A communication signal waveform generation method based on SAGAN
Through the self-attention generation adversarial network based on SAGAN, the high cost problem of complex communication signal reconstruction is solved, efficient generation and learning of signals of specific frame structures is realized, and the complexity and cost of traditional methods are reduced.
Patent Information
- Application Number
- CN202310119491.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-15
- Publication Date
- 2025-08-22
- Estimated Expiration
- 2043-02-15
AI Technical Summary
It is difficult for the prior art to effectively reconstruct complex communication signals from non-cooperatives in modern electromagnetic environments, especially those with specific frame structures. Traditional methods require complex parameter measurements and analysis, resulting in high cost and low efficiency.
A self-attention generation adversarial network based on SAGAN is adopted to construct a training data set with a specific frame structure, and an optimized objective function of the generator and discriminator is designed to reduce parameter measurement and analysis of the target signal by generating adversarial training.
Flexible reconstruction of complex signals is realized, labor and time costs are reduced, and learning ability of signal frame structure, position, length and modulation style are improved, so that signals with long range correlation can be generated.
Smart Images
Figure CN116260728B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of wireless communication signal simulation generation, and in particular to a communication signal waveform generation method based on SAGAN (Self-Attention Generative Adversarial Network). Background Art
[0002] Simulating complex electromagnetic environments is an important research area in the current communications field, and communication signal generation is a key step. There are two main approaches to generating communication signals for non-cooperative parties. One is based on parameter measurement and analysis: by capturing target signals in space, the target signal's parameters, such as code rate, modulation style, and carrier frequency, are estimated, and the signal is reconstructed based on the estimated results. The other approach is based on blind reconnaissance and blind interference: by sparsely sampling the target signal, the sparse characteristics of the signal are obtained, and then the sparse characteristics of the signal are used to reconstruct the signal. However, modern electromagnetic environments contain a large number of new and complex communication signals. Traditional methods make it difficult to reconstruct complex communication signals with specific frame structures from non-cooperative parties in space. Summary of the Invention
[0003] The technical problem to be solved by this invention is to provide a communication signal waveform generation method based on SAGAN, which avoids the shortcomings of the above-mentioned background technology. The model of this invention does not require complex parameter measurement and analysis of the target signal, which can reduce manpower and time costs. The model is flexible and easy to train, and can solve the problem of complex signal reconstruction difficulties of non-cooperative parties.
[0004] The technical solution adopted by the present invention is:
[0005] A communication signal waveform generation method based on SAGAN includes the following steps:
[0006] (1) Construct a training dataset with specific frame structure and different signal-to-noise ratios;
[0007] (2) Based on the self-attention generative adversarial network architecture, a network model for one-dimensional signal reconstruction is constructed;
[0008] (3) Design the optimization objective functions of the generator and discriminator in the network model respectively;
[0009] (4) Using the training data set obtained in step (1), the network model is trained based on the training optimization strategy to obtain a trained generator;
[0010] (5) Input random noise into the trained generator to generate reconstructed signal sample data.
[0011] Furthermore, the specific method of step (1) is:
[0012] The constructed frame structure signal samples include a preamble, zero or one postamble, and one or more payloads. There is a pilot block between adjacent payloads. The pilot block evenly divides the information of all payloads. The frame structure signal samples are set to have different signal-to-noise ratios.
