A radio data augmentation method and system for signal modulation identification

Abstract

The invention relates to the field of radio communication technology, and specifically discloses a radio data augmentation method and system for signal modulation identification, comprising: obtaining a data set to be augmented; the data set to be augmented comprises real data and modulation category labels; building a radio data augmentation model; the radio data augmentation model comprises a generator and a discriminator; training the radio data augmentation model by using an adaptive moment estimation optimization algorithm to obtain a trained target generator; inputting the data category labels and quantities to be generated into the target generator to obtain target generated data; and merging the target generated data and real data under each modulation category label to obtain an augmented data set for signal modulation identification.

Description

Technical Field

[0001] The present invention relates to the field of radio communication technology, and in particular to a radio data augmentation method and system for signal modulation identification. Background Art

[0002] At present, modulation recognition methods based on deep learning have achieved remarkable results in improving the classification performance of radio modulation data. However, such methods face two major challenges in practical applications: first, the performance of deep learning models is highly dependent on a large amount of high-quality radio modulation sample data; second, the quality of radio modulation data directly affects the performance of modulation classification models. In actual application scenarios, obtaining a large amount of reliable radio modulation sample data often faces problems such as high cost and operational difficulties, which seriously restricts the application and development of deep learning models in the field of radio modulation recognition.

[0003] To solve the above problems, scholars have gradually shifted their research focus to radio data augmentation technology. At present, there are two main technical routes in this field: the first is to transfer image or text data enhancement methods to radio modulation data augmentation, including operations such as rotating signal data at a specific angle, slicing and splicing data, and adding Gaussian noise; the second is to use deep learning models, such as autoencoders, generative adversarial networks and their derivative networks, to generate constellation diagrams or directly generate signal data by learning the distribution of original data.

[0004] However, there are still obvious deficiencies in existing research: on the one hand, there is a relative lack of data augmentation methods specifically for radio signal modulation recognition; on the other hand, due to the differences in the modulation categories involved in radio signal modulation recognition in different application scenarios, existing methods fail to fully consider the distribution characteristics of radio modulation data itself and the distribution differences between data of different modulation categories. Solving these problems has important theoretical and practical significance for improving the performance and applicability of radio modulation recognition systems. Summary of the invention

[0005] In view of the above problems, the object of the present invention is to provide a radio data augmentation method and system for signal modulation identification, which can enhance and expand the data when the data is insufficient to improve the accuracy of signal modulation identification.

[0006] The present invention provides a radio data augmentation method for signal modulation identification, comprising: Acquire a data set to be augmented; the data set to be augmented includes real data and modulation category labels; Constructing a radio data augmentation model; the radio data augmentation model includes a generator and a discriminator; The radio data augmentation model is trained using an adaptive moment estimation optimization algorithm to obtain a trained target generator; Inputting the data category label and quantity to be generated into the target generator to obtain target generation data; The target generated data and real data under each modulation category label are merged to obtain the augmented data set for signal modulation recognition.

[0007] In a possible implementation, obtaining the data set to be augmented includes: Obtain the files to be augmented for each modulation category; Generate a corresponding modulation category label according to the index of the file to be augmented; A data set to be augmented is constructed according to the file to be augmented and the modulation category label.

[0008] In a possible implementation manner, constructing the radio data augmentation model includes: Using a linear layer, four concatenated module and two parallel A module constructs said generator; Among them, the The module includes a conditional batch normalization layer CBN, a nonlinear activation layer LeakyReLU and a linear layer Linear; the conditional batch normalization layer CBN includes batch normalization and linear transformation Linear; The discriminator is constructed using five linear layers and four non-linear activation layers.

[0009] In a possible implementation manner, the training the radio data augmentation model by using an adaptive moment estimation optimization algorithm includes: The modulation category label is input into the embedding layer for embedding operation to obtain the category vector, and the formula is as follows: ; in, is the embedding operation of the embedding layer, is the modulation category label, is the category vector.

