proposed various data augmentation and generation techniques. These include using classic path loss models, such as COST231-Hata, and multipath fading distributions, such as Rayleigh, Rician, and Nakagami, to generate pseudo-channel samples. While these can reflect macroscopic fading characteristics, they tend to overlook time-frequency domain details, making it difficult to meet the requirements of high-precision modeling. Interpolation and resampling techniques are also used: these extend a small number of CSI samples through temporal or spatial interpolation. These methods are simple and do not incur the overhead of generating a model, but when the number of original samples is extremely small, interpolation errors can cause the new samples to deviate from the true scores. Alternatively, classic machine learning generation methods can be used: these utilize techniques such as dictionary learning and sparse coding to construct a channel feature dictionary and generate new data based on a limited number of samples. These methods can partially preserve the channel structure, but they rely on manually designed features, limiting the generation effect.

[0004] In recent years, GANs have been widely used in the field of few-shot augmentation, learning data distribution through adversarial training. Although typical GAN ​​schemes can generate realistic samples, they often produce pseudo-samples that violate power spectrum or time delay diffusion laws when physical constraints are lacking. Summary of the Invention

[0005] This invention addresses the challenge of insufficient sample generation for nodes deployed in unknown or data-scarce channel environments. It also addresses the lack of a systematic, end-to-end solution for augmenting Channel State Information (CSI) data with small samples, a solution that can guarantee the statistical accuracy of generated samples even when data is extremely scarce or missing, without requiring extensive prior knowledge or manual feature design. Therefore, the technical solution this invention aims to provide is an efficient method for augmenting small-sample CSI data to support subsequent physical layer authentication and performance evaluation models.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is a method for augmenting small-sample channel state information data based on generative adversarial networks, comprising:

[0007] By designing a generator that includes a deconvolutional upsampling layer and a dedicated physical constraint module, random noise is mapped into high-fidelity time-domain and frequency-domain signal samples that conform to the power spectrum and spectral distribution of the wireless channel.

[0008] By introducing a time-frequency joint convolution structure at the discriminator end, the network can distinguish the authenticity of input samples from both time-domain and frequency-domain features. The gradient penalty mechanism ensures that the overall network satisfies 1-Lipschitz continuity, thereby effectively suppressing mode collapse and improving the quality of generated samples during training.

[0009] After co-optimization of the generator and discriminator, it is possible to synthesize statistically consistent channel data with multiple fading characteristics, even with only limited measured CSI data.

[0010] Subsequently, these synthetic samples are fused with the original scarce data to form an augmented dataset to support the training of subsequent physical layer authentication, channel estimation, or other deep learning models.

[0011] To verify the universality and stability of this scheme, this invention compared the power spectral density, autocorrelation function, and constellation diagram of the generated samples and the original samples in various classic channel distributions such as Rayleigh, Rician, and Nakagami. Furthermore, the performance changes during the model training process were evaluated through experiments in different training rounds. The results show that the generated samples are highly consistent in various statistical indicators and can effectively improve the accuracy and robustness of subsequent classification or regression models.

[0012] Wasserstein GAN (WGAN) with physical constraints: To improve generation stability and constrain physical properties, researchers introduced gradient penalty and physical constraint modules such as spectrum matching and power normalization into the WGAN architecture, so that the generated data can better reflect the real channel environment while satisfying statistical characteristics.

[0013] The beneficial effect of this invention is that, by introducing a physically constrained WGAN model, it achieves the generation of high-quality, high-reliability small-sample channel data under conditions of extremely scarce or missing measured channel data, and has the following significant advantages:

[0014] High consistency: The generated samples strictly follow the channel power spectrum and frequency domain distribution characteristics, improving the consistency between simulation data and the real environment;

[0015] Accuracy: The time-frequency joint discriminator can effectively identify spurious samples and avoid model collapse and pattern ambiguity;

[0016] Stability: The fusion of synthetic samples and limited experimental data expands the training set size, significantly improving the performance and robustness of subsequent deep learning or traditional classification and regression models;

[0017] Low dependency: It achieves an end-to-end small sample channel augmentation process without relying on a large amount of prior knowledge or manual feature design. Attached Figure Description

[0018] Figure 1 WGAN flowchart;

[0019] Figure 2 For generator network structure;

[0020] Figure 3 To compare the generated results with the power spectral density of the original data

