Small sample radiation source data enhancement and individual identification method based on equipotential constellation diagram

Abstract

The present invention proposes a method for data enhancement and individual identification of small-sample radiation sources based on equipotential constellation diagrams, comprising the following steps: preprocessing the collected I / Q signals emitted by the radiation source, and mapping the one-dimensional I / Q signal sequence data onto a two-dimensional equipotential constellation diagram image. An improved generative adversarial network is used to perform data enhancement on the two-dimensional equipotential constellation diagram of the I / Q signal sample, and a structural similarity measurement method is used to screen the generated equipotential constellation diagram to optimize the quality of the generated samples. A residual neural network recognition model is constructed, and the generated sample data set is merged with the original data set for model training. Finally, the obtained radiation source recognition model identifies the radiation source device in the small-sample scenario. By comprehensively applying technologies such as mapping the I / Q signal sequence data to images, data enhancement and screening, the model provides an effective solution for individual identification of radiation sources in small-sample scenarios.

Description

Technical Field

[0001] The present invention belongs to the application field of radiation source individual identification technology, and in particular relates to a small sample radiation source data enhancement and individual identification method based on an equipotential constellation diagram. Background Art

[0002] Radiation source identification technology essentially uses specific parameters within RF signals, such as frequency, phase, and amplitude, to analyze and process the signals to extract RF fingerprint features associated with the signal source. These RF fingerprint features can be matched with data in a pre-established signal database to identify the device to which the signal belongs. Each radiation source's RF fingerprint is unique. This is because even radiation sources from the same manufacturer, batch, and model can have subtle differences in the components of their transmitter analog circuits, which contribute to the uniqueness of their fingerprint features. These RF fingerprint features can be used to identify and authenticate the source of the signal, effectively distinguishing the electronic device emitting the signal. Deep learning methods are currently widely used in the radiation source identification problem, achieving a series of breakthroughs. Compared with traditional methods, deep learning's end-to-end recognition architecture and powerful feature extraction capabilities make it a mainstream model for radiation source identification.

[0003] However, most deep learning-based methods for identifying individual radiation sources typically require large amounts of labeled data. In practice, however, labeled data for radiation sources is often difficult to obtain, with only a small number of labeled samples available. This makes it difficult for neural networks to accurately extract the characteristics of radiation sources, and underfitting is prone to occur during network training, resulting in unsatisfactory recognition results. Therefore, improving the accuracy of individual radiation source identification with small sample sizes has always been a challenge in combining deep learning with individual radiation source identification.

[0004] Therefore, the present application focuses on solving the problem of individual identification of radiation sources in small sample scenarios, and proposes a method for data enhancement and individual identification of small sample radiation sources based on equipotential constellation diagrams. In the proposed method, the I / Q signals emitted by the radiation source collected in the scene are preprocessed, and the I / Q signal data in the form of a one-dimensional sequence is converted into an equipotential constellation diagram in the form of a two-dimensional image. The equipotential constellation diagram is adversarially generated to achieve data enhancement, and the generated equipotential constellation diagram is screened using an image similarity calculation method, and finally input into a residual neural network for feature extraction and classification. The small sample radiation source data enhancement and individual identification scheme based on equipotential constellation diagrams proposed in this application further solves the problem of difficulty in obtaining sufficient labeled data for model training in small sample scenarios. Summary of the Invention

[0005] The purpose of the embodiment of the present application is to provide a small sample radiation source data enhancement and individual identification method based on an equipotential constellation diagram, so as to solve the problems of the existing radiation source individual identification model based on deep learning, such as underfitting of the model, poor feature extraction ability and unsatisfactory recognition effect when the amount of label data is small.

