A radar active jamming small sample identification method

Abstract

This invention belongs to the field of radar signal processing technology and relates to a method for identifying radar active interference in small samples. The method includes constructing a simulated radar active interference sample set and performing short-time Fourier transform processing, dividing it into a support set and a query set according to the small sample learning method; embedding multi-scale asymmetric features of the sample input into a network to extract multi-scale and directional features; correcting the category prototype center through an improved prototype network and prototype correction method; generating a distance scaling coefficient using a distance sensing module; completing interference category discrimination and updating model parameters through weighted distance measurement; and using the trained model to classify and identify test samples. This invention solves the problems of inaccurate category prototype representation and poor adaptability of distance measurement in small sample radar active interference identification through the collaborative design of prototype correction and distance sensing guidance.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, specifically to a method for identifying radar active interference in small sample sizes, applicable to the identification of radar active interference types in complex electromagnetic environments under small sample conditions. Background Technology

[0002] As a core sensing device in modern defense and civilian fields, radar operates in increasingly complex environments. The presence of various active interferences seriously affects the radar's ability to detect and track targets. Therefore, radar active interference identification has become a key aspect of radar anti-jamming technology.

[0003] In existing technologies, radar active interference identification methods based on deep learning rely on large-scale labeled samples for model training. However, in real-world environments, interference signals are diverse, complex, and difficult to obtain, often making it difficult to acquire sufficient labeled samples. This results in traditional deep learning methods having poor generalization performance and low recognition accuracy in small-sample scenarios.

[0004] To address the problem of small sample recognition, meta-learning methods such as prototype networks have been applied to radar active interference identification. These methods use the mean of the features of the support set samples as the category prototype and classify based on the distance between the query sample and the category prototype. However, existing prototype network methods have two major drawbacks: First, the category prototype is calculated using only a simple mean, which is easily affected by anomalous samples in the support set, leading to inaccurate prototype representation and thus affecting classification accuracy. Second, the distance metric uses a fixed method, failing to consider the different correlations between different query samples and the category prototype, and cannot adaptively adjust the distance weights, resulting in insufficient robustness when facing interference categories with different feature distributions.

[0005] In view of the shortcomings of the existing technologies, how to improve the accuracy and robustness of radar active interference identification by optimizing the category prototype representation and distance measurement method under small sample conditions has become an urgent technical problem to be solved, and it is also the original intention of the development of this invention. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying radar active interference in small samples. This method optimizes the category prototype representation through prototype correction and achieves adaptive distance measurement with the help of distance perception guidance, thereby solving the above-mentioned technical problems and improving the recognition performance under small sample conditions.

[0007] To achieve the objectives of this invention, the following technical solutions will be adopted.

[0008] A method for identifying small-sample radar active interference includes the following steps:

[0009] S1. Construct a radar active interference simulation sample set, and perform short-time Fourier transform processing on the sample set. Divide the samples into support set and query set according to the small sample learning method to generate tasks for model training and testing.

[0010] S2. Input the support set and query set samples into the multi-scale asymmetric feature embedding network to extract the multi-scale and directional features of radar active interference and obtain the corresponding feature vector.

[0011] S3. Input the feature vectors of the support set samples into the improved prototype network, and correct the prototype center of each interference category through the prototype correction method to obtain the corrected category prototype.

[0012] S4. Based on the feature vectors of the query set samples and the support set samples, use the distance-aware module to generate the distance scaling coefficient between the query samples and each category;

[0013] S5. Use the distance scaling factor to weight the distance between the query sample and each category of correction prototype to complete the interference category discrimination, and update the model parameters through backpropagation of the loss function until the model converges.

[0014] S6. Use the converged model to classify and identify the radar active interference samples in the test task, and output the corresponding interference type results.

[0015] Furthermore, the few-shot learning method is the N-way K-shot paradigm, where N is the number of interference categories in each task, K is the number of samples of each category in the support set, and the value of K ranges from 5 to 10, which meets the application requirements of few-shot scenarios.

[0016] Furthermore, the multi-scale asymmetric feature embedding network adopts a structure that combines multi-scale convolution with directional attention enhancement, which can simultaneously capture multi-scale detailed features and directional features of radar active interference signals, solving the problems of incomplete single-scale feature extraction and insufficient feature discrimination.

[0017] Furthermore, the prototype correction method includes the following steps:

[0018] 1. Calculate the initial prototype center for each interference category;

[0019] 2. Remove samples with abnormal features from the support set through anomaly sample detection;

[0020] 3. The prototype center is revised based on valid samples to make the category prototype more closely match the true feature distribution.