[0021] Figure 4 A comparison graph of autocorrelation functions;

[0022] Figure 5 Comparison with constellation charts;

[0023] Figure 6 Comparison of the probability density functions of the original channel and the simulated channel for the Nakagami distribution with shape parameter m=1;

[0024] Figure 7 Comparison of the original and simulated channel constellation diagrams for the Nakagami distribution with shape parameter m=1;

[0025] Figure 8 Comparison of the probability density functions of the original and simulated channels for the Nakagami distribution with shape parameter m=5;

[0026] Figure 9 Comparison of the original and simulated channel constellation diagrams for the Nakagami distribution with shape parameter m=5;

[0027] Figure 10 Comparison of model simulation performance under different training epochs; (a) epoch=1; (b) epoch=10; (c) epoch=20. Detailed Implementation

[0028] Generative Adversarial Networks (GANs) consist of two parts: a generator and a discriminator, which are trained through a game-like process. The generator attempts to capture the distribution of real data and generate fake samples that are difficult to distinguish from real samples. The discriminator then distinguishes real data from generated data among the input samples.

[0029] The GAN architecture of this invention employs an improved design with joint spatiotemporal constraints, consisting of a time-frequency dual-domain discriminator (TFDiscriminator) and a physical constraint generator (Complex Generator). Compared to traditional GANs, this model introduces a frequency domain feature analysis path into the discriminator, improving the statistical realism of generated samples through joint time-frequency domain feature discrimination. Furthermore, a physical constraint module is embedded in the generator, preserving channel physical characteristics through forced power normalization and spectrum matching mechanisms. The entire training process utilizes the Wasserstein GAN (WGAN) framework with gradient penalty to ensure training stability.

[0030] In traditional GAN ​​architectures, JS divergence is used to measure the generation distribution P. g The difference between the JS divergence and the true distribution Pr is significant, but the gradient of the JS divergence is zero when the two distributions do not overlap, leading to training instability. WGAN outputs do not use the Sigmoid function but instead use the Wasserstein distance, and the update value is limited to a range on each update to alleviate the mode collapse problem and achieve more stable convergence. Then, by introducing a gradient penalty in the discriminator loss, 1-Lipschitz continuity is effectively maintained, improving the quality and diversity of generated samples.

[0031]

[0032] W is the Wasserstein distance, sup is the minimum upper bound, and ||f|| L ≤1 indicates that the Lipschitz constant of the 1-Lipschitz function f does not exceed 1. Let x be a sample obtained by sampling from the true distribution Pr, and let f be the expectation of the 1-Lipschitz function.

[0033] The overall structure of the scheme is as follows Figure 1 As shown in the diagram. First, the original channel data is input. The raw data is normalized using the Standard Scaler algorithm to standardize the data and eliminate outliers. It's worth noting that using the Min MaxScaler algorithm can prevent the generator from learning correctly because extreme values ​​or long-tailed distributions in the channel can cause the normalized values ​​to concentrate in a certain range, making it impossible for the generator to generate data properly. Then, the generator receives random noise z and generates fake channel data G(z). By maximizing the discriminator's false positive rate, it learns a generation strategy that more closely approximates the real channel distribution. The discriminator then distinguishes between the real and generated data, improving its ability to differentiate between real and generated samples by minimizing the false positive rate.

[0034] The generator aims to learn the underlying distribution of channel data from the noise space and generate fake samples with temporal and frequency domain features. In this invention, the generator combines a deconvolutional network (DC) with a physical constraint module. The DC is used to progressively upsample random noise and recover the details of the time-series signal layer by layer. The DC first maps the random noise to a high-dimensional feature space through linear transformation and reshape, then forms an initial feature tensor through one-dimensional deconvolution (ConvTranspose1d), and finally uses the GELU activation function to enhance nonlinear expressiveness. Subsequently, feature upsampling is achieved through four layers of one-dimensional deconvolution, with each layer doubling the feature length, ultimately outputting two-dimensional features of the synthesized channel. The physical constraint module introduces spectral consistency constraints to ensure the physical rationality of the generated channel data, making the generated data conform to the target power spectral density in the frequency domain.

[0035] The workflow is as follows: First, random noise is input and mapped to a high-dimensional feature space through linear transformation and shape reshaping. The input random noise vector with a dimension of 100 is expanded into an initial feature tensor with 256 channels × 64 time steps. The GELU activation function is used to enhance the nonlinear expressive power.