[0006] The present invention provides a small sample radiation source data enhancement and individual identification method based on an equipotential constellation diagram, comprising the following steps:

[0007] (1) Preprocess the I / Q signals emitted by the radiation source collected in the scene and convert the I / Q signal data in the form of a one-dimensional sequence into an equipotential constellation diagram in the form of a two-dimensional image:

[0008] Specifically, the representation of the I / Q signal emitted by the radiation source device is as follows:

[0009] Assume that the data transmitted by the radiation source is an I / Q modulated signal, which consists of an I carrier and a Q carrier, where the Q carrier is obtained by rotating the I carrier 90 degrees:

[0010]

[0011]

[0012] This maps the one-dimensional I / Q signal emitted by the radiation source onto a two-dimensional constellation diagram. In the constellation diagram, the I carrier and Q carrier of each sampling point correspond to the horizontal and vertical coordinates of the constellation diagram, respectively. The angle of a point relative to the origin represents the phase offset of its carrier waveform relative to the reference phase, while the distance from the point to the origin represents the signal amplitude. Due to channel noise and other interference factors, the amplitude and phase information of the signal can be affected, making it impossible for the sampling points to be concentrated at a single point in the constellation diagram. Therefore, different areas of the constellation diagram have different sampling point densities.

[0013] Based on the constellation diagram, a sliding window is used to traverse the constellation diagram in turn to calculate the number of data sample points in different areas of the constellation diagram. A rectangular window function of size W×H is used, where W represents the width of the rectangle and H represents the height of the rectangle. The point density ρ is used. (W,H) (i, j) indicates how many sample points there are in the sliding window.

[0014]

[0015] where dots (W,H) (i, j) is the number of points in the rectangular window, and N is the length of the I / Q waveform. The constellation diagram is colored according to the point density of different areas, where yellow represents higher density areas, green represents medium density areas, and blue represents lower density areas. The specific calculation is as follows:

[0016]

[0017] Where x1 and y1 represent the upper left corner of the current coordinate of the rectangular window function, x2 and y2 represent the lower right corner of the current coordinate of the rectangular window function, W0 and H0 represent the coordinates of the upper left corner of the constellation diagram, W1 and H1 represent the coordinates of the lower right corner of the constellation diagram, ρ(i, j) is the relative point density, dots(i, j) represents a sampling point, and its coordinates in the constellation diagram are (i, j).

[0018] (2) Based on the traditional generative adversarial network (GAN), a local discriminator is introduced, and the global discriminator loss and the local discriminator loss are combined to improve the quality of the generated equipotential constellation diagram, and data enhancement is performed on the I / Q signal samples represented in the form of two-dimensional images to achieve the expansion of small sample data sets.

[0019] The improved generative adversarial network (PatchGAN) is used to enhance the data of the equipotential constellation map to expand the small sample data set:

[0020] Specifically, this step may include the following sub-steps:

[0021] (2.1) GAN consists of a generator model and a discriminator model. The two are trained alternately in the network, learning and improving together in the confrontation, and finally reaching a Nash equilibrium.

[0022] Generator G learns the feature distribution P of the true equipotential constellation x , and maps out a generated equipotential constellation G(z) that is similar to the feature distribution of the real equipotential constellation. The discriminator D needs to learn the feature distribution P of the real equipotential constellation. x And the generated equipotential constellation feature distribution P G(z) To identify whether the input equipotential constellation is real or generated, that is, to perform binary classification on x and G(z). When the discriminator cannot distinguish the source of the equipotential constellation, the two reach a state of equilibrium and the network completes training.

[0023] GAN's adversarial game, mutual progress, and final Nash equilibrium are achieved by relying on the adversarial loss function, which is the objective function of the initial GAN.

[0024]

[0025] By leveraging the idea of ​​partial derivatives of multivariate functions, we can discuss the discriminator when the generator G is fixed: Here, D(x) represents the discriminator's confidence in the distribution of the true equipotential constellation features, which must be as large as possible; D(G(z)) represents the discriminator's confidence in the generated equipotential constellation, which must be as small as possible. This means that the second sub-formula as a whole must be as large as possible. In summary, when G is fixed, D needs to be maximized. Similarly, when D is fixed, G needs to be minimized. This means that minimizing the generator loss while maximizing the discriminator loss achieves a Nash equilibrium, and the network completes training.

