Radiation source individual open set identification method based on bispectrum characteristics

Abstract

The invention discloses a radiation source individual open set identification method based on bispectrum characteristics. The method comprises the following steps: firstly, preprocessing an electromagnetic radiation source signal, performing bispectrum estimation to obtain a bispectrum, calculating a contour integral bispectrum value according to an integral path, and inputting the obtained integral bispectrum value into a feature extraction module formed by introducing a one-dimensional residual convolutional network into a space attention module; and the feature extraction module extracts multi-scale individual fingerprint features to obtain an activation vector, Weibull model fitting prediction and OpenMax correction calculation are carried out in combination with a statistical extremum theory, and finally identification of an unknown category of electromagnetic radiation source individual in an open set environment is realized. According to the method, the calculation amount is reduced, the calculation efficiency is improved, and the identification accuracy of the electromagnetic radiation source individuals in the open set environment is improved.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, and in particular relates to a method for identifying individual open sets of radiation sources based on bispectral features. Background Technology

[0002] With the rapid development of wireless communication technology and the widespread application of various electronic devices, the electromagnetic spectrum environment is becoming increasingly complex, making the need for accurate identification and management of radiation sources more urgent. Accurate identification of individual electromagnetic radiation sources, as a technology capable of distinguishing different physical electromagnetic devices, demonstrates significant application value in fields such as spectrum management, network security, and tracing illegal signals by analyzing the unique, subtle features (i.e., radio frequency fingerprints) in the signals emitted by these devices, resulting from differences in hardware manufacturing.

[0003] Traditional methods for accurately identifying electromagnetic radiation sources rely primarily on macroscopic parameters such as modulation patterns and carrier frequencies. These methods can only identify the type of electromagnetic signal, not distinguish between different devices of the same model. The rise of deep learning technology has brought new opportunities for the accurate identification of individual electromagnetic radiation sources. Deep learning-based methods automatically learn and extract high-level features through neural networks, effectively capturing subtle differences in electromagnetic signals. Currently, many scholars both domestically and internationally have long been focusing on researching deep learning-based technologies for the accurate identification of individual electromagnetic radiation sources. However, in practical applications, electromagnetic radiation sources are diverse and constantly evolving, making it impractical to collect and train signals from all possible sources. When faced with unknown categories of radiation sources, closed-set identification methods may classify unknown sources as known categories. This "class bias" problem severely restricts the practicality and reliability of electromagnetic radiation source identification systems.

[0004] To address this issue, Open-Set Recognition (OSR) technology emerged. OSR can effectively distinguish between known and unknown categories even when the training data does not contain all possible categories. Its core idea is to set decision boundaries so that the system can not only correctly identify known categories but also detect new, unseen categories. Currently, research on accurate identification of individual electromagnetic radiation sources mainly focuses on detecting individuals of known categories (i.e., closed-set signals). Identifying individual electromagnetic radiation sources in open-set environments remains a pressing problem. Summary of the Invention

[0005] The purpose of this invention is to provide an open-set identification method for individual radiation sources based on bispectral features, thereby improving the accuracy of identification of individual electromagnetic radiation sources in an open-set environment.

[0006] To achieve the objective of this invention, a method for identifying individual open sets of radiation sources based on bispectral features is provided, comprising the following steps:

[0007] S1. Preprocess the individual signals of electromagnetic radiation sources to obtain a training set containing sample data of N types of known electromagnetic radiation source signals and a test set containing sample data of N types of known electromagnetic radiation source signals and 1 type of unknown electromagnetic radiation source signals.

[0008] S2. Input the training set into the feature extraction module to obtain the activation feature vectors of known electromagnetic radiation source signals; construct a Weibull distribution model of the activation vectors of individual samples of known electromagnetic radiation sources using the corresponding class centroid distance and statistical extremum theory of the activation feature vectors of the known electromagnetic radiation source signals; the feature extraction module includes a one-dimensional residual convolutional network and a spatial attention model;

[0009] S3. Input the test samples in the test set into the feature extraction module to obtain the activation feature vector; then, after fitting and predicting the activation feature vector of the test samples through the Weibull distribution model of the known category electromagnetic radiation source signal, obtain the corrected weight value.

