A Radio Signal Generation Method Based on Generative Adversarial Networks

Abstract

The present invention relates to a radio signal generation method based on a generative adversarial network. By using the generative adversarial network, the present invention constructs a complete generator and discriminator system, divides the original signal in a certain proportion as the training samples of the network structure, adjusts the network structure parameters by using the training data, designs a new loss metric function according to the authenticity of the original signal data, further measures the similarity loss between the real signal data and the synthetic signal data, and uses this loss function to train the generator. Finally, the learning of the original modulated radio signal is completed, and new synthetic signal data is generated by adding random noise. The synthetic signal generated by the present invention is highly similar to the original signal and has a high recognition rate in the signal receiver.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and radio, and in particular to a radio signal generation method based on a generative adversarial network. Background Art

[0002] OFDM (Orthogonal Frequency Division Multiplexing) is a modulation technique for digital communication. It divides a high-speed data stream into multiple subcarriers for transmission. The modulation method for each subcarrier can be QPSK, 16QAM, 64QAM, etc. By combining multiple low-speed signals into a high-speed signal, a high-speed data stream is formed. This high-speed data stream is sent to the receiver, and the receiver demodulates and restores the original digital data through the reverse process, achieving efficient spectrum utilization. OFDM technology is widely used in fields such as digital television, wireless local area network (WLAN), and mobile communication.

[0003] The generative adversarial network (GAN) is a deep learning model proposed by Ian Goodfellow et al. in 2014. It can generate highly realistic new data, including images, audio, video, etc. The core idea of GAN is to gradually let the generator learn the feature distribution of real data in an adversarial manner, so as to generate new data similar to the real data. During the training process of the GAN model, the generator and the discriminator compete with each other. The goal of the generator is to generate data that is realistic enough to deceive the discriminator so that it cannot distinguish between real data and generated data. The goal of the discriminator is to distinguish between real data and generated data and judge the source of the data as accurately as possible. Through this competitive relationship, the generator continuously optimizes its generation ability until the generated data cannot be distinguished by the discriminator. Currently, the generative adversarial network is mainly used in fields such as images, videos, and audio. However, the forged signals generated by the existing technology have a large difference from the real signals and are easily recognized. Therefore, using the generative adversarial network to generate signals is a feasible method. Summary of the Invention

[0004] The purpose of the present invention is to solve the technical problem of low reception rate of synthetic signals in the fields of artificial intelligence and signal generation, and to provide a radio signal generation method based on a generative adversarial network. This method does not require complex processing of the original signal, can adaptively extract the features of different signals, and generate signals of the same type. After testing, the generated signals can achieve a reception rate approximate to that of the original signals.

[0005] To achieve the above object, the technical solution of the present invention is: a radio signal generation method based on a generative adversarial network. A signal transmitter generates an original signal, adds an interference signal to the original signal, and then puts it into the generative adversarial network model for learning. Through the training of the generative adversarial network model, a synthetic data signal similar to the original signal is generated. The method specifically includes: generating an OFDM signal; constructing a generative adversarial network model; training the generative adversarial network model.

[0006] The content of the present invention mainly focuses on establishing a generative adversarial neural network, which mainly consists of a generator network and a discriminator network. As Figure 1 shown, the generator inputs real signal data, synthesizes signal data according to its distribution and characteristics, and adds random noise. The discriminator distinguishes whether the signal samples come from the training set or the signal generator to train the signal generator. The new signal generated by it can replace the original signal, and specifically includes the following steps:

[0007] 1. Generate an OFDM signal: Generate an OFDM signal through an OFDM signal transmitter, where the bit signal-to-noise ratio EbNoVec = -10, the modulation order M = 4, the number of bits transmitted per symbol k = log2(M), the total number of subcarriers numSC = 2048, the symbol signal-to-noise ratio snrVec = EbNoVec + 10 * log10(k) + 10 * log10(numDC / numSC), and numDC is the number of data subcarriers in OFDM. The OFDM signal is expressed as x = [x1,...x i ,...x n , x i = (a i , b i ) is the i-th data of x, and a i and b i are the real part and the imaginary part of x i respectively.