[0026] (2.2) However, traditional GAN ​​has problems such as unstable model convergence and low quality of generated images, which will cause a large amount of data features to be lost in the generated equipotential constellation map, affecting the performance of subsequent feature extraction and classification. Therefore, an improved generative adversarial network, PatchGAN, is introduced, in which the local discriminator distinguishes whether each square n×n area with a side length of n in the image is true or false. This discriminator is run convolutionally on the image, and all responses are averaged to provide a final output of D. Where N can be much smaller than the full size of the image and still produce high-quality results. Smaller PatchGANs have fewer parameters, run faster, and can be applied to arbitrarily large images. The output of the PatchGAN discriminator is a matrix, each value represents the true or false corresponding to a block in the original image, and the average is taken as the result after equalization. The loss function of PatchGAN can be expressed as:

[0027]

[0028] Among them L adv (D ij (x),1) represents the probability value of the n×n local area in the i-th row and j-th column of the input equipotential constellation being correctly judged as the true equipotential constellation, L adv (D ij (G(z)), 0) represents the probability value that the n×n local area in the i-th row and j-th column of the generated image is correctly judged as the generated equipotential constellation diagram.

[0029] (2.3) The equipotential constellation is input into the local discriminator and the global discriminator at the same time to obtain their respective discrimination results. The two discriminators are updated only by their own losses, and the generator is updated by the sum of the losses of the two discriminators. This takes into account both the overall outline and local details of the generated equipotential constellation.

[0030] L total =L GAN +L PatchGAN

[0031] (3) Using the structural similarity (SSIM) metric, the equipotential constellation generated in step (2) is screened to optimize the quality of the enhanced sample set:

[0032] Specifically, in order to further improve the quality of the equipotential constellation maps input into the neural network, an image similarity calculation method, structural similarity (SSIM), is adopted to screen images with more similar feature distributions to the original equipotential constellation maps from the generated equipotential constellation maps.

[0033] The basic idea of ​​the SSIM algorithm is to divide the equipotential constellation into small blocks, calculate the brightness, contrast, and structure information of each block, and then combine this information to calculate the similarity between equipotential constellations. When calculating brightness, contrast, and structure information, SSIM uses a Gaussian weighting function to weight the pixels surrounding each pixel. This makes the algorithm robust to noise and distortion in the equipotential constellations.

[0034] SSIM can be expressed as follows:

[0035]

[0036] Where x and y represent two equipotential constellations, μ represents the mean grayscale value, σ represents the variance of the grayscale value, and c is a nonzero constant to prevent the denominator from being zero. The output of the SSIM algorithm is a value between 0 and 1, indicating the degree of similarity between the two equipotential constellations, where 1 indicates identical equipotential constellations and 0 indicates completely different equipotential constellations.

[0037] (4) A residual neural network is used to construct a radiation source recognition model, and the generated sample data set is merged with the original data set for model training. Feature extraction and classification are performed on the equipotential constellation data set after data enhancement and screening:

[0038] The residual network includes:

[0039] (4-1) The network is composed of basic blocks, which are composed of convolutional layers, batch normalization layers and activation functions;

[0040] (4-2) Residual connections are introduced between the basic blocks described in step (4-1) so that the input information of the previous layer is directly added to the output of the subsequent layer, building a channel for information to flow directly in the network.

[0041] (4-3) The residual network uses a series of convolutional layers and residual blocks to extract image features and maps these features to category probabilities through fully connected layers and softmax activation functions.

[0042] Specifically, input image: the image is input into the residual network model for classification. The equipotential constellation map can be represented as a three-dimensional tensor with width, height and number of channels (such as RGB image has 3 channels).