[0010] S4. The activation feature vector of the test sample is corrected using the corrected weight value to obtain the corrected activation feature vector.

[0011] S5. The modified activation feature vector is used to identify individuals from unknown radiation sources through a scoring function.

[0012] A non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the above-described method for identifying individual radiation sources.

[0013] The significant advancement of this invention compared to existing technologies lies in:

[0014] (1) This invention identifies individual electromagnetic radiation source signals by extracting bispectral features. Bispectral features have good anti-noise performance and can retain detailed information that reflects the phase changes of individual radiation source fingerprint features to the greatest extent. Furthermore, by introducing a dimensionality reduction method, the bispectral estimation results can be integrated into a one-dimensional integral bispectral feature function value by performing contour integration on the bispectral estimation results, thus avoiding redundant information and reducing the amount of computation.

[0015] (2) This invention uses a one-dimensional residual convolutional network as a feature extraction module to avoid the loss of shallow features; and the kernel size is designed in layers, with kernels from large to small to gradually expand the receptive field and extract multi-scale radio frequency individual fingerprint features; at the same time, the spatial attention mechanism is integrated in the feature extraction module, so that the network pays more attention to the features of important positions in the integral bispectral sequence.

[0016] (3) This invention combines statistical extreme value theory to construct a Weibull distribution model, and adopts OpenMax instead of the traditional softmax layer as the output layer. By correcting the activation vector parameters, it distinguishes between individual signals of known radiation sources and individual signals of unknown radiation sources, and realizes accurate identification of individual electromagnetic radiation sources in an open set environment.

[0017] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of the steps of the present invention;

[0020] Figure 2 This is the contour integral path diagram of the present invention;

[0021] Figure 3 This is a diagram of the residual convolutional network structure of the present invention;

[0022] Figure 4 This is a structural diagram of the residual module of the present invention, wherein... Figure 4 (a) is a structural diagram of the first three residual modules. Figure 4 (b) is a schematic diagram of the structure of the last three residual modules. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This invention provides a method for identifying individual open-set radiation sources based on bispectral features, combined with... Figure 1 This includes the following steps:

[0025] S1. Preprocess the individual signals of electromagnetic radiation sources to obtain a training set containing sample data of N types of known electromagnetic radiation source signals and a test set containing sample data of N types of known electromagnetic radiation source signals and 1 type of unknown electromagnetic radiation source signals.

[0026] S2. Input the training set into the feature extraction module to obtain the activation feature vectors of known electromagnetic radiation source signals; construct a Weibull distribution model of the activation vectors of individual samples of known electromagnetic radiation sources using the corresponding class centroid distance of the activation feature vectors of the known electromagnetic radiation source signals and the Extreme Value Theory (EVT); the feature extraction module uses a one-dimensional residual convolutional network and a spatial attention model, combined with... Figure 3 The one-dimensional residual convolutional network consists of six cascaded residual modules. The bispectral feature sequence obtained in step 1 is first input into a convolutional layer with a kernel size of 1×8. The first convolutional layer uses a small kernel to capture short-term local features, such as phase abrupt changes or edges in the bispectral sequence. The kernel sizes of the residual modules are set to 1×16 and 1×32, respectively. Figure 4 ,; among which, the first 3 convolutional modules are as follows Figure 4 As shown in (a), the kernel size of the convolutional layer is 1×16; the last three convolutional modules are as follows: Figure 4 As shown in (b), the kernel size of the convolutional layer is 1×32. Larger kernels can capture global features. The convolutional layers are designed in layers, with the kernel size increasing from small to large, gradually expanding the receptive field and extracting detailed fingerprint features of individuals from multi-scale radiation sources. At the same time, the residual module uses a skip connection network structure to superimpose shallow and deep features, effectively avoiding the loss of shallow features and gradient explosion during network training.