[0036] Then, feature upsampling is achieved through four layers of deconvolution, with each deconvolution layer having a stride of 2 to double the feature length, ultimately outputting two-dimensional features (real part + imaginary part) over 1024 time steps. Specifically, after the deconvolutional network, a dimensionality permutation Permute is followed by a Physical Constraint module, which uses power normalization and spectral matching loss, where S... target The preset channel power spectrum ensures that the generated signal conforms to the propagation characteristics of the wireless channel.

[0037] Physical constraints are then applied, the spectrum of the generated data is calculated and compared with the target spectrum, the loss is calculated using mean squared error (MSE), and the generator parameters are updated accordingly. Finally, a forged channel data sequence is output. The figure below illustrates the network structure of the generator, from input random noise to the generation of a complex channel matrix.

[0038] The discriminator's task is to perform binary classification of the input data by learning the time-domain and frequency-domain features of the data. To better capture the characteristics of wireless channel data in the time and frequency domains, the discriminator in this paper is designed as a time-frequency joint convolutional network.

[0039] First, temporal features of the channel data are extracted using a temporal convolution module with a kernel size of 3. This part employs the Leaky ReLU activation function, and downsampling is performed layer by layer to improve feature abstraction. The Leaky ReLU activation function introduces a small slope in the negative interval, preventing permanent neuron inactivation and improving the discriminator's performance. After two convolutions, one-dimensional adaptive average pooling is used, and the output size is fixed to a specified target length. Subsequently, a flattening operation is performed to flatten the multidimensional data into a one-dimensional vector, outputting the temporal features.

[0040] Next, a frequency domain convolution module is used to perform a Fourier transform on the input data to obtain spectral features. After a logarithmic transform, frequency domain information is extracted through a convolutional layer. Subsequently, one-dimensional adaptive average pooling and flattening operations are used to unify the feature dimensions of both. Finally, a fully connected layer is used to fuse the time domain and frequency domain features, and the rectified result is output through a fully connected layer.

[0041] This time-frequency joint discrimination method enables the discriminator to capture the characteristics of channel data more comprehensively, thereby effectively improving the ability to distinguish between real and fake samples. It is also more stable and avoids the gradient vanishing problem, ensuring the effectiveness of the discriminator.

[0042] To demonstrate the model's ability to scale up small samples, this section compares the distribution of data samples generated by the WGAN model under different dataset types with the original datasets. First, experiments were conducted based on the Rayleigh channel dataset generated earlier. Because the dataset was split during simulation, with the first 80% of the data input into the model for training, the actual validation compared the remaining 20% ​​after the split with the generated data to verify the model's temporal relevance.

[0043] Figure 3 The figure shows a comparison of the power spectral density of the channel dataset generated by WGAN and the original dataset. The orange portion represents the power spectral density of the generated dataset, and the blue portion represents the power spectral density of the original dataset. The dashed line depicts the envelope of the dataset. Simulation experiments show that the variance of the channel in the original dataset is σ = 2.12, while that in the generated dataset is σ = 2.11. It can be seen that the distributions of the two datasets are basically consistent.

[0044] Figure 4The graph shows a comparison of the simulated and actual autocorrelation functions. The horizontal axis represents the lag order, ranging from 0 to 200, representing the autocorrelation of the signal at different lag steps. 0 indicates the signal's correlation with itself. 1 indicates the correlation between the current signal and the previous time step. 200 indicates the correlation between the current signal and 200 time steps ago. The vertical axis represents the autocorrelation value, which is the normalized autocorrelation value, generally ranging from [-1, 1]. It reflects the similarity of the signal over time; the closer the value is to 1, the more similar the signal is between the current time step and the lag time step; the closer the value is to 0, the lower the correlation; a negative value indicates a negative correlation at that lag time step. The graph shows that the generated channel has a similar correlation to the original channel.

[0045] Figure 5 This section presents a comparison of simulated and actual constellation diagrams, a tool used to visualize signal quality. The horizontal axis represents the in-phase (real part) of the signal, and the vertical axis represents the quadrature (imaginary part). Each point represents the position of a modulation symbol in the IQ plane. Constellation diagrams can reflect whether the data generated by GANs retains the modulation characteristics and channel effects of the original signal. Figure 5 (a) shows the constellation diagram of the original channel. Figure 5 (b) shows the constellation diagram of the simulated channel. It can be seen that the simulated channel retains the modulation characteristics and channel effects of the original channel, and the two are basically consistent.