[0021] Furthermore, the abnormal sample detection utilizes an intra-class correlation matrix to quickly identify and remove abnormal samples from the support set, thus avoiding interference from abnormal samples to the prototype center.

[0022] Furthermore, the distance-aware module dynamically generates a distance scaling factor by calculating the feature correlation between the query sample and the prototype of each category, enabling the distance metric to adaptively match the degree of association between different samples and categories.

[0023] Furthermore, the weighted distance metric adjusts the original distance using a distance scaling factor to minimize the weighted distance for query samples that are strongly correlated with the category prototype, thereby improving the accuracy of category discrimination.

[0024] Furthermore, the loss function adopts the cross-entropy loss function, and all parameters of the multi-scale asymmetric feature embedding network, the improved prototype network, and the distance-aware module are updated through backpropagation until the model converges.

[0025] The beneficial effects of this invention are:

[0026] 1. This invention effectively solves the technical problems of category prototypes being easily interfered with by abnormal samples and having inaccurate representation in existing prototype networks by using a prototype correction method. This makes the category prototypes more closely match the true feature distribution of the interference signals, significantly improves the representation accuracy of the category prototypes, and provides a reliable basis for classification and discrimination.

[0027] 2. This invention introduces a distance sensing module to dynamically generate a distance scaling coefficient, thereby achieving adaptive adjustment of the distance measurement. This solves the problem of poor adaptability of existing fixed distance measurement methods, and can better match the differences in feature distribution of different interference categories, thus improving the robustness of recognition.

[0028] 3. The multi-scale asymmetric feature embedding network of the present invention can fully extract the multi-scale and directional features of radar active interference, solve the problems of incomplete single-scale feature extraction and insufficient feature discrimination, and provide high-quality feature support for subsequent prototype correction and distance measurement.

[0029] 4. The technical solution of this invention is specifically designed for small sample scenarios. It does not require large-scale sample labeling and can still maintain high recognition accuracy and robustness under small sample conditions. It is suitable for radar active interference identification in complex electromagnetic environments and has important engineering application value. It has significant progress compared with existing technologies. Attached Figure Description

[0030] Figure 1 This is a flowchart of a radar active interference small sample identification method according to the present invention;

[0031] Figure 2 This is a schematic diagram of the structure of the multi-scale asymmetric convolution module described in this invention;

[0032] Figure 3This is a schematic diagram of the distance sensing module structure described in this invention. Detailed Implementation

[0033] The present invention will be further described in conjunction with the accompanying drawings and embodiments.

[0034] As an embodiment 1 of the present invention, such as Figure 1 As shown, a method for identifying small-sample radar active interference includes the following steps:

[0035] S1: Construct a radar active interference simulation sample set, and perform short-time Fourier transform processing on the sample set. Divide the samples into support set and query set according to the small sample learning method to generate tasks for model training and testing.

[0036] The specific process includes the following steps:

[0037] S11: Based on the LFM radar transmitted signal, MATLAB simulation software was used to simulate radar active jamming signals, generating radar active jamming signals of 10 types, and constructing radar active jamming simulation data samples. The radar active jamming types include single jamming and compound jamming types. Single jamming types include: Noise Amplitude Modulation Jamming (NAMJ), Noise Frequency Modulation Jamming (NFMJ), Smart Noise Jamming (SNJ), Range Dense False Target Jamming (RMT), Intermittent Sample-and-Relay Jamming (ISRJ), Spectral Dispersion Jamming (SMSP), and Comb Jamming (COMB). The compound jamming types are ISRJ+SMSP, ISRJ+COMB, and SMSP+COMB, respectively.

[0038] In this embodiment, the 10 different types of active radar jamming signals are set in 2dB increments within a noise-to-interference ratio (NNR) range of −8dB to 10dB. 50 active radar jamming signal samples are generated for each NNR condition, resulting in 500 jamming signals for each jamming type. See Table 1 for the active radar jamming parameter settings.

[0039] Table 1: Radar Active Interference Parameter Settings

[0040] S12: Perform short-time Fourier transform processing on the radar active interference signals in the simulation data sample set to obtain a two-dimensional time-frequency image as the input sample for subsequent feature extraction. The expression for its short-time Fourier transform is:

[0041] ;

[0042] In the formula, For interference signals at time and frequency Time-frequency representation at the location, The variable is the integral variable, representing the time shift parameter. This is an interference signal. For window functions, This indicates the conjugate operation. The imaginary unit;

[0043] S13: The two-dimensional time-frequency image is divided into a meta-training set, a meta-validation set, and a meta-test set in a ratio of 7:2:1; in each meta-task, for each type of radar active interference, the support set and query set are further divided in a ratio of 1:4.