[0027] The number of input and output channels for each convolutional layer is shown in the figure. A batch normalization (BN) layer is added after each convolutional layer to accelerate convergence, and a dropout layer is added to prevent overfitting and improve stability. After feature normalization via the 6th residual module, the features are multiplied by the coefficients from the spatial attention module, making the network focus more on features at important temporal positions in the bispectral sequence. The resulting features are then passed through an average pooling layer and a fully connected layer to output the activation vector AV.

[0028] S3. Input the test samples in the test set into the feature extraction module to obtain the activation feature vector; then, after fitting and predicting the activation feature vector of the test samples through the Weibull distribution model of the known category electromagnetic radiation source signal, obtain the corrected weight value.

[0029] S4. The activation feature vector of the test sample is corrected using the corrected weight value to obtain the corrected activation feature vector.

[0030] S5. The modified activation feature vector is used to identify individuals from unknown radiation sources through a scoring function (OpenMax).

[0031] The preprocessing in S1 involves estimating the bispectral data of the detected electromagnetic radiation source signal to obtain a bispectral map. To avoid information redundancy and reduce computational complexity, the bispectral map is then integrally processed to obtain integrated bispectral values. The integrated bispectral value data is then divided into a training set and a test set.

[0032] The bispectral image obtained in S1 is a two-dimensional feature with a large data volume, and direct application would lead to complex computations. To overcome this difficulty, an integral bispectral method is introduced to transform the two-dimensional bispectral image into a one-dimensional function. Among them, contour integral is one of the most common integral bispectral methods, which specifically includes the following steps:

[0033] S11. Perform bispectral estimation on the electromagnetic radiation source signal s(n);

[0034] S12. Extract the contour integral bispectral value based on the bispectral estimation;

[0035] The discrete signal of the electromagnetic radiation source after sampling s(n) for:

[0036]

[0037] Where N represents the total number of signal sampling points, w(n) represents Gaussian white noise, and w(n) and s(n) are independent of each other; the discrete signal x(n) is divided into K segments, and the third-order cumulant of each segment is calculated, as shown in the following formula:

[0038] ;

[0039] Where i and j are the delay points, , , The number of time sampling points for each segment. ;

[0040] Expanding the above formula, we get:

[0041] ;

[0042] If the mean of both the signal and the Gaussian white noise is zero, then... ;

[0043] The summation and average of the third-order cumulants of the K segments are taken as the third-order cumulant of the signal, as shown in the following formula:

[0044] ;

[0045] Bispectral estimation is the estimation of the third-order cumulant of a signal. Perform a second-order Fourier transform, as shown in the following equation:

[0046] ;

[0047] Because w(n) is Gaussian white noise, This is essentially negligible. Therefore, the bispectral estimation result of signal x(n) is basically determined by the third-order cumulant of the radiation source signal itself. The decision was made. This demonstrates that bispectral analysis can effectively suppress the influence of Gaussian white noise on the signal, thereby maximizing the preservation of the individual characteristics of the radiation source signal.

[0048] The integration path of the contour integral bispectrum is a square centered at the origin, combined with... Figure 2 Each circle of solid black lines The path represents the contour integral, with each black dot indicating the corresponding bispectral estimate. It can be seen that the contour integral bispectral representation eliminates redundant information without missing bispectral values, capturing crucial target information while preserving signal phase information. The contour integral bispectral value can be denoted as...

[0049] ;

[0050] in, This represents the sum of bispectral estimates along each integration path. Since individual fingerprint features are subtle, a 1024-point FFT resolution is chosen to preserve more details in the bispectral integration, hence L=1024. The bispectral value of each contour integration is a one-dimensional feature vector of length 1024.