[0046] Furthermore, to demonstrate that the model can simulate different parameters for different distributions, relevant experiments were conducted for verification. The experiments generated channels with different shape parameters under the Nakagami distribution and compared them with the original data for verification.

[0047] Figure 6 , Figure 7 Simulation results for the channel with shape parameter m=1 and m=1.02 for the Nakagami distribution are presented respectively. Figure 8 , Figure 9 Simulation results for the channel with shape parameter m=5 for the Nakagami distribution are presented, and the simulated channel m=5.17 is also shown. It can be seen that the model can simulate channel parameters under arbitrary conditions, proving the model's reliability.

[0048] To demonstrate the changes during model training Figure 10The diagram shows a comparison between simulation results and actual data at different training epochs. With one training epoch, the shape parameter of the simulated dataset generated by the model is 1.05; with 10 training epochs, the shape parameter is 1.03; and with 20 training epochs, the shape parameter is 1.02. This demonstrates the effectiveness of model training in changing the data.

Claims

1. A method for augmenting small-sample channel state information data based on generative adversarial networks, characterized in that, include: The input random noise is progressively upsampled through a deconvolutional network. Each deconvolutional layer is followed by an activation function to enhance nonlinear expressive power. A physical constraint module is connected after the output layer of the deconvolutional network. The physical constraint module outputs synthetic channel state information that conforms to the power spectrum and spectrum distribution of the wireless channel through power normalization and spectrum matching loss. The deconvolutional network and the physical constraint module together form a generator. Synthetic channel state information is introduced as input into a discriminator with a joint time-frequency convolutional structure. The discriminator judges the authenticity of the synthetic channel state information from two dimensions: time domain features and frequency domain features. The gradient penalty mechanism ensures that the generator network composed of the generator and the discriminator satisfies 1-Lipschitz continuity. Synthetic channel state information is fused with the original channel state information data to form an augmented dataset, which supports the training of subsequent physical layer authentication, channel estimation or other deep learning models.

2. The method as described in claim 1, characterized in that, The deconvolutional network process is as follows: first, random noise is mapped to a high-dimensional feature space through linear transformation and shape reshaping for spatial expansion; then, an initial feature tensor is formed through one-dimensional deconvolution; and finally, the GELU activation function is used to enhance the nonlinear expression capability. Subsequently, feature upsampling is achieved through four layers of one-dimensional deconvolution, with each layer of one-dimensional deconvolution doubling the feature length, ultimately outputting the two-dimensional features of the synthesized channel.

3. The method as described in claim 2, characterized in that, A random noise vector of dimension 100 is expanded into an initial feature tensor of 256 channels × 64 time steps through linear transformation and shape reshaping. The nonlinear expressive power is enhanced by GELU activation function and then passed through four layers of deconvolution with a stride of 2 for each layer, finally outputting two-dimensional features of 1024 time steps.

4. The method as described in claim 1, characterized in that, During network training, the spectrum of the synthesized channel state information is compared with the target spectrum, the loss is calculated using mean square error, and the generator parameters are updated accordingly.

5. The method as described in claim 1, characterized in that, The processing procedure of the discriminator with the time-frequency joint convolution structure is as follows: The temporal features of the synthesized channel state information are extracted by the temporal convolution module, and then a one-dimensional vector of a specified target length is output as the temporal features through one-dimensional adaptive average pooling and flattening operations. The synthesized channel state information is Fourier transformed by the frequency domain convolution module to obtain spectral features. Then, the frequency domain information is extracted by logarithmic transformation and convolutional layer. Subsequently, one-dimensional adaptive average pooling and flattening operations are used to output a one-dimensional vector of a specified target length as frequency domain features. Finally, the time-domain features and frequency-domain features are fused through a fully connected layer, and after rectification, the binary classification result of whether it is true or false is output through the fully connected layer.

6. The method as described in claim 5, characterized in that, In the temporal convolution module, the kernel size is 3, and the Leaky ReLU activation function is used. The Leaky ReLU activation function avoids permanent neuron deactivation by introducing a small slope in the negative interval.

Images