[0008] 2. Construct a generative adversarial network model

[0009] Define the generator model: As Figure 2 shown, the generator consists of five functional layers: an input layer, an embedding layer, a feature fusion layer, an LSTM layer, and a regression layer. The generator first takes the encoded real OFDM signal data and random noise as inputs. For the real part and the imaginary part of the signal, each pair of them is embedded using a multi-layer perceptron (MLP) to obtain a 64-dimensional vector: e i = φ s (Δa i , Δb i ; W es ), where φ s is the rectified linear unit (ReLU) of the MLP with an activation function for embedding data. is the mean value of a i , and is the mean value of b i . W es is the weight matrix of the multi-layer perceptron for embedding signal data. After the embedding process, all vectors and random noise are further concatenated, and then they are fused into a 100-dimensional vector using a dense layer. Then, using the distribution and type characteristics of each data, and finally through feature fusion, to support the modeling of the network and the generation of signals. In the LSTM modeling layer, a many-to-many LSTM structure is used, taking a sequence with a specific time step as input and generating a sequence with the same time step as output. Given the size of the fused features, 100 units are allocated in the LSTM model, and the fused features are fed into the model: H = LSTM(F; W lstm ), where F represents the fused features of all data in the signal data, that is, F = [f0, f1,..., f maxlength-1 , f i is the fused feature vector of the i-th data, H is the output of the LSTM model, having the same time step dimension as the input, that is, H = [h0, h1,..., h maxlength-1 , h i is the modeling output of f i , and W lstm is the weight matrix of the LSTM model. Finally, the synthetic signal data decoded from the output H of the LSTM modeling layer is processed, where each feature vector h i in H is a 100-dimensional vector containing the distribution characteristics of the synthetic signal data. A dense layer with two units and the hyperbolic tangent function (tanh) are used to decode the deviations of the real and imaginary parts of the signal data. In addition, to further expand the output range to ensure that its range covers all possible deviation values, a dense layer and the softmax function are used to restore the one-hot representation of these data: (Δa′ i , Δb′ i ) = D s (h i ; W ds ), where Δa′ i and Δb′ i are the deviations of the real and imaginary parts of the i-th synthetic signal data, D s represents the dense layer of the tanh or softmax function with the decoded distribution feature attributes, and W ds is the decoding weight matrix of these dense layers (the decoding weight matrix is shared among all data).

[0010] Define the discriminator model: As shown in Figure 3 , the discriminator has a network structure similar to that of the generator, consisting of an input layer, an embedding layer, a feature fusion layer, and an LSTM layer. The differences from the generator are as follows: (1) The input layer of the discriminator only takes signal data as input without adding random noise. (2) The discriminator uses a many-to-one LSTM model, which takes features with time steps as input and outputs a scalar: where F represents the fused features of all signal data, F = [f0, f1,..., f maxlength-1 , f i is the fused feature vector of the i-th data, is the weight matrix of the LSTM model, and h is the output scalar of the LSTM model. (3) A single-unit dense layer with a sigmoid function is used to perform binary classification (real data or synthetic data) on the scalar output: O d = D bc (h; W bc ), where D bc is a single-unit dense layer with a sigmoid function, W bc is its weight matrix, and O d is the final output of the discriminator.

[0011] Define the loss function: The original generative adversarial network is designed to optimize the following objective function:

[0012]

[0013] where p data (x) represents the distribution of real sample data, p z (z) represents the prior of the noise variable, D(x) represents the probability that x comes from p data (x), and G(z) represents the mapping from p z (z) to p data (x). The goal of the generator is to minimize The goal of the discriminator is to maximize This results in a two-player min-max game. According to the objective function O(D, G), the loss function of the discriminator can be considered as the binary cross-entropy (BDE) loss function (L BCE ), which will also be used to train the generator. Different from the original generative adversarial network, the present invention requires real signal data as input. For this purpose, a new loss metric function is designed to further measure the similarity loss between real signal data and synthetic signal data, and the generator is trained using this loss function. The loss function is defined as follows:

[0014] Loss(y p, t r , t s ) = αL BCE (y p ) + βL s (t r , t s ) + γL c (t r , t s )

[0015] Among them, y p represents the prediction result of the discriminator on the data, t r and t s respectively represent the real signal data and the corresponding synthesized signal data, L BCE is the original binary cross-entropy loss from the discriminator, L s and L c are respectively the distribution similarity loss and the classification similarity loss between the real data and the generated data, α, β, γ are the weights of these losses, and can be allocated differently for different scenarios. In the present invention, Softmax cross-entropy (SCE) is adopted as the loss function for L s and L c . The Softmax function: exp(x) represents the exponential function of ex, a k is the k-th input signal in the output layer, exp(a k ) represents the exponential function of a k . The denominator represents that there are n output signals (neurons) in the output layer, and calculates the sum of the exponentials of all input signals in the output layer. y k is the output of the k-th neuron. Since they are all regarded as multi-classification problems in this framework, SCE can be used for optimization. During the model training process, Loss(y p , t s , t s ) will update the weights of the generator to improve the quality of the synthesized signal data.