[0051] The one-dimensional residual convolutional network of S2 is composed of several cascaded residual modules. Each residual module includes a convolutional layer, a batch normalization layer, a ReLU layer, and a Dropout layer. The feature vector after passing through the several residual modules and being normalized is multiplied by the coefficients of the spatial attention module. The feature vector is then passed through a fully connected layer to obtain the activation feature vector of the known electromagnetic radiation source signal.

[0052] The construction of the Weibull distribution model of S2 includes the following steps;

[0053] S21. Record the activation vectors (AV) of N known categories of electromagnetic radiation source signals.

[0054] S22. Determine the first [type] based on the activation feature vector of the known category of electromagnetic radiation source signals. Mean Activation Vectors (MAV, or centroids) of each category.

[0055] S23. Based on the average activation vector, determine the distances from the activation vectors of all samples in the nth class to the centroids of their corresponding classes, and sort them from largest to smallest to obtain the nth class. Distance distribution of each category;

[0056] S24. According to the statistical extreme value theory, select the first t maxima for Weibull fitting, and establish the th... A Weibull model of activation vectors for each category;

[0057] S25. Repeat the above steps until the Weibull distribution model of the activation vectors of individual samples of all known categories of radiation sources is obtained.

[0058] Step 2 is specifically shown in the following formula:

[0059] S21, ;

[0060] in, Indicates the category to which it belongs, with a value range of 100. ; Indicates the first Class common One sample; Indicates the first of this category There are 10 samples, with a value range of 1000. ; Let be the set of activation vectors for all samples in the nth class; Let be the activation vector value of the j-th sample in the n-th category;

[0061] S22, ;

[0062] in, Indicates the first The center point of the distribution space of individual samples of a known class of radiation sources, and the edge sample points that are farther away from the center point constitute the boundary of the class;

[0063] S23, ; ;

[0064] in, Let be the distance from the activation vector of the j-th sample of this category to the centroid of the corresponding category. Let be the distance from the activation vector of the Mth sample in this category to the centroid of the corresponding category. For the first Distance distribution of each category;

[0065] S24, ; ;

[0066] in, For the first The probability density function (PDF) of the Weibull model for each class activation vector. For the location parameters of the Weibull model of the nth category, Let n be the shape parameters of the Weibull model for the nth category. Let n be the Weibull model scaling parameter for the nth category. To Integrating yields the cumulative distribution function. For the first Distance distribution of each category The value in.

[0067] S3 specifically includes the following steps:

[0068] S31. Record the test samples in the test set. The activation feature vector obtained after inputting the feature extraction module:

[0069] S32, Based on the test sample The components in the activation feature vector are sorted in descending order, and the index values ​​of each component after sorting are recorded as the index values ​​of the largest component.

[0070] S33, Based on the sample to be tested The activation feature vector determines its distance from the center point of the distribution space of individual samples of each known class of radiation sources:

[0071] S34. Based on the probability density function of the Weibull model of the known class, determine the distance between the aforementioned center points that appears in the first... The probability in the distance distribution of class samples;

[0072] S35. Determine the test sample based on the distance between the center points and the cumulative distribution function of each known class. The corrected weight values ​​for each component of the activated feature vector corresponding to the category;

[0073] S36. The test sample is determined according to the corrected weight value. Each component in the activation feature vector is modified;

[0074] S37. Based on the corrected weight value, in the sample to be tested... Add the first to the activation feature vector 3D pseudo-activation feature vector components;

[0075] S38, according to the aforementioned first The test sample is obtained from the pseudo-activation feature vector components. The corrected activation feature vector.