[0043] Basic Block: The core of the residual network model is the basic block, which is a residual unit composed of multiple convolutional layers. Each basic block includes the following steps: Convolutional Layer: Performs a convolution operation on the input feature map to convert it into a higher-level feature representation. Batch Normalization: Normalizes the results of the convolution operation to accelerate training. Activation Function: Apply a nonlinear activation function (such as ReLU) to activate each element in the feature map.

[0044] Residual Connection: Residual connections are introduced between basic blocks to transfer information across layers. Residual connections construct a skip connection by adding the input of the previous layer directly to the output of the subsequent layer, allowing information to flow directly in the network and helping to alleviate the gradient vanishing problem.

[0045] Residual Block: A residual block composed of multiple basic blocks is a repeating module of the residual network model. These blocks build a deeper network structure by stacking multiple basic blocks to better capture complex features in the image.

[0046] Global average pooling layer: At the end of the residual block, a global average pooling layer is used to reduce the dimensionality of the feature map. The global average pooling operation reduces the spatial dimension of each feature map to 1x1, retaining the average value of each channel, thereby reducing the number of parameters.

[0047] Fully connected layer: After the global average pooling layer, a fully connected layer is added to map features to category labels. The fully connected layer consists of one or more hidden layers with a large number of neurons, and is finally connected to a softmax layer with the number of output neurons of the same category.

[0048] Softmax activation function: The final layer uses a softmax activation function to convert the output into a probability distribution, representing the predicted probability of each category.

[0049] The residual network uses a series of convolutional layers and residual blocks to extract image features and maps these features to class probabilities through fully connected layers and a softmax activation function. Residual connections help address the vanishing gradient problem as network depth increases, enabling the training of deeper network structures and achieving better classification performance for equipotential constellations. Finally, the neural network model trained on the data augmented dataset is obtained.

[0050] (5) Use the small sample radiation source recognition model trained with the above enhanced dataset to identify the radiation source of the scene.

[0051] The technical solutions provided by the embodiments of the present application may have the following beneficial effects:

[0052] As can be seen from the above examples, the method proposed in this application solves the problems of underfitting and poor recognition accuracy that deep learning models often encounter when using small sample sizes. This method uses an equipotential constellation to convert one-dimensional signals into two-dimensional images, and then applies an improved GAN for data augmentation. This solves the problem of poor fitting when traditional GANs directly generate signal data.

[0053] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0055] Figure 1 The present invention is a flowchart of a method for enhancing small sample radiation source data and identifying individuals based on an equipotential constellation diagram according to an exemplary embodiment. DETAILED DESCRIPTION

[0056] Exemplary embodiments are described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different drawings represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with this application.

[0057] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. As used in this application and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0058] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0059] Figure 1 is a flowchart of a lightweight radiation source individual identification method based on adaptive width learning according to an exemplary embodiment. Figure 1 As shown, the method is applied in a terminal and may include the following steps:

[0060] (1) Preprocessing the I / Q signals emitted by the radiation source collected in the scene, converting the I / Q signal data in the form of a one-dimensional sequence into an equipotential constellation diagram in the form of a two-dimensional image;

[0061] (2) Based on the traditional generative adversarial network, a local discriminator is introduced, and the global discriminator loss and the local discriminator loss are combined to improve the quality of the generated equipotential constellation map, and the I / Q signal samples represented in the form of two-dimensional images are enhanced to achieve the expansion of small sample data sets;

[0062] (3) Using the structural similarity measurement method, the equipotential constellation graph generated in step (2) is screened to optimize the quality of the generated sample set;

[0063] (4) A residual neural network is used to construct a radiation source recognition model, and the generated sample dataset is merged with the original dataset for model training;

[0064] (5) Using the radiation source recognition model trained in step (4), the radiation source devices in the small sample scene are identified.