[0044] S2: Input the support set and query set samples into the multi-scale asymmetric feature embedding network to extract the multi-scale and directional features of radar active interference and obtain the corresponding feature vector;

[0045] The specific process includes the following steps:

[0046] S21: Input the time-frequency image into a multi-scale asymmetric feature embedding network to obtain the support set feature vector and the query set feature vector. The embedding network sequentially includes a Conv0 layer, a MaxPool layer, a cascaded MACB layer, an AvgPool layer, and a Flatten layer. See Table 2.

[0047] S22: The MACB cascaded layer is composed of three identical multi-scale asymmetric convolutional modules connected in series, such as... Figure 2 As shown, each of the multi-scale asymmetric convolutional modules includes a first multi-scale parallel convolutional layer, a second asymmetric parallel convolutional layer, and a third convolutional fusion layer. Each convolutional layer is followed by a batch normalization layer and an activation layer. A TA attention mechanism is introduced at the end of the module, and residual connections are set to output feature maps.

[0048] S23: The kernel size of the first multi-scale parallel convolutional layer is 1×1, 3×3 and 5×5, the kernel size of the second asymmetric parallel convolutional layer is 3×3, 1×3 and 3×1, and the kernel size of the third convolutional fusion layer is 3×3.

[0049] S3: Input the feature vectors of the support set samples into the improved prototype network, and correct the prototype center of each interference category through the prototype correction method to obtain the corrected category prototype.

[0050] The specific process includes the following steps:

[0051] S31: Extract the support set feature vectors, and group the support set sample feature vectors according to the interference category to obtain the support feature vector set corresponding to each category. The expression for the support feature vector set is:

[0052] ;

[0053] In the formula, Indicates the first The set of support set feature vectors corresponding to each type of interference. This indicates the number of samples in the support set for each type of interference. Indicates matrix transpose. .

[0054] S32: Construct an intra-class correlation matrix for the support feature vector set of each category, and set the diagonal elements of the intra-class correlation matrix to zero to obtain the intra-class similarity matrix among the support samples of that category. The expression is:

[0055] ;

[0056] In the formula, Indicates the first The set of support set feature vectors corresponding to each type of interference. The operation represents extracting the diagonal elements of a matrix;

[0057] S33: Calculate the average relevance score for each supporting sample based on the intra-class similarity representation, using the following formula:

[0058] , ;

[0059] In the formula: Indicates the first Class support set The average relevance score of each supporting sample Indicates the first The feature vector of the nth support sample and the nth support sample feature vector The intra-class similarity representation value between the feature vectors of each supporting sample.

[0060] S34: Normalize the average relevance score to obtain the intra-class weights, calculated using the following formula:

[0061] , ;

[0062] In the formula: Indicates the first Class support set The intra-class weights of each supporting sample.

[0063] S35. Using the intra-class weights, perform a weighted summation of the support feature vectors for that class to obtain the corrected prototype center of that class. The calculation formula is as follows:

[0064] ;

[0065] In the formula: Indicates the first Class correction prototype center, Indicates the first The feature vectors of the supporting samples.

[0066] S4: Based on the feature vectors of the query set samples and support set samples, the distance-aware module generates distance scaling coefficients between the query samples and each category, such as... Figure 3 As shown;

[0067] The specific process includes the following steps:

[0068] S41: Project the feature vectors of the query set and support set onto the attention subspace, and use the attention mechanism to obtain the inter-class correlation matrix between the query samples and support samples. The calculation formula is as follows:

[0069] ;

[0070] ;

[0071] In the formula, , , These represent the query matrix, key matrix, and value matrix, respectively. Represents the feature vector matrix of the query set. This represents the eigenvector matrix of the support set. , , Let each represent a learnable linear mapping parameter matrix. Represents the feature dimension of the attention subspace. This represents the inter-class correlation matrix between the query sample and the supporting samples;

[0072] S42: Generate a distance scaling factor matrix between the query sample and each category based on the inter-class correlation matrix. The calculation formula is:

[0073] ;

[0074] In the formula, It is a two-layer fully connected network. For the Sigmoid function;

[0075] S5: Use distance scaling factor to weight the distance between the query sample and each category of correction prototype to complete the interference category discrimination, and update the model parameters through backpropagation of loss function;

[0076] The specific process includes the following steps:

[0077] S51: Based on the distance scaling factor matrix, the distance between the query sample and the correction prototype center of each interference category is weighted and measured to obtain the weighted distance between the query sample and each interference category. The calculation formula is as follows:

[0078] ;

[0079] In the formula, Indicates the query sample in the distance scaling factor matrix With the Distance scaling factor between classes The distance is Euclidean.