[0010] In a possible implementation manner, the training the radio data augmentation model by using an adaptive moment estimation optimization algorithm further includes: Generate random noise, and input the random noise and the category vector into a generator to obtain generated data; Calculate the total loss of the generator according to the spectral weight matrix and the spectral error, and update the parameters of the generator according to the total loss of the generator; Inputting the real data and the category vector into the discriminator respectively to obtain a first discrimination result and a second discrimination result; A total loss of a discriminator is calculated according to the first discrimination result and the second discrimination result, and a parameter of the discriminator is updated according to the total loss of the discriminator.

[0011] In a possible implementation manner, inputting the random noise and the category vector into a generator to obtain generated data includes: The random noise and the category vector are concatenated to obtain a conditional vector, and the formula is as follows: ; in, Represents a splicing operation, is random noise, is the category vector, is the conditional vector; The random noise is input into the linear layer to obtain a noise vector with a dimension of 256. The formula is as follows: ; in, It is a linear layer The weight matrix of It is a linear layer The bias vector of is the noise vector; The noise vector and the conditional vector are sequentially passed through four series The module performs feature extraction to obtain the target features. The formula is as follows: ; Input the target features into The module obtains I-channel data of generated data and inputs the target feature into The module obtains Q-channel data of the generated data, and combines the I-channel data of the generated data and the Q-channel data of the generated data to obtain the generated data. The formula is as follows: ; in, To generate data, To generate data for I channel, To generate the Q-path data, is the target feature, For splicing operation.

[0012] In a possible implementation, calculating the total generator loss according to the spectrum weight matrix and the spectrum error includes: Calculate frequency domain loss according to the spectrum weight matrix and the spectrum error , the formula is as follows: ; in, is the spectral weight matrix, is the spectral error, is a mean operation, is an element-wise multiplication operation: The adversarial loss is calculated according to the following formula : ; in, is a mean operation, is the discriminator for generating data Output: The total loss of the generator is calculated according to the adversarial loss and the frequency domain loss, and the formula is as follows: ; in, is the total loss of the generator, To combat losses, is the frequency domain loss.

[0013] In a possible implementation, the frequency domain loss is calculated according to the spectrum weight matrix and the spectrum error. Also includes: The generated data and real data Perform one-dimensional discrete Fourier transform respectively to obtain the spectrum of the generated data and the spectrum of real data ; The spectrum selection weight matrix is ​​calculated according to the following formula ; ; in, represents the Euclidean norm, is the scaling factor, is the spectrum matrix, To obtain the minimum value operation, To obtain the maximum value operation; The spectrum error is calculated according to the following formula : ; in, is the square of the Euclidean norm.

[0014] In a possible implementation manner, calculating the total loss of the discriminator according to the first discrimination result and the second discrimination result includes: The gradient penalty is calculated according to the following formula: ; in, is the weight coefficient of the gradient penalty, is the interpolation data between the real data and the generated data, for right The partial derivative of is the square of the Euclidean norm, is the modulation category label; The total loss of the discriminator is calculated according to the following formula : ; in, is the mean operation, is the gradient penalty, is the first judgment result, is the second discrimination result.

[0015] The present invention also provides a radio data augmentation system for signal modulation identification, which executes any of the radio data augmentation methods, including: An acquisition module is used to acquire a data set to be augmented; the data set to be augmented includes real data and a modulation category label; A radio data augmentation model construction module, used to construct a radio data augmentation model; the radio data augmentation model includes a generator and a discriminator; A training module, used for training the radio data augmentation model using an adaptive moment estimation optimization algorithm to obtain a trained target generator; A data generation module, used for inputting the data category label and quantity to be generated into the target generator to obtain target generation data; The merging module is used to merge the target generated data and the real data under each modulation category label to obtain an augmented data set for signal modulation identification.