[0016] 3. Training the generative adversarial network model

[0017] Take the signal generated by the transmitter as the training sample input of the generative adversarial network model. According to the real signal data, a new signal is generated through the generator model in the network. Take the generated new signal data as the input of the discriminator, and the discriminator makes a judgment. After the network model iterates multiple times, finally the similarity loss reaches an ideal state, and the generative adversarial network can generate data highly similar to the real signal data.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The method of the present invention does not require complex processing of the original signal, can adaptively extract the features of different signals, and generate signals of the same type; After testing, the generated signals can achieve a reception rate approximate to that of the original signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow block diagram for implementing the present invention.

[0020] Figure 2 It is a generator network model.

[0021] Figure 3 It is a discriminator network model.

[0022] Figure 4 It is the data of the rxSig part.

[0023] Figure 5 It is the data of the data1 part.

[0024] Figure 6 It is the error rate errorStates. DETAILED DESCRIPTION OF THE INVENTION

[0025] Next, with reference to the drawings, taking the orthogonal frequency division multiplexing (OFDM) signal as an example, the technical solution of the present invention will be specifically described.

[0026] 1. Construct an original signal data set:

[0027] Generate OFDM signals through an OFDM signal generator, where the bit signal-to-noise ratio EbNoVec = -10, the symbol signal-to-noise ratio snrVec = EbNoVec + 10 * log10(k) + 10 * log10(numDC / numSC), the signal data after channel coding is data_encoded, and data_encoded is modulated by QPSK to obtain the modulated signal data rxSig. There are a total of 53664 data, and some of the data are as Figure 4 shown, which are used as the main training and test data for the generator and discriminator in the generative adversarial network.

[0028] 2. Construct a generative adversarial network model

[0029] Create a generator model (such as Figure 2As shown in the figure: The generator consists of five functional layers: an input layer, an embedding layer, a feature fusion layer, an LSTM layer, and a regression layer. First, the generator takes the encoded real OFDM signal data and random noise as inputs. Each element in the noise data is a number randomly generated from a normal distribution with a mean of 0, a standard deviation of 1, and a shape of (1027, 100). This value is obtained through multiple experiments and achieves a balance between adversarial performance and generation time. Then, multi-layer perceptron embedding is used. For the real and imaginary parts of the signal, each pair is embedded using a multi-layer perceptron (MLP) to obtain 64-dimensional vectors. After the embedding process, all vectors and random noise are further concatenated, and then they are fused into a 100-dimensional vector using a dense layer. Finally, using the distribution and type characteristics of each data, through feature fusion, it supports the modeling of the network and the generation of signals. In the LSTM modeling layer, a many-to-many LSTM structure is used. A sequence with a specific time step is used as the input, and a sequence with the same time step is generated as the output. Given the size of the fused features, 100 units are allocated in the LSTM model, and the fused features are fed into the model.

[0030] Create a discriminator model (as Figure 3 shown): The discriminator consists of an input layer, an embedding layer, a feature fusion layer, and an LSTM layer. Finally, a single-unit dense layer with a sigmoid function is used for binary classification (real data or synthetic data) of the scalar output.

[0031] Define the loss function: Loss(y p , t r , t s ) = αL BCE (y p ) + βL s (t r , t s ) + γL c (t r , t s ), where α, β, and γ are the weights of the loss, set to 2, 1, and 1 respectively. The loss functions of L s and L c adopt Softmax cross-entropy (SCE).

[0032] 3. Train the network model

[0033] Convert the real data signal rxSig.mat file into a csv file, split its real and imaginary parts into two columns of data for storage. In addition, to further expand the output range to ensure that it covers all possible deviation values, calculate the average values of the real and imaginary parts of the real signal data and the deviations between the real data and the average values, and save them for training together with the real data. The training is carried out for a total of 2000 epochs. After the training is completed, the generator can generate synthetic data, named data1. Some of the data is as Figure 5 shown.