[0080] S52: Calculate the predicted probability of the query sample belonging to each interference category based on the weighted distance. The calculation formula is as follows:

[0081] ;

[0082] In the formula, Indicates query sample The true label;

[0083] S53: Construct a classification loss function based on the true label of the query sample and the predicted probability. The calculation formula is as follows:

[0084] ;

[0085] S6: The network parameters use the Adam optimizer during the model training phase, with an initial learning rate set to 0.0001. The trained model is then used to classify and identify radar active interference samples in the test task, outputting the corresponding interference type results.

[0086] As an embodiment 2 of the present invention, such as Figure 1 As shown, a method for identifying small-sample radar active interference includes the following steps:

[0087] S1: Construct a radar active interference simulation sample set, and perform short-time Fourier transform processing on the sample set to divide it into a known interference category set and an unknown interference category set. Following the rules for small-sample learning tasks, extract a support set and a query set from the known interference category set for model training. During the testing phase, samples from the unknown interference category set are mixed into the query set to construct a test task capable of identifying unknown interference.

[0088] The specific process includes the following steps:

[0089] S11: Based on the LFM radar transmitted signal, MATLAB simulation software was used to simulate radar active jamming signals, generating radar active jamming signals of 10 types, and constructing radar active jamming simulation data samples. The radar active jamming types include single jamming and composite jamming types. Single jamming types include: Noise Amplitude Modulation Jamming (NAMJ), Noise Frequency Modulation Jamming (NFMJ), Smart Noise Jamming (SNJ), Range Dense Target Jamming (RMT), Intermittent Sample-and-Relay Jamming (ISRJ), Spectral Dispersion Jamming (SMSP), and Comb Jamming (COMB). Composite jamming types are ISRJ+SMSP, ISRJ+COMB, and SMSP+COMB, respectively. Among the above 10 types of radar active jamming, NAMJ, NFMJ, SNJ, RMT, and ISRJ are selected as known jamming; SMSP, COMB, ISRJ+SMSP, ISRJ+COMB, and SMSP+COMB are selected as unknown jamming.

[0090] In this embodiment, the 10 different types of radar active jamming signals are set in 2dB increments within a noise-to-interference ratio (NIRS) range of −8dB to 10dB. 50 radar active jamming signal samples are generated for each NIRS condition, resulting in 500 jamming signals for each jamming type. See Table 3 for the radar active jamming parameter settings.

[0091] Table 3. Radar active jamming parameter settings:

[0092] S12: Perform short-time Fourier transform processing on the radar active interference signals in the simulation data sample set to obtain a two-dimensional time-frequency image as the input sample for subsequent feature extraction. The expression for its short-time Fourier transform is:

[0093] ;

[0094] In the formula: For interference signals at time and frequency Time-frequency representation at the location, The variable is the integral variable, representing the time shift parameter. This is an interference signal. For window functions, This indicates the conjugate operation. The imaginary unit;

[0095] S13: The two-dimensional time-frequency image is divided into a meta-training set, a meta-validation set, and a meta-test set in a ratio of 7:2:1; in each meta-task, for each type of radar active interference, the support set and query set are further divided in a ratio of 1:4.

[0096] S2: Input the support set and query set samples into the multi-scale asymmetric feature embedding network to extract the multi-scale and directional features of radar active interference and obtain the corresponding feature vector;

[0097] The specific process includes the following steps:

[0098] S21: Input the time-frequency image into a multi-scale asymmetric feature embedding network to obtain the support set feature vector and the query set feature vector. The embedding network consists of a Conv0 layer, a MaxPool layer, a cascaded MACB layer, an AvgPool layer, and a Flatten layer. See Table 4 for the parameters of the multi-scale asymmetric feature embedding network.

[0099] Table 4. Parameters of the Multi-Scale Asymmetric Feature Embedding Network:

[0100] S22: The MACB cascaded layer is composed of three identical multi-scale asymmetric convolutional modules connected in series, such as... Figure 2 As shown, each of the multi-scale asymmetric convolutional modules includes a first multi-scale parallel convolutional layer, a second asymmetric parallel convolutional layer, and a third convolutional fusion layer. Each convolutional layer is followed by a batch normalization layer and an activation layer. A TA attention mechanism is introduced at the end of the module, and residual connections are set to output feature maps.