[0016] The radio data augmentation method and system for signal modulation recognition provided by the present invention generate "fake" data with the same distribution characteristics as the original data based on a generative adversarial network, so as to achieve the purpose of augmenting the original data, that is, to enhance and expand the data when the data is insufficient to improve the accuracy of signal modulation recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a flow chart of a radio data augmentation method provided by an embodiment of the present invention; Figure 2 A network structure diagram of a radio data augmentation model provided by an embodiment of the present invention; Figure 3The embodiments of the present invention provide The structure diagram of the module; Figure 4 A structural diagram of a conditional batch normalization layer CBN provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a radio data augmentation system provided by an embodiment of the present invention; Figure 6 A comparison chart of modulation recognition accuracy of original data provided by an embodiment of the present invention and augmented data using different augmentation methods under different signal-to-noise ratios. DETAILED DESCRIPTION

[0018] The following detailed description of the embodiments of the present invention is further described in detail in conjunction with the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to exemplarily illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention, that is, the present invention is not limited to the preferred embodiments described, and the scope of the present invention is defined by the claims.

[0019] In the description of the present invention, it should be noted that, unless otherwise specified, “plurality” means two or more than two; the terms “first”, “second”, etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance; for ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0020] Figure 1 A flow chart of a radio data augmentation method provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a radio data augmentation method for signal modulation identification, comprising: Step S110, obtaining a data set to be augmented; The data set to be augmented includes real data and modulation category labels; In a possible implementation, a to-be-augmented file of each modulation category is obtained; a corresponding modulation category label is generated according to an index of the to-be-augmented file; and a to-be-augmented data set is constructed according to the to-be-augmented file and the modulation category label.

[0021] Step S120, constructing a radio data augmentation model IQ-SpectralGAN; the radio data augmentation model includes a generator and a discriminator; In one possible implementation, a linear layer, four cascaded module and two parallel Module build generator; where, The module includes a conditional batch normalization layer CBN, a nonlinear activation layer LeakyReLU and a linear layer Linear. Its structure is shown in the following figure. Figure 3As shown, the conditional batch normalization layer CBN includes batch normalization and linear transformation Linear, and its structure diagram is as follows Figure 4 As shown; the discriminator is constructed using five linear layers and four non-linear activation layers.

[0022] Step S130, training the radio data augmentation model using an adaptive moment estimation optimization algorithm to obtain a trained target generator; Wherein, N is a positive integer; In a possible implementation, the modulation category label is input into the embedding layer for embedding operation to obtain The category vector of dimension is as follows: ; in, is the embedding operation of the embedding layer, which can be expressed as , mapping discrete category labels into low-dimensional dense vectors. is the modulation category label, is the category vector, modulating the category label The number of categories is .

[0023] Generate random noise, and input the random noise and the category vector into the generator to obtain generated data; Calculate the total loss of the generator according to the spectral weight matrix and the spectral error, and update the parameters of the generator according to the total loss of the generator; The real data and the category vector are input into the discriminator respectively to obtain the first discrimination result and the second discrimination result; The total loss of the discriminator is calculated according to the first discrimination result and the second discrimination result, and the parameters of the discriminator are updated according to the total loss of the discriminator.

[0024] Specifically, the real data and the category vector Input the first five linear layers and four nonlinear activation layers to get the preliminary discrimination results ; Then the preliminary judgment results and the category vector Multiply each element and sum them to get the judgment result . Identification results The same logic applies.

[0025] In one possible implementation, random noise and a category vector are input into a generator to generate data include: The random noise and the category vector are concatenated to obtain the conditional vector. The formula is as follows: ; in, Represents a splicing operation, is random noise, is the category vector, is the conditional vector; Input random noise into the linear layer to obtain a noise vector with a dimension of 256. The formula is as follows: ; in, It is a linear layer The weight matrix of It is a linear layer The bias vector of is the noise vector; The noise vector and the conditional vector are sequentially passed through four series The module performs feature extraction to obtain the target features. The formula is as follows: ; Among them, each Each module receives the features and conditional vectors output by the previous layer. , so that the output characteristics of the module are related to the given conditions.