[0065] It can be seen from the above embodiments that the present application proposes a data enhancement scheme based on equipotential constellation diagrams and GANs, which solves the problem of underfitting of the model and poor recognition effect in deep learning scenarios with small amounts of labeled data, and further optimizes the problems of poor detail performance and unstable model convergence of traditional GANs when generating images; the proposed small-sample radiation source individual recognition model can achieve relatively ideal model recognition accuracy when there is only a small amount of labeled data.

[0066] This paper converts the collected I / Q signals into equipotential constellations, which can be represented in an image data format while conveying deep statistical information from the original wireless signal waveform. This allows the use of high-performance image classification algorithms in computer vision to classify equipotential constellations, thereby achieving the classification of radiation sources. In traditional GANs, the generated equipotential constellations exhibit poor detail, which impacts the subsequent neural network's ability to extract equipotential constellation features. Therefore, improvements are made to the discriminator, introducing PatchGAN, which improves the detail of the generated equipotential constellations. By further screening the equipotential constellations generated by PatchGAN and selecting images with more similar feature distributions to the original equipotential constellations, the categorical characteristics of the generated images can be more prominent, which is beneficial for improving the performance of subsequent neural network training.

[0067] In the specific implementation of step (1), the I / Q signals emitted by the radiation source collected in the scene are preprocessed, and the I / Q signal data in the form of a one-dimensional sequence is converted into an equipotential constellation diagram in the form of a two-dimensional image;

[0068] Specifically, the representation of the I / Q signal emitted by the radiation source device is as follows:

[0069] Assume that the data transmitted by the radiation source is an I / Q modulated signal, which consists of an I carrier and a Q carrier, where the Q carrier is obtained by rotating the I carrier 90 degrees:

[0070]

[0071]

[0072] This maps the one-dimensional I / Q signal emitted by the radiation source onto a two-dimensional constellation diagram. In the constellation diagram, the I carrier and Q carrier of each sampling point correspond to the horizontal and vertical coordinates of the constellation diagram, respectively. The angle of a point relative to the origin represents the phase offset of its carrier waveform relative to the reference phase, while the distance from the point to the origin represents the signal amplitude. Due to channel noise and other interference factors, the amplitude and phase information of the signal can be affected, making it impossible for the sampling points to be concentrated at a single point in the constellation diagram. Therefore, different areas of the constellation diagram have different sampling point densities.

[0073] Based on the constellation diagram, a sliding window is used to traverse the constellation diagram in turn to calculate the number of data sample points in different areas of the constellation diagram. A rectangular window function of size W×H is used, where W represents the width of the rectangle and H represents the height of the rectangle. The point density ρ is used. (W,H) (i, j) indicates how many sample points there are in the sliding window.

[0074]

[0075] where dots (W,H) (i, j) is the number of points in the rectangular window, and N is the length of the I / Q waveform. The constellation diagram is colored according to the point density of different areas, where yellow represents higher density areas, green represents medium density areas, and blue represents lower density areas. The specific calculation is as follows:

[0076]

[0077] Where x1 and y1 represent the upper left corner of the current coordinate of the rectangular window function, x2 and y2 represent the lower right corner of the current coordinate of the rectangular window function, W0 and H0 represent the coordinates of the upper left corner of the constellation diagram, W1 and H1 represent the coordinates of the lower right corner of the constellation diagram, ρ(i, j) is the relative point density, dots(i, j) represents a sampling point, and its coordinates in the constellation diagram are (i, j).

[0078] In the specific implementation of step (2), the improved generative adversarial network (PatchGAN) is used to perform data enhancement on the equipotential constellation map to expand the small sample data set:

[0079] Specifically, this step may include the following sub-steps:

[0080] (2.1) GAN consists of a generator model and a discriminator model. The two are trained alternately in the network, learning and improving together in the confrontation, and finally reaching a Nash equilibrium.