[0076] The specific formula for S3 is as follows:

[0077] S31, ;

[0078] in, The range of values ​​is , for The Each component value This is the activation feature vector obtained after inputting into the feature extraction module;

[0079] S32, Based on the test sample Activation feature vector Sort the components in descending order, and then sort the values ​​of each component. The index value n is recorded sequentially as the index value of the i-th largest component. ;

[0080] S33, ;

[0081] in, For the test sample The distance between the activation feature vector and the center point of each known class. For the test sample Activation feature vector;

[0082] S34. Determine the distance between the aforementioned center points based on the Weibull model probability density function of the known class. Appeared in the The probability in the distance distribution of class samples;

[0083] S35, ;

[0084] in, The corrected weight values ​​for each component corresponding to its category. For the test sample The index value of the nth largest component of each component of the activation feature vector. The number of categories is known;

[0085] S36 ;

[0086] in, for The Middle The corrected value of each component, For the product of Hadamard, To adjust the weight values;

[0087] S37 ;

[0088] in, For the first 3D pseudo-activation feature vector components

[0089] S38 ;

[0090] in, For the sample to be tested The corrected activation feature vector.

[0091] The scoring function of S5 uses the OpenMax function to transform feature values ​​into class probabilities, and determines the class of a sample based on the maximum value of the probability distribution, as shown in the following formula:

[0092] ;

[0093] in, The sample to be tested belongs to the first The probability value of the category, To distinguish categorical variables.

[0094] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0095] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying individual open sets of radiation sources based on bispectral features, characterized in that, Includes the following steps: S1. Preprocess the individual signals of electromagnetic radiation sources to obtain a training set containing sample data of N types of known electromagnetic radiation source signals and a test set containing sample data of N types of known electromagnetic radiation source signals and 1 type of unknown electromagnetic radiation source signals. S2. Input the training set into the feature extraction module to obtain the activation feature vectors of known electromagnetic radiation source signals; construct a Weibull distribution model of the activation vectors of individual samples of known electromagnetic radiation sources using the corresponding class centroid distance and statistical extremum theory of the activation feature vectors of the known electromagnetic radiation source signals; the feature extraction module includes a one-dimensional residual convolutional network and a spatial attention model; S3. Input the test samples in the test set into the feature extraction module to obtain the activation feature vector; The corrected weight value is obtained by fitting and predicting the activation feature vector of the sample to be tested through the Weibull distribution model of the known category of electromagnetic radiation source signal; S4. The activation feature vector of the test sample is corrected using the corrected weight value to obtain the corrected activation feature vector. S5. The modified activation feature vector is used to identify individuals from unknown radiation sources through a scoring function.

2. The method for identifying individual radiation sources based on bispectral features according to claim 1, characterized in that, The preprocessing in S1 involves estimating the bispectral data of the detected electromagnetic radiation source signal to obtain a bispectral map; then, the bispectral map is integrated along the contour lines to obtain integrated bispectral values, and the integrated bispectral value data is divided into a training set and a test set.

3. The method for identifying individual radiation sources based on bispectral features according to claim 1, characterized in that, The one-dimensional residual convolutional network of S2 is composed of several cascaded residual modules. Each residual module includes a convolutional layer, a batch normalization layer, a ReLU layer, and a Dropout layer. The feature vector after passing through the several residual modules and being normalized is multiplied by the coefficients of the spatial attention module. The feature vector is then passed through a fully connected layer to obtain the activation feature vector of the known electromagnetic radiation source signal.

4. The method for identifying individual radiation sources based on bispectral features according to claim 3, characterized in that, The number of residual modules is six, the kernel size of the first three convolutional layers is 1×16, and the kernel size of the last three convolutional layers is 1×32.

5. The method for identifying individual radiation sources based on bispectral features according to claim 1, characterized in that, The construction of the Weibull distribution model of S2 includes the following steps; S21. Record the activation feature vectors of N known categories of electromagnetic radiation source signals; S22. Determine the first [type] based on the activation feature vector of the known category of electromagnetic radiation source signals. The average activation vector of each category; S23. Based on the average activation vector, determine the distances from the activation vectors of all samples in the nth class to the centroids of their corresponding classes, and sort them from largest to smallest to obtain the nth class. Distance distribution of each category; S24. According to the statistical extreme value theory, select the first t maxima for Weibull fitting, and establish the th... A Weibull model of activation vectors for each category; S25. Repeat the above steps until the Weibull distribution model of the activation vectors of individual samples of all known categories of radiation sources is obtained.