[0026] Enter the target feature The module obtains the I-channel data of the generated data and inputs the target feature into The module obtains Q-channel data of generated data, and combines I-channel data of generated data and Q-channel data of generated data to obtain generated data. The formula is as follows: ; in, To generate data, To generate data for I channel, To generate the Q-path data, is the target feature, For splicing operation.

[0027] In one possible implementation, calculating the total generator loss based on the spectrum weight matrix and the spectrum error includes: Calculate frequency domain loss based on spectral weight matrix and spectral error , the formula is as follows: ; in, is the spectral weight matrix, is the spectral error, is a mean operation, is an element-wise multiplication operation: The adversarial loss is calculated according to the following formula : ; in, is a mean operation, is the discriminator for generating data Output: The total loss of the generator is calculated based on the adversarial loss and frequency domain loss. The formula is as follows: ; in, is the total loss of the generator, To combat losses, is the frequency domain loss.

[0028] In one possible implementation, the frequency domain loss is calculated based on the spectrum weight matrix and the spectrum error. Also includes: The generated data and real data Perform one-dimensional discrete Fourier transform respectively to obtain the spectrum of the generated data and the spectrum of real data ; The spectrum selection weight matrix is ​​calculated according to the following formula ; ; in, represents the Euclidean norm, is the scaling factor, is the spectrum matrix, To obtain the minimum value operation, To obtain the maximum value operation; The spectrum error is calculated according to the following formula : ; in, is the square of the Euclidean norm.

[0029] In a possible implementation, calculating the total loss of the discriminator according to the first discrimination result and the second discrimination result includes: The gradient penalty is calculated according to the following formula: ; in, is the weight coefficient of the gradient penalty, is the interpolation data between the real data and the generated data, for right The partial derivative of is the square of the Euclidean norm, is the modulation category label; The total loss of the discriminator is calculated according to the following formula : ; in, is the mean operation, is the gradient penalty, is the first judgment result, is the second discrimination result.

[0030] Step S140: input the data category label and quantity to be generated into the target generator to obtain the target generated data. ; Step S150: Merge the target generation data under each modulation category label With real data , and obtain the augmented data set for signal modulation identification.

[0031] Figure 5 A schematic diagram of the structure of a radio data augmentation system provided by an embodiment of the present invention is shown in FIG. Figure 5 As shown, the present invention also provides a radio data augmentation system for signal modulation identification, which executes any radio data augmentation method, including: Acquisition module 1, used to acquire the data set to be augmented; the data set to be augmented includes real data and modulation category labels; A radio data augmentation model construction module 2 is used to construct a radio data augmentation model; the radio data augmentation model includes a generator and a discriminator; A training module 3 is used to train the radio data augmentation model using an adaptive moment estimation optimization algorithm to obtain a trained target generator; The data generation module 4 is used to input the data category label and quantity to be generated into the target generator to obtain the target generation data; The merging module 5 is used to merge the target generated data and the real data under each modulation category label to obtain an augmented data set for signal modulation identification.

[0032] The following is a description with reference to specific embodiments.

[0033] The experiment was verified on the IQ signal dataset RML2016.10a generated by GNU Radio software platform simulation, as shown in Table 1. Table 1 is the detailed information of the dataset RML2016.10a. The constellation diagram of the dataset RML2016.10a is shown in Figure 2 The hardware and software configurations used in the experiment are shown in Table 2. The hyperparameter settings for the IQ-SpectralGAN model training of the radio data augmentation model in the experiment are shown in Table 3, and other hyperparameters are default values.

[0034] Table 1

[0035] Table 2

[0036] Table 3

[0037] The accuracy is used in the experiment to measure the performance of the generalization ability of the network model. The recognition accuracy can be calculated according to the following formula: .

[0038] In this experiment, the number of modulation categories of the RML2016.10a data set is 11, and there are 20,000 signal samples in each category. The training set, validation set and test set are divided into 6:2:2 in each modulation category. According to the method proposed in the present invention, 1,500 signal samples are generated for each of the 11 modulation categories of the RML2016.10a data set, and they are added to the original training set to obtain the augmented training set. The division of the data set before and after augmentation is shown in Table 4.