[0081] Generator G learns the feature distribution P of the true equipotential constellation x , and maps out a generated equipotential constellation G(z) that is similar to the feature distribution of the real equipotential constellation. The discriminator D needs to learn the feature distribution P of the real equipotential constellation. x And the generated equipotential constellation feature distribution P G(z) To identify whether the input equipotential constellation is real or generated, that is, to perform binary classification on x and G(z). When the discriminator cannot distinguish the source of the equipotential constellation, the two reach a state of equilibrium and the network completes training.

[0082] GAN's adversarial game, mutual progress, and final Nash equilibrium are achieved by relying on the adversarial loss function, which is the objective function of the initial GAN.

[0083]

[0084] By leveraging the idea of ​​partial derivatives of multivariate functions, we can discuss the discriminator when the generator G is fixed: Here, D(x) represents the discriminator's confidence in the distribution of the true equipotential constellation features, which must be as large as possible; D(G(z)) represents the discriminator's confidence in the generated equipotential constellation, which must be as small as possible. This means that the second sub-formula as a whole must be as large as possible. In summary, when G is fixed, D needs to be maximized. Similarly, when D is fixed, G needs to be minimized. This means that minimizing the generator loss while maximizing the discriminator loss achieves a Nash equilibrium, and the network completes training.

[0085] (2.2) However, traditional GANs have problems such as unstable model convergence and low quality of generated images, which will cause a large amount of data features to be lost in the generated equipotential constellation map, affecting the performance of subsequent feature extraction and classification. Therefore, an improved generative adversarial network, PatchGAN, is introduced, in which a local discriminator distinguishes whether each square n×n area with a side length of n in the image is true or false. This discriminator is run convolutionally on the image, and all responses are averaged to provide a final output of D. Where N can be much smaller than the full size of the image and still produce high-quality results. Smaller PatchGANs have fewer parameters, run faster, and can be applied to arbitrarily large images. The output of the PatchGAN discriminator is a matrix, where each value represents the true or false corresponding to a block in the original image, and the average is taken as the result after equalization. The loss function of PatchGAN can be expressed as:

[0086]

[0087] Among them L adv (D ij (x),1) represents the probability value of the n×n local area in the i-th row and j-th column of the input equipotential constellation being correctly judged as the true equipotential constellation, L adv (D ij (G(z)), 0) represents the probability value that the n×n local area in the i-th row and j-th column of the generated image is correctly judged as the generated equipotential constellation diagram.

[0088] (2.3) The equipotential constellation is input into the local discriminator and the global discriminator at the same time to obtain their respective discrimination results. The two discriminators are updated only by their own losses, and the generator is updated by the sum of the losses of the two discriminators. This takes into account both the overall outline and local details of the generated equipotential constellation.

[0089] L total =L GAN +L PatchGAN

[0090] In the specific implementation of step (3), the structural similarity (SSIM) measurement method is used to screen the equipotential constellation diagram generated in step (2) to optimize the quality of the enhanced sample set:

[0091] Specifically, in order to further improve the quality of the equipotential constellation maps input into the neural network, an image similarity calculation method, structural similarity (SSIM), is adopted to screen images with more similar feature distributions to the original equipotential constellation maps from the generated equipotential constellation maps.

[0092] The basic idea of ​​the SSIM algorithm is to divide the equipotential constellation into small blocks, calculate the brightness, contrast, and structure information of each block, and then combine this information to calculate the similarity between equipotential constellations. When calculating brightness, contrast, and structure information, SSIM uses a Gaussian weighting function to weight the pixels surrounding each pixel. This makes the algorithm robust to noise and distortion in the equipotential constellations.

[0093] SSIM can be expressed as follows:

[0094]

[0095] Where x and y represent two equipotential constellations, μ represents the mean grayscale value, σ represents the variance of the grayscale value, and c is a nonzero constant to prevent the denominator from being zero. The output of the SSIM algorithm is a value between 0 and 1, indicating the degree of similarity between the two equipotential constellations, where 1 indicates identical equipotential constellations and 0 indicates completely different equipotential constellations.