[0034] 4. Test the reception effect

[0035] Put the synthetic data data1 generated by the generative adversarial network into the OFDM signal receiver for testing. After data1 undergoes inverse resource mapping, inverse constellation mapping, and inverse interleaving coding, the data_deinterleav data is obtained. Some of the data is shown in the figure. Compare data_deinterleav with the signal data_encoded that has undergone channel coding in the original transmitter to obtain the error rate errorStats, as Figure 6 shown. From the error rate errorStats, it can be seen that there are a total of 79872 pieces of data, and 26883 of them are incorrect. The error rate is 0.3366, indicating that the recognition rate of the synthetic data data1 generated by the generative adversarial network in the receiver is 66.34%, which has good performance.

[0036] The above are the preferred embodiments of the present invention. All changes made according to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention in terms of the functions and effects produced belong to the protection scope of the present invention.

Claims

1. A radio signal generation method based on a generative adversarial network, characterized in that, The signal transmitter generates an original signal, adds an interference signal to the original signal, and then puts it into the generative adversarial network model for learning. Through the training of the generative adversarial network model, a synthetic data signal similar to the original signal is generated. The method specifically includes: generating an OFDM signal; constructing a generative adversarial network model; training the generative adversarial network model; The generative adversarial network model consists of a generator and a discriminator, and the specific construction method is as follows: (1) The generator consists of five functional layers: an input layer, an embedding layer, a feature fusion layer, an LSTM layer, and a regression layer. The generator first takes the encoded real OFDM signal and random noise as inputs. For the real and imaginary parts of the signal, each pair of them is embedded using a multi-layer perceptron (MLP) to obtain 64-dimensional vectors. After the embedding process, all the vectors and random noise are concatenated, and then they are fused into a 100-dimensional vector using a dense layer. Then, using the distribution and type features of each data, through feature fusion, it supports the modeling of the network and the generation of signals. In the LSTM layer, a many-to-many LSTM structure is used, taking a sequence with a predetermined time step as input and generating a sequence with the same time step as output. Given the size of the fused features, 100 units are allocated in the LSTM layer, and the fused features are fed into the LSTM layer. Finally, the synthetic signal data H output from the LSTM layer is decoded; (2) The discriminator consists of an input layer, an embedding layer, a feature fusion layer, and an LSTM layer. The difference from the generator is that the input layer of the discriminator only takes signal data as input without adding random noise. The discriminator uses a many-to-one LSTM structure, taking features with time steps as input and outputting a scalar. A single-unit dense layer with a sigmoid function is used to perform binary classification on the scalar output, that is, classifying real data or synthetic data; Each eigenvector h in the synthetic signal data H i is a 100-dimensional vector containing the distribution characteristics of the synthetic signal data, using a Dense layer with two units and the hyperbolic tangent function tanh to decode the deviation of the real and imaginary parts of the signal data; in addition, using the Dense layer and the Softmax function to recover the data; The training method of the generative adversarial network model is specifically as follows: (1) Random noise is added during the training process. Each element in the random noise data is a number randomly generated from a normal distribution with a mean of 0, a standard deviation of 1, and a shape of (1027, 100); (2) The loss function is designed as Loss(y p , t r , t s ) = αL BCE (y p ) + βL s (t r , t s ) + γL c (t r , t s ), where y p represents the discriminator's prediction result for the data, t r and t s represent the real signal data and the corresponding synthesized signal data respectively, L BCE is the original binary cross-entropy loss from the discriminator, L s and L c are the distribution similarity loss and the classification similarity loss between the real data and the generated data respectively, α, β, γ are the weights of these losses, and different allocations are made for different scenarios; L s and L c has a loss function of Softmax Cross Entropy (SCE).

2. The radio signal generation method based on a generative adversarial network according to claim 1, wherein Generation of the OFDM signal: An OFDM signal is generated by an OFDM signal transmitter, where the bit signal-to-noise ratio EbNoVec = -10, the modulation order M = 4, the number of bits transmitted per symbol k = log2(M), the total number of subcarriers numSC = 2048, the symbol signal-to-noise ratio snrVec = EbNoVec + 10*log10(k) + 10*log10(numDC / numSC), numDC is the number of data subcarriers in OFDM, and the OFDM signal is expressed as x = [x1,...x i ,...x n , x i = (a i , b i ) is the i-th data of x, a i and b i are the real and imaginary parts of x i respectively.

Images