[0013] Furthermore, the network model in step (2) includes a generator and a discriminator, where:
[0014] The generator consists of a fully connected layer, six deconvolution layers, and two self-attention layers. There are two types of deconvolution layers: a small deconvolution layer with a kernel size of 3, a stride of 1, and an expansion padding of 1; and a large deconvolution layer with a kernel size of 4, a stride of 2, and an expansion padding of 1. The input of the generator is a random noise with a length of 1 channel and 64 pixels. The random noise first passes through a fully connected layer, and the output data length becomes 8192. The data is then affine transformed into 64 channels and 64 lengths. The transformed data enters the deconvolution layer and passes through three small deconvolution layers and a large deconvolution layer in sequence. Deconvolution layer, a self-attention layer, a large deconvolution layer, a self-attention layer, a large deconvolution layer, and finally the target generated data are obtained; each time the data passes through a small deconvolution layer, the number of output channels is doubled and the length remains unchanged; each time the data passes through a large deconvolution layer, the number of output channels is reduced to half of the original and the length is doubled; each time the data passes through a self-attention layer, the number of output channels and the length of the data remain unchanged; to maintain the stability of model training, the data is batch normalized after passing through the fully connected layer. In addition, except for the last deconvolution layer, the data after passing through the remaining deconvolution layers are first spectrally normalized and then batch normalized;
[0015] The discriminator consists of a fully connected layer, five convolutional layers, and two self-attention layers. There are two types of convolutional layers: a small convolutional layer with a kernel size of 3, a stride of 1, and an expansion padding of 1; and a large convolutional layer with a kernel size of 4, a stride of 2, and an expansion padding of 1. The input of the discriminator is the sample or the data generated by the generator. The input data passes through a small convolutional layer, two large convolutional layers, a self-attention layer, a large convolutional layer, a self-attention layer, and a large convolutional layer in sequence. Each time the data passes through a small convolutional layer, the output is The number of channels becomes twice as large as before, and the length remains unchanged. Every time the data passes through a large convolutional layer, the output channel becomes twice as large as before, and the length becomes half of the original. Every time the data passes through a self-attention layer, the number of output channels and the data length remain unchanged. Finally, the output data of the convolutional layer is affine transformed and input into the fully connected layer. The output length of the data after affine transformation is 512 after passing through the first fully connected layer, and the output length after passing through the second fully connected layer is 1. In order to keep the model training stable, the data is batch normalized after passing through the fully connected layer. The data after the convolutional layer is first spectrally normalized and then batch normalized.
[0016] Furthermore, in the optimization objective function of step (3), Wasserstein distance is used as the loss function. Wasserstein distance represents the minimum cost required to transform a certain distribution P into a target distribution Q.
[0017] The optimization objective function is designed as follows:
[0018] Note x p is the amount of data at different locations in the distribution P, denoted by x q is the amount of data at different locations in distribution Q, and γ is the matrix represented by the current movement plan. Then the total cost of the current movement plan to move data from distribution P to distribution Q is B. The relationship between them is as follows:
[0019]
[0020] Among them, γ(x p ,x q ) represents the specific push behavior, that is, how much data from distribution P will be pushed to distribution Q in this push;
[0021] Then, calculate the distance of the current push, that is, the distance || x from distribution P to distribution Q in this push p -x q The cost of each push is the multiplication of the amount of data moved and the distance moved.
[0022] Add up the cost of each move to get the total cost of the move plan;
[0023] The minimum of the total cost of all movement plans is the bulldozing distance between distribution P and distribution Q, which is calculated as follows:
[0024] W(P,Q)=inf γ∈Π(P,Q) E (x,y)～γ [||xy||]
[0025] Where Π(P,Q) represents the set of all possible joint distributions of distribution P and distribution Q, and inf represents the lower bound;
[0026] When using the bulldozer distance to measure the loss, the objective function of the discriminator is:
[0027]
[0028] The objective function of the generator is:
[0029]
[0030] Among them, D∈1-Lipschitz means that the discriminator satisfies the 1-Lipschitz constraint.
[0031] Furthermore, the specific method of step (4) is:
[0032] The following process is performed until the number of iterations is reached:
[0033] (4-1) From the probability distribution function P of the data set data (x) randomly selects a sample set consisting of m samples x={x (1) ,x (2) ,...,x (m)};
[0034] (4-2) From the prior noise probability distribution function P z Randomly take a noise vector consisting of m noise data in (x)
[0035] (4-3) Input the noise vector z1 to the generator and output the generated signal
[0036] (4-4) Substitute the sample x and the generated signal Input discriminator, the discriminator gives high score to sample x and generates signal assigning low scores;
[0037] (4-5) Fixed generator updates the discriminator parameters θ D ;
[0038] (4-6) From P z Randomly take a noise vector consisting of m noise data in (x)
[0039] (4-7) Input the noise vector z2 to the generator and output the generated signal
[0040] (4-8) will generate a signal Input the discriminator, the generator's goal is to generate a signal that allows the discriminator to give a higher score;
[0041] (4-9) Fixed discriminator updates generator parameters θ G .