6. The method for identifying individual radiation sources based on bispectral features according to claim 5, characterized in that, Step 2 is specifically shown in the following formula: S21、 ； in, Indicates the category to which it belongs, with a value range of 100. ; Indicates the first Class common One sample; Indicates the first of this category There are 10 samples, with a value range of 1000. ; Let be the set of activation vectors for all samples in the nth class; Let be the activation vector value of the j-th sample in the n-th category; S22、 ； in, Indicates the first The center point of the distribution space of individual samples of a known class of radiation sources, and the edge sample points that are farther away from the center point constitute the boundary of the class; S23、 ； ； in, Let be the distance from the activation vector of the j-th sample of this category to the centroid of the corresponding category. Let be the distance from the activation vector of the Mth sample in this category to the centroid of the corresponding category. For the first Distance distribution of each category; S24、 ； ； in, For the first The probability density function of a Weibull model with activation vectors for each category. For the location parameters of the Weibull model of the nth category, Let n be the shape parameters of the Weibull model for the nth category. Let n be the Weibull model scaling parameter for the nth category. To Integrating yields the cumulative distribution function. For the first Distance distribution of each category The value in.

7. The method for identifying individual open sets of radiation sources based on bispectral features according to claim 6, characterized in that, S3 specifically includes the following steps: S31. Record the test samples in the test set. The activation feature vector obtained after inputting the feature extraction module: S32, Based on the test sample The components in the activation feature vector are sorted in descending order, and the index values ​​of each component after sorting are recorded as the index values ​​of the largest component. S33, Based on the sample to be tested The activation feature vector determines its distance from the center point of the distribution space of individual samples of each known class of radiation sources: S34. Based on the probability density function of the Weibull model of the known class, determine the distance between the aforementioned center points that appears in the first... The probability in the distance distribution of class samples; S35. Determine the test sample based on the distance between the center points and the cumulative distribution function of each known class. The corrected weight values ​​for each component of the activated feature vector corresponding to the category; S36. The test sample is determined according to the corrected weight value. Each component in the activation feature vector is modified; S37. Based on the corrected weight value, in the sample to be tested... Add the first to the activation feature vector 3D pseudo-activation feature vector components; S38, according to the aforementioned first The test sample is obtained from the pseudo-activation feature vector components. The corrected activation feature vector.

8. The method for identifying individual radiation sources based on bispectral features according to claim 7, characterized in that, The specific formula for S3 is as follows: S31、 ； in, The range of values ​​is , for The Each component value This is the activation feature vector obtained after inputting into the feature extraction module; S32, Based on the test sample Activation feature vector Sort the components in descending order, and then sort the values ​​of each component. The index value n is recorded sequentially as the index value of the i-th largest component. ; S33、 ； in, For the test sample The distance between the activation feature vector and the center point of each known class. For the test sample Activation feature vector; S34. Determine the distance between the aforementioned center points based on the Weibull model probability density function of the known class. Appeared in the The probability in the distance distribution of class samples; S35、 ； in, The corrected weight values ​​for each component corresponding to its category. For the test sample The index value of the nth largest component of each component of the activation feature vector. The number of categories is known; S36、 ； in, for The Middle The corrected value of each component, For Hadama product, To adjust the weight values; S37、 ； in, For the first 3D pseudo-activation feature vector components S38、 ； in, For the sample to be tested The corrected activation feature vector.

9. The method for identifying individual radiation sources based on bispectral features according to claim 8, characterized in that, The scoring function of S5 uses the OpenMax function to transform feature values ​​into class probabilities, and determines the class of a sample based on the maximum value of the probability distribution, as shown in the following formula: ； in, The sample to be tested belongs to the first The probability value of the category, To distinguish categorical variables.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1 to 9.

Images