[0101] S23: The kernel size of the first multi-scale parallel convolutional layer is 1×1, 3×3 and 5×5, the kernel size of the second asymmetric parallel convolutional layer is 3×3, 1×3 and 3×1, and the kernel size of the third convolutional fusion layer is 3×3.

[0102] S3: Input the feature vectors of the support set samples into the improved prototype network, and correct the prototype center of each interference category through the prototype correction method to obtain the corrected category prototype.

[0103] The specific process includes the following steps:

[0104] S31: Extract the support set feature vectors, and group the support set sample feature vectors according to the interference category to obtain the support feature vector set corresponding to each category. The expression for the support feature vector set is:

[0105] ;

[0106] In the formula, Indicates the first The set of support set feature vectors corresponding to each type of interference. This indicates the number of samples in the support set for each type of interference. Indicates matrix transpose. ;

[0107] S32: Construct an intra-class correlation matrix for the support feature vector set of each category, and set the diagonal elements of the intra-class correlation matrix to zero to obtain the intra-class similarity matrix among the support samples of that category. The expression is:

[0108] ;

[0109] In the formula, Indicates the first The set of support set feature vectors corresponding to each type of interference. The operation represents extracting the diagonal elements of a matrix;

[0110] S33: Calculate the average relevance score for each supporting sample based on the intra-class similarity representation, using the following formula:

[0111] , ;

[0112] In the formula: Indicates the first Class support set The average relevance score of each supporting sample Indicates the first The feature vector of the nth support sample and the nth support sample feature vector The intra-class similarity representation value between the feature vectors of each supporting sample.

[0113] S34: Normalize the average relevance score to obtain the intra-class weights, calculated using the following formula:

[0114] , ;

[0115] In the formula: Indicates the first Class support set The intra-class weights of each supporting sample.

[0116] S35. Using the intra-class weights, perform a weighted summation of the support feature vectors for that class to obtain the corrected prototype center of that class. The calculation formula is as follows:

[0117] ;

[0118] In the formula: Indicates the first Class correction prototype center, Indicates the first The feature vectors of the supporting samples.

[0119] S4: Based on the feature vectors of the query set samples and support set samples, the distance-aware module generates distance scaling coefficients between the query samples and each category, such as... Figure 3 As shown;

[0120] The specific process includes the following steps:

[0121] S41: Project the feature vectors of the query set and support set onto the attention subspace, and use the attention mechanism to obtain the inter-class correlation matrix between the query samples and support samples. The calculation formula is as follows:

[0122] ;

[0123] ;

[0124] In the formula, , , These represent the query matrix, key matrix, and value matrix, respectively. Represents the feature vector matrix of the query set. This represents the eigenvector matrix of the support set. , , Let each represent a learnable linear mapping parameter matrix. Represents the feature dimension of the attention subspace. This represents the inter-class correlation matrix between the query sample and the supporting samples;

[0125] S42: Generate a distance scaling factor matrix between the query sample and each category based on the inter-class correlation matrix. The calculation formula is:

[0126] ;

[0127] In the formula, It is a two-layer fully connected network. For the Sigmoid function;

[0128] S5: Use distance scaling factor to weight the distance between the query sample and each category of correction prototype to complete the interference category discrimination, and update the model parameters through backpropagation of loss function;

[0129] The specific process includes the following steps:

[0130] S51: Based on the distance scaling factor matrix, the distance between the query sample and the correction prototype center of each interference category is weighted and measured to obtain the weighted distance between the query sample and each interference category. The calculation formula is as follows:

[0131] ;

[0132] In the formula, Indicates the query sample in the distance scaling factor matrix With the Distance scaling factor between classes The distance is Euclidean.

[0133] S52: Calculate the predicted probability of the query sample belonging to each interference category based on the weighted distance. The calculation formula is as follows:

[0134] ;

[0135] In the formula, Indicates query sample The true label;

[0136] S53: Construct a classification loss function based on the true label of the query sample and the predicted probability. The calculation formula is as follows:

[0137] ;

[0138] S6: The network parameters use the Adam optimizer during the model training phase, with an initial learning rate set to 0.0001. The trained model is then used to classify and identify radar active interference samples in the test task, outputting the corresponding interference type results.