[0042] Compared with the prior art, the present invention has the following advantages:
[0043] 1. The present invention is based on a signal reconstruction model of a self-attention generative adversarial network. The generative adversarial network does not require complex parameter measurement and analysis of the signal during the signal generation process. Through adversarial training of the generator and the discriminator, the signal distribution generated by the generator will gradually approach the distribution of the sample signal. This method reduces manpower and time costs. For some new and complex signals and signals that are difficult to measure and analyze, traditional methods cannot reconstruct the target signal, but this method can still generate the signal.
[0044] 2. The signal reconstruction model of the present invention is based on the self-attention generative adversarial network. For communication signals with a specific frame structure, there is a long-range correlation in a signal burst length. The convolution kernel of the convolutional network cannot completely cover the target with a long-range correlation area, which will cause the learned target to lack completeness. The self-attention mechanism can enable the model to pay attention to the global information of the signal, thereby improving the GAN's ability to capture long-range data correlations. It can enable the model to fully learn the frame structure number, position, length and specific symbol information of the target signal, and can learn the modulation style, frequency bandwidth and other characteristics of the signal. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 Figure 2 is a schematic diagram of the sample signal burst structure.
[0046] Figure 2 It is a schematic diagram of the sample burst structure.
[0047] Figure 3 It is a schematic diagram of the generator network.
[0048] Figure 4 It is a schematic diagram of the discriminator network.
[0049] Figure 5 This is a schematic diagram of Wasserstein distance.
[0050] Figure 6 This is a flow chart of a communication signal waveform generation method based on SAGAN in an embodiment of the present invention.
[0051] Figure 7 This is a schematic diagram of the self-attention generative adversarial network model architecture. DETAILED DESCRIPTION
[0052] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0053] Reference Figure 6 and Figure 7 , a communication signal waveform generation method based on SAGAN, comprising the following steps:
[0054] (1) Annotate the actual collected wireless communication signals and construct training data sets with specific frame structures and different signal-to-noise ratios;
[0055] (2) Construct a network for one-dimensional signal reconstruction based on the self-attention generative adversarial network architecture;
[0056] (3) Design optimization objective functions in the generator and discriminator respectively;
[0057] (4) Using the training data set obtained in step (1), the model is trained based on the training optimization strategy to obtain a trained generator;
[0058] (5) Random noise is input into the generator to generate reconstructed signal sample data.
[0059] The specific method for constructing the training data set in step (1) is as follows:
[0060] The general burst structure of linear modulation of the sample signal is as follows Figure 1 As shown in Figure 1. A transmission burst consists of four parts: a preamble (pre), a postamble (post), a pilot block (p), and a payload (load). Each burst contains a preamble, possibly a postamble, and one or more payload sections. The repetition, size, spacing, and content of the pilot blocks are configurable, and the payload sections between them are uniform in size.
[0061] The simulation is carried out using one of the common burst formats. QPSK is selected from the linear modulation burst waveform to generate a training data set. The sample burst length is 64 symbols, of which the preamble is 12 symbols, the postamble is 8 symbols, there are 2 pilot blocks, each of 4 symbols, and the payload is 36 symbols, which are evenly divided by the pilot blocks, each of 12 symbols. The sample burst structure is as follows: Figure 2 shown.
[0062] In constructing the sample set, the symbols for the preamble, postamble, and pilot blocks are fixed, using a specific sequence of 0s and 1s. The data payload is randomly generated, using a random sequence of 0s and 1s. The baseband signal undergoes QPSK modulation. To ensure transmission within the channel, it undergoes shaping filtering to ensure a limited bandwidth. The signal is then shifted to an intermediate frequency (IF) for transmission. The signal in the channel is sampled 8x, with one sample consisting of 512 complex sampling points. The signal is then reduced to zero IF, ultimately generating signal samples with an independent frame structure that conforms to the operating mode and modulation pattern specified by the protocol.