[0039] Table 4

[0040] In order to verify the effectiveness of the radio data augmentation method for signal modulation recognition proposed in the present invention, the modulation recognition network MCLDNN is trained using the training sets before and after augmentation, compared with the traditional rotation enhancement method. MCLDNN is a classic network for modulation recognition. Then the model parameters with the highest recognition accuracy on the validation set are taken as the optimal model parameters, and finally the test set is input into the modulation recognition network MCLDNN with the optimal model parameters to obtain the recognition accuracy under each modulation category. The results are shown in Table 5 below.

[0041] Figure 6 A comparison chart of modulation recognition accuracy of original data provided by an embodiment of the present invention and augmented data using different augmentation methods under different signal-to-noise ratios, Figure 6 The horizontal axis is the signal-to-noise ratio, and the vertical axis is the recognition accuracy of the data on the modulation recognition network MCLDNN. Table 5 shows the average recognition accuracy and the highest recognition accuracy of the original data and the augmented data with different augmentation methods. Figure 6 As can be seen from Table 5, the data augmented by the radio data augmentation model IQ-SpectralGAN model of the present invention has a better recognition accuracy than the original data and the rotated augmented data. This shows that the radio data augmentation method for signal modulation recognition proposed in the present invention can better generate "new" data for signal modulation recognition and improve the accuracy of signal modulation recognition.

[0042] Table 5

[0043] The radio data augmentation method and system for signal modulation identification provided by the present invention have the following technical effects: (1) The generator in the radio data augmentation model IQ-Spectral GAN ​​proposed in this invention incorporates category label information in each layer. This not only introduces category information at the input stage of the network, but also utilizes the category information at each level of the generator, so that the network can more fully generate corresponding I and Q signals based on category features, thereby better guiding the data generation process and improving the correlation and accuracy of the generated data with the desired category modulation signal.

[0044] (2) The present invention adds frequency domain loss during the training process, which enables the network to focus not only on spatial domain information but also on frequency domain information during training, which is conducive to capturing the complex frequency structure and key frequency information in the communication modulated IQ signal. Through the dynamic spectrum weighting mechanism, the model can focus on learning the frequency part that plays a key role in signal quality and accuracy, thereby guiding the generated data to match the real data in frequency domain characteristics, helping the network to learn high-frequency features, overcoming the limitation of traditional generative adversarial networks that only focus on spatial domain information, and enhancing the network's comprehensive characterization ability of modulated signals.

[0045] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A radio data augmentation method for signal modulation identification, characterized in that: include: Acquire a data set to be augmented; the data set to be augmented includes real data and modulation category labels; Constructing radio data augmentation models; The radio data augmentation model includes a generator and a discriminator; The radio data augmentation model is trained using an adaptive moment estimation optimization algorithm to obtain a trained target generator; Inputting the data category label and quantity to be generated into the target generator to obtain target generation data; The target generated data and real data under each modulation category label are merged to obtain the augmented data set for signal modulation recognition.

2. The radio data augmentation method according to claim 1, characterized in that: The obtaining of the data set to be augmented comprises: Obtain the files to be augmented for each modulation category; Generate a corresponding modulation category label according to the index of the file to be augmented; A data set to be augmented is constructed according to the file to be augmented and the modulation category label.

3. The radio data augmentation method according to claim 1, characterized in that: The constructing of the radio data augmentation model comprises: Using a linear layer, four concatenated Module and two parallel A module constructs said generator; Among them, the The module includes a conditional batch normalization layer CBN, a nonlinear activation layer LeakyReLU and a linear layer Linear; the conditional batch normalization layer CBN includes batch normalization and linear transformation Linear; The discriminator is constructed using five linear layers and four non-linear activation layers.

4. The radio data augmentation method according to claim 1, characterized in that: The training of the radio data augmentation model by using an adaptive moment estimation optimization algorithm comprises: The modulation category label is input into the embedding layer for embedding operation to obtain the category vector, and the formula is as follows: ； in, is the embedding operation of the embedding layer, is the modulation category label, is the category vector.