[0096] In the specific implementation of step (4), a residual network is used to extract features and classify the equipotential constellation dataset after data enhancement and screening.

[0097] Specifically, input image: the image is input into the residual network model for classification. The equipotential constellation map can be represented as a three-dimensional tensor with width, height and number of channels (such as RGB image has 3 channels).

[0098] Basic Block: The core of the residual network model is the basic block, which is a residual unit composed of multiple convolutional layers. Each basic block includes the following steps: Convolutional Layer: Performs a convolution operation on the input feature map to convert it into a higher-level feature representation. Batch Normalization: Normalizes the results of the convolution operation to accelerate training. Activation Function: Apply a nonlinear activation function (such as ReLU) to activate each element in the feature map.

[0099] Residual Connection: Residual connections are introduced between basic blocks to transfer information across layers. Residual connections construct a skip connection by adding the input of the previous layer directly to the output of the subsequent layer, allowing information to flow directly in the network and helping to alleviate the gradient vanishing problem.

[0100] Residual Block: A residual block composed of multiple basic blocks is a repeating module of the residual network model. These blocks build a deeper network structure by stacking multiple basic blocks to better capture complex features in the image.

[0101] Global average pooling layer: At the end of the residual block, a global average pooling layer is used to reduce the dimensionality of the feature map. The global average pooling operation reduces the spatial dimension of each feature map to 1x1, retaining the average value of each channel, thereby reducing the number of parameters.

[0102] Fully connected layer: After the global average pooling layer, a fully connected layer is added to map features to category labels. The fully connected layer consists of one or more hidden layers with a large number of neurons, and is finally connected to a softmax layer with the number of output neurons of the same category.

[0103] Softmax activation function: The final layer uses a softmax activation function to convert the output into a probability distribution, representing the predicted probability of each category.

[0104] The residual network uses a series of convolutional layers and residual blocks to extract image features and maps these features to class probabilities through fully connected layers and a softmax activation function. Residual connections help address the vanishing gradient problem as network depth increases, enabling the training of deeper network structures and achieving better classification performance for equipotential constellations. Finally, the neural network model trained on the data augmented dataset is obtained.

[0105] In the specific implementation of step (5), the radiation source of the scene is identified using the small sample radiation source identification model trained with the above-mentioned enhanced data set.

[0106] Those skilled in the art will readily conceive of other embodiments of the present application after considering the specification and practicing the contents disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of this application and include common knowledge or customary techniques in the art that are not disclosed in this application.

[0107] It will be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagrams, characterized in that: The steps include: (1) Preprocessing the I / Q signals emitted by the radiation source collected in the scene, converting the I / Q signal data in the form of a one-dimensional sequence into an equipotential constellation diagram in the form of a two-dimensional image; (2) Based on the traditional generative adversarial network, a local discriminator is introduced, and the global discriminator loss and the local discriminator loss are combined to improve the quality of the generated equipotential constellation map, and the I / Q signal samples represented in the form of two-dimensional images are enhanced to achieve the expansion of small sample data sets; (3) Using the structural similarity measurement method, the equipotential constellation graph generated in step (2) is screened to optimize the quality of the generated sample set; (4) A residual neural network is used to construct a radiation source recognition model, and the generated sample dataset is merged with the original dataset for model training; (5) Using the radiation source recognition model trained in step (4), the radiation source devices in the small sample scene are identified.

2. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagram according to claim 1 is characterized in that: The I / Q signal emitted by the radiation source in step (1) consists of an I carrier and a Q carrier, wherein the Q carrier is obtained by rotating the I carrier by 90 degrees.

3. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagram according to claim 1 is characterized in that: In the step (1), the I / Q signal data in the form of a one-dimensional sequence is converted into an equipotential constellation diagram in the form of a two-dimensional image, wherein the I carrier and the Q carrier corresponding to each sampling point are respectively represented as the horizontal coordinate and the vertical coordinate of the constellation diagram.

4. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagrams according to claim 3 is characterized in that: Based on the constellation diagram, a sliding window is used to traverse the constellation diagram in turn to calculate the number of data sample points in different areas of the constellation diagram; a rectangular window function of size W×H is used, where W represents the width of the rectangle and H represents the height of the rectangle; the point density ρ is used (W,H) (i, j) indicates how many sample points there are in the sliding window; Where (i, j) is the center coordinate of the rectangular window in the constellation diagram, dots (W,H) (i, j) is the number of points in the rectangular window, and N is the length of the I / Q waveform.

5. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagram according to claim 4 is characterized in that: The constellation map is colored according to the point density of different areas, where yellow represents higher density areas, green represents medium density areas, and blue represents lower density areas.

6. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagrams according to claim 1 is characterized in that: In step (2), based on the traditional generative adversarial network, the discriminator architecture is improved and a local discriminator is introduced. The local discriminator distinguishes whether each square n×n area with a side length of n in the image is true or false; the above local discriminator is run on the equipotential constellation map, and all discriminator responses are averaged to provide the final output. The final output is a matrix, each value represents the true or false of the corresponding block in the original image, and the average value is taken as the result after equalization. Its loss function is expressed as: Among them L adv (D ij (x),1) represents the probability value of the n×n local area in the i-th row and j-th column of the input true equipotential constellation being correctly judged as the true equipotential constellation, L adv (D ij (G(z)), 0) represents the probability value that the n×n local area in the i-th row and j-th column of the generated equipotential constellation is correctly determined as the generated equipotential constellation.

7. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagrams according to claim 6 is characterized in that: After using the generator to generate an equipotential constellation, the local discriminator and the global discriminator are input simultaneously to obtain their respective discrimination results. The two discriminators are updated only by their own losses, and the generator is updated using the sum of the two discriminator losses. The loss function of the global discriminator is expressed as: Where D(x) represents the output of the discriminator for the real sample, 1-D(G(z)) represents the output of the discriminator for the generated sample, and P data(x) represents the real data distribution, P G(z) Represents the generated data distribution; The overall loss function of the improved generative adversarial network is: L total =L GAN +L PatchGAN Among them L total Denotes the overall loss function of the improved generative adversarial network, L GAN represents the loss function of the global discriminator, L PatchGAN represents the loss function of the local discriminator.

8. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagrams according to claim 1 is characterized in that: The specific process of screening in step (3) is as follows: The equipotential constellation map is divided into small blocks, and then the brightness, contrast and structural information of each block are calculated. Then, this information is combined to calculate the similarity between the equipotential constellations. When calculating the brightness, contrast and structural information, SSIM uses a Gaussian weighting function to weight the pixels around each pixel. SSIM is expressed as follows: Among them, x and y represent two equipotential constellation diagrams, μ represents the average value of the image grayscale, σ represents the variance of the image grayscale, and c is a non-zero constant to prevent the denominator from being zero; the output of the SSIM algorithm is a value between 0 and 1, which represents the similarity between the two equipotential constellation diagrams, where 1 represents exactly the same equipotential constellation diagrams, and 0 represents completely different equipotential constellation diagrams.

9. The method for enhancing small sample radiation source data and identifying individuals based on equipotential constellation diagrams according to claim 1, characterized in that: The residual network in step (4) includes: (4-1) The network is composed of basic blocks, which are composed of convolutional layers, batch normalization layers and activation functions; (4-2) Residual connections are introduced between the basic blocks described in step (4-1) so that the input information of the previous layer is directly added to the output of the subsequent layer, building a channel for information to flow directly in the network. (4-3) The residual network uses a series of convolutional layers and residual blocks to extract image features and maps these features to category probabilities through fully connected layers and softmax activation functions.

Images