[0063] The specific design method of the self-attention generative adversarial network model architecture in step (2) is as follows:
[0064] The generator is constructed by combining a fully connected layer (FC), a deconvolution layer (Dconv) and a self-attention layer. The input of the generator is 1-channel 64-length random noise. In order to enable the network to represent more features, the length becomes 8192 after a fully connected layer, and then it is batch normalized and then activated by the ReLU function. The data is then affine transformed into 64 channels and 64 lengths. The transformed data enters the deconvolution layer. The convolution kernel size (kernel_size) of the first deconvolution layer is 3, the stride (stride) is 1, and the padding (padding) is 1. After the first convolution kernel, the output channel is 128 and the output data length is 64. In order to make the generator have enough ability to fit the target distribution and compete with the discriminator, the generator has a total of 6 deconvolution layers and 2 self-attention layers. The data must be spectrally normalized after each deconvolution, and then batch normalized. Except for the last layer, the activation function is ReLU. The self-attention layer is placed after the second to last convolution layer and the third to last convolution layer respectively. The number of channels and length of the data input to the self-attention layer do not change after output. After these 6 deconvolution and two self-attention layers, the output is 2 channels and 512 length. Finally, the data output by the Tanh activation function is the generated signal. The generator structure is as follows Figure 3 shown.
[0065] The discriminator is constructed by combining a fully connected layer, a convolution layer (Convolution Layer, Conv) and a self-attention layer. In order to enable the discriminator to better distinguish the feature differences between the sample signal and the generated signal, the discriminator first uses a small convolution kernel of length 3 and a step size of 1. In this way, the discriminator can extract more subtle features of the signal without losing information. Then, a convolution kernel of length 4 and a step size of 2 is used to increase the receptive field. The discriminator inputs a generated signal or sample signal with a length of 2 channels and 512. After passing through 5 convolution layers and two self-attention layers, it outputs 512 channels and 32 length data. After passing through two fully connected layers, it outputs 1 channel and 1 length. The structure of the discriminator is as follows: Figure 4 shown.
[0066] In traditional GANs, the JS divergence is generally used to measure the distribution distance between generated data and real data. If there is no overlap in the distribution space between the two, the JS divergence value will always be log2, and the discriminator will be unable to pass the gradient to the generator, making it difficult to optimize the generative model. This model uses the Wasserstein distance as the loss function. The Wasserstein distance is also called the EM distance (Earth Mover's Distance). The Wasserstein distance can be expressed as the average of the sum of all distances moved when pushing a pile of dirt that conforms to a certain distribution P to another pile of dirt with a target distribution Q.
[0067] like Figure 5 As shown in the figure, there are many ways to transform a pile of soil with distribution P into a pile of soil with distribution Q. Each movement method is called a "moving plan". Each moving plan will generate a corresponding Wasserstein distance. The Wasserstein distance generated by all moving plans is calculated, and the one with the smallest value is selected as the EM distance of the P and Q distributions.
[0068] Note x p is the amount of data at different locations in the distribution P, denoted by x q is the amount of data at different locations in distribution Q, and γ is the matrix representing the current moving plan. Then the total cost of the current moving plans to move data from distribution P to distribution Q is B. The relationship between them is as follows:
[0069]
[0070] Among them, γ(x p ,x q) represents the specific push behavior, that is, how much data from distribution P should be pushed to distribution Q in this push, and then calculates the current push distance, that is, the distance || x from distribution P to distribution Q in this push p -x q The cost of each move is obtained by multiplying the amount of data moved by the distance moved in each push. The total cost of the move plan is obtained by adding up the costs of each move.
[0071] The minimum value of the total cost of all movement plans is the bulldozer distance between distribution P and distribution Q. The EM distance formula between distribution P and distribution Q is as follows:
[0072] W(P,Q)=inf γ∈Π(P,Q) E (x,y)～γ [||xy||]
[0073] Where Π(P,Q) represents the set of all possible joint distributions of distribution P and distribution Q, inf represents the lower bound, E represents the mathematical expectation, and (x,y)~γ represents a combination of (x,y) from the bulldozing behavior γ.