5. The radio data augmentation method according to claim 4, characterized in that: The using of the adaptive moment estimation optimization algorithm to train the radio data augmentation model further comprises: Generate random noise, and input the random noise and the category vector into a generator to obtain generated data; Calculate the total loss of the generator according to the spectral weight matrix and the spectral error, and update the parameters of the generator according to the total loss of the generator; Inputting the real data and the category vector into the discriminator respectively to obtain a first discrimination result and a second discrimination result; A total loss of a discriminator is calculated according to the first discrimination result and the second discrimination result, and a parameter of the discriminator is updated according to the total loss of the discriminator.

6. The radio data augmentation method according to claim 5, characterized in that: The step of inputting the random noise and the category vector into a generator to obtain generated data comprises: The random noise and the category vector are concatenated to obtain a conditional vector, and the formula is as follows: ； in, Represents a splicing operation, is random noise, is the category vector, is the conditional vector; The random noise is input into the linear layer to obtain a noise vector with a dimension of 256. The formula is as follows: ； in, It is a linear layer The weight matrix of It is a linear layer The bias vector of is the noise vector; The noise vector and the conditional vector are sequentially passed through four series The module performs feature extraction to obtain the target features. The formula is as follows: ； Input the target features into The module obtains I-channel data of generated data and inputs the target feature into The module obtains Q-channel data of the generated data, and combines the I-channel data of the generated data and the Q-channel data of the generated data to obtain the generated data. The formula is as follows: ； in, To generate data, To generate data for I channel, To generate the Q-path data, is the target feature, For splicing operation.

7. The radio data augmentation method according to claim 5, characterized in that: The calculation of the total generator loss according to the spectrum weight matrix and the spectrum error includes: Calculate frequency domain loss according to the spectrum weight matrix and the spectrum error , the formula is as follows: ； in, is the spectral weight matrix, is the spectral error, is a mean operation, is an element-wise multiplication operation: The adversarial loss is calculated according to the following formula : ； in, is a mean operation, is the discriminator for generating data Output: The total loss of the generator is calculated according to the adversarial loss and the frequency domain loss, and the formula is as follows: ； in, is the total loss of the generator, To combat losses, is the frequency domain loss.

8. The radio data augmentation method according to claim 7, characterized in that: The frequency domain loss is calculated according to the spectrum weight matrix and the spectrum error Also includes: The generated data and real data Perform one-dimensional discrete Fourier transform respectively to obtain the spectrum of the generated data and the spectrum of real data ; The spectrum selection weight matrix is ​​calculated according to the following formula ; ； in, represents the Euclidean norm, is the scaling factor, is the spectrum matrix, To obtain the minimum value operation, To obtain the maximum value operation; The spectrum error is calculated according to the following formula : ； in, is the square of the Euclidean norm.

9. The radio data augmentation method according to claim 5, characterized in that: The calculating the total loss of the discriminator according to the first discrimination result and the second discrimination result comprises: The gradient penalty is calculated according to the following formula: ； in, is the weight coefficient of the gradient penalty, is the interpolation data between the real data and the generated data, for right The partial derivative of is the square of the Euclidean norm, is the modulation category label; The total loss of the discriminator is calculated according to the following formula : ； in, is the mean operation, is the gradient penalty, is the first judgment result, is the second discrimination result.

10. A radio data augmentation system for signal modulation identification, executing the radio data augmentation method according to any one of claims 1 to 9, characterized in that: include: An acquisition module is used to acquire a data set to be augmented; the data set to be augmented includes real data and a modulation category label; A radio data augmentation model construction module, used to construct a radio data augmentation model; the radio data augmentation model includes a generator and a discriminator; A training module, used for training the radio data augmentation model using an adaptive moment estimation optimization algorithm to obtain a trained target generator; A data generation module, used for inputting the data category label and quantity to be generated into the target generator to obtain target generation data; The merging module is used to merge the target generated data and the real data under each modulation category label to obtain an augmented data set for signal modulation identification.

Images