[0074] According to the above principle, when using EM distance to measure the loss, the objective function of the discriminator in GAN is:
[0075]
[0076] Among them, Π(P,Q) represents the set of all possible joint distributions of distribution P and distribution Q, inf represents the lower bound, E represents the mathematical expectation, x~P G Indicates that the data x comes from the data distribution P generated by the generator G ;
[0077] The objective function of the generator is:
[0078]
[0079] Among them, D∈1-Lipschitz means that the discriminator satisfies the 1-Lipschitz constraint, E represents the mathematical expectation, x~P G Indicates that the data x comes from the data distribution P generated by the generator G .
[0080] The model optimization training method in step (4) is specifically as follows:
[0081] During network training, the generator and discriminator are trained alternately to update their parameters. To make the training more stable, the generator is first fixed to train the discriminator, and then the discriminator is fixed to train the generator. In one iteration, the model parameters of the generator and discriminator are updated once respectively. The training process is as follows:
[0082] (4-1) From the probability distribution function P of the data set data (x) randomly selects a sample set consisting of m samples x={x (1) ,x (2) ,...,x (m)};
[0083] (4-2) From the prior noise probability distribution function P z Randomly take a noise vector consisting of m noise data in (x)
[0084] (4-3) Input the noise vector z1 to the generator and output the generated signal
[0085] (4-4) Substitute the sample x and the generated signal Input discriminator, the discriminator gives high score to sample x and generates signal assigning low scores;
[0086] (4-5) Fixed generator updates the discriminator parameters θ D ;
[0087] (4-6) From P z Randomly take a noise vector consisting of m noise data in (x)
[0088] (4-7) Input the noise vector z2 to the generator and output the generated signal
[0089] (4-8) will generate a signal Input the discriminator, the generator's goal is to generate a signal that allows the discriminator to give a higher score;
[0090] (4-9) Fixed discriminator updates generator parameters θ G .
[0091] In summary, the present invention constructs a communication signal dataset with a specific frame structure, designs a generator and discriminator network architecture containing a self-attention mechanism, and designs appropriate optimization objective functions for the generator and discriminator to achieve algorithm convergence. The self-attention mechanism of the present invention can improve the model's ability to capture long-range correlations in data. Through game training between the generator and discriminator, the generator model can learn the characteristics of specific frame signals, thereby realizing signal reconstruction.
Claims
1. A communication signal waveform generation method based on SAGAN, characterized in that: The following steps are involved: (1) Construct a training dataset with specific frame structure and different signal-to-noise ratios; (2) Based on the self-attention generative adversarial network architecture, a network model for one-dimensional signal reconstruction is constructed; (3) Design the optimization objective functions of the generator and discriminator in the network model respectively; in the optimization objective function, the Wasserstein distance is used as the loss function. The Wasserstein distance represents the minimum cost required to transform a certain distribution P into the target distribution Q; The optimization objective function is designed as follows: Note x p is the amount of data at different locations in the distribution P, denoted by x q is the amount of data at different locations in distribution Q, and γ is the matrix represented by the current movement plan. Then the total cost of the current movement plan to move data from distribution P to distribution Q is B. The relationship between them is as follows: Among them, γ(x p ,x q ) represents the specific push behavior, that is, how much data from distribution P will be pushed to distribution Q in this push; Then, calculate the distance of the current push, that is, the distance || x from distribution P to distribution Q in this push p -x q The cost of each push is the multiplication of the amount of data moved and the distance moved. Add up the cost of each move to get the total cost of the move plan; The minimum of the total cost of all movement plans is the bulldozing distance between distribution P and distribution Q, which is calculated as follows: W(P,Q)=inf γ∈Π(P,Q) E (x,y)～γ [||x-y||] Where Π(P,Q) represents the set of all possible joint distributions of distribution P and distribution Q, inf represents the lower bound, E represents the mathematical expectation, and (x,y)~γ represents a combination of (x,y) from the bulldozing behavior γ. When using the bulldozer distance to measure the loss, the objective function of the discriminator is: The objective function of the generator is: Among them, D∈1-Lipschitz means that the discriminator satisfies the 1-Lipschitz constraint, E represents the mathematical expectation, x~P G Indicates that the data x comes from the data distribution P generated by the generator G ; (4) Using the training data set obtained in step (1), the network model is trained based on the training optimization strategy to obtain a trained generator; (5) Input random noise into the trained generator to generate reconstructed signal sample data.
2. The method for generating a communication signal waveform based on SAGAN according to claim 1, wherein: The specific method of step (1) is: The constructed frame structure signal samples include a preamble, zero or one postamble, and one or more payloads. There is a pilot block between adjacent payloads. The pilot block evenly divides the information of all payloads. The frame structure signal samples are set to have different signal-to-noise ratios.
3. The communication signal waveform generation method based on SAGAN according to claim 1, characterized in that: The network model in step (2) includes a generator and a discriminator, where: The generator consists of a fully connected layer, six deconvolution layers, and two self-attention layers. There are two types of deconvolution layers: a small deconvolution layer with a kernel size of 3, a stride of 1, and an expansion padding of 1; and a large deconvolution layer with a kernel size of 4, a stride of 2, and an expansion padding of 1. The input of the generator is a random noise with a length of 1 channel and 64 pixels. The random noise first passes through a fully connected layer, and the output data length becomes 8192. The data is then affine transformed into 64 channels and 64 lengths. The transformed data enters the deconvolution layer and passes through three small deconvolution layers and a large deconvolution layer in sequence. Deconvolution layer, a self-attention layer, a large deconvolution layer, a self-attention layer, a large deconvolution layer, and finally the target generated data are obtained; each time the data passes through a small deconvolution layer, the number of output channels is doubled and the length remains unchanged; each time the data passes through a large deconvolution layer, the number of output channels is reduced to half of the original and the length is doubled; each time the data passes through a self-attention layer, the number of output channels and the length of the data remain unchanged; to maintain the stability of model training, the data is batch normalized after passing through the fully connected layer. In addition, except for the last deconvolution layer, the data after passing through the remaining deconvolution layers are first spectrally normalized and then batch normalized; The discriminator consists of a fully connected layer, five convolutional layers, and two self-attention layers. There are two types of convolutional layers: a small convolutional layer with a kernel size of 3, a stride of 1, and an expansion padding of 1; and a large convolutional layer with a kernel size of 4, a stride of 2, and an expansion padding of 1. The input of the discriminator is the sample or the data generated by the generator. The input data passes through a small convolutional layer, two large convolutional layers, a self-attention layer, a large convolutional layer, a self-attention layer, and a large convolutional layer in sequence. Each time the data passes through a small convolutional layer, the output is The number of channels becomes twice as large as before, and the length remains unchanged. Every time the data passes through a large convolutional layer, the output channel becomes twice as large as before, and the length becomes half of the original. Every time the data passes through a self-attention layer, the number of output channels and the data length remain unchanged. Finally, the output data of the convolutional layer is affine transformed and input into the fully connected layer. The output length of the data after affine transformation is 512 after passing through the first fully connected layer, and the output length after passing through the second fully connected layer is 1. In order to keep the model training stable, the data is batch normalized after passing through the fully connected layer. The data after the convolutional layer is first spectrally normalized and then batch normalized.
4. The method for generating a communication signal waveform based on SAGAN according to claim 1, wherein: The specific method of step (4) is: The following process is performed until the number of iterations is reached: (4-1) From the probability distribution function P of the data set data (x) randomly selects a sample set consisting of m samples x={x (1) ,x (2) ,...,x (m) }; (4-2) From the prior noise probability distribution function P z Randomly take a noise vector consisting of m noise data in (x) (4-3) Input the noise vector z1 to the generator and output the generated signal (4-4) Substitute the sample x and the generated signal Input discriminator, the discriminator gives high score to sample x and generates signal assigning low scores; (4-5) Fixed generator updates the discriminator parameters θ D ; (4-6) From P z Randomly take a noise vector consisting of m noise data in (x) (4-7) Input the noise vector z2 to the generator and output the generated signal (4-8) will generate a signal Input the discriminator, the generator's goal is to generate a signal that allows the discriminator to give a higher score; (4-9) Fixed discriminator updates generator parameters θ G .