[0139] Experimental Results and Creativity Verification

[0140] This embodiment 1 conducts simulation experiments under the 10-way 5-shot and 10-way 10-shot paradigms, and compares them with existing small-sample radar active interference identification methods (prototype network PN, relational network RN). The experimental results are shown in Table 5, which shows the radar active interference identification results under different methods.

[0141] Table 5. Radar active interference identification results under different methods:

[0142] Experimental results show that the recognition accuracy of the method of the present invention is significantly higher than that of the existing comparative methods in both the 10-way 5-shot and 10-way 10-shot paradigms. Specifically, the accuracy is improved by 14.55% compared with the RN method in the 10-way 5-shot paradigm and by 12.69% compared with the RN method in the 10-way 10-shot paradigm. This fully demonstrates that the present invention effectively solves the shortcomings of the prior art through the synergistic improvement of prototype correction and distance perception guidance, significantly improves the recognition performance under small sample conditions, has outstanding substantive features and significant progress, and meets the requirements of patent inventiveness.

[0143] This embodiment 2 conducts a 5-way 5-shot paradigm simulation experiment for complex scenarios with unknown interference, and compares it with existing small-sample radar active interference identification methods (prototype network PN, relational network RN). The experimental results are shown in Table 6, which shows the identification results of different methods in unknown interference scenarios.

[0144] Table 6. Recognition results of different methods in unknown interference scenarios:

[0145] Experimental results show that the recognition accuracy of the present invention is significantly higher than that of existing comparative methods. Compared with the PN and RN methods, the average recognition accuracy of the present invention is improved by 5.39% and 9.33% respectively under known interference conditions; and by 5.71% and 10.70% respectively under unknown interference conditions. This fully demonstrates that the present invention, through the synergistic improvement of prototype correction and distance-aware guidance, effectively enhances the model's ability to recognize unknown interference without reducing the recognition performance of known interference. This compensates for the shortcomings of existing technologies in generalization performance under unknown interference, possessing outstanding substantive features and significant progress, and meeting the requirements of patent inventiveness.

[0146] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for identifying small-sample radar active interference, characterized in that: Includes the following steps: S1. Construct a radar active interference simulation sample set, and perform short-time Fourier transform processing on the sample set. Divide the samples into support set and query set according to the small sample learning method to generate tasks for model training and testing. S2. Input the support set and query set samples into the multi-scale asymmetric feature embedding network to extract the multi-scale and directional features of radar active interference and obtain the corresponding feature vector. S3. Input the feature vectors of the support set samples into the improved prototype network, and correct the prototype center of each interference category through the prototype correction method to obtain the corrected category prototype. S4. Based on the feature vectors of the query set samples and the support set samples, use the distance-aware module to generate the distance scaling coefficient between the query samples and each category; S5. Use the distance scaling factor to weight the distance between the query sample and each category of correction prototype to complete the interference category discrimination, and update the model parameters through backpropagation of the loss function until the model converges. S6. Use the converged model to classify and identify the radar active interference samples in the test task, and output the corresponding interference type results.

2. The radar active interference small sample identification method according to claim 1, characterized in that: The few-shot learning method is the N-way K-shot paradigm, where N is the number of interference categories in each task, K is the number of samples of each category in the support set, and the value of K ranges from 5 to 10, which meets the application requirements of few-shot scenarios.

3. The radar active interference small sample identification method according to claim 2, characterized in that: The multi-scale asymmetric feature embedding network adopts a structure that combines multi-scale convolution with directional attention enhancement, and simultaneously captures multi-scale detailed features and directional features of radar active interference signals to solve the problems of incomplete single-scale feature extraction and insufficient feature discrimination.

4. The radar active interference small sample identification method according to claim 3, characterized in that: The prototype correction method includes the following steps: S41. Calculate the initial prototype center for each interference category; S42. Remove anomalous feature samples from the support set through anomaly sample detection; S43. Based on valid samples, the prototype center is revised to make the category prototype more closely match the true feature distribution.

5. The radar active interference small sample identification method according to claim 4, characterized in that: The abnormal sample detection utilizes the intra-class correlation matrix to quickly identify and remove abnormal samples from the support set, thus avoiding interference from abnormal samples to the prototype center.

6. The radar active interference small sample identification method according to claim 5, characterized in that: The distance-aware module dynamically generates a distance scaling factor by calculating the feature correlation between the query sample and the prototype of each category, so that the distance metric can adaptively match the degree of association between different samples and categories.

7. The radar active interference small sample identification method according to claim 6, characterized in that: The weighted distance metric adjusts the original distance using a distance scaling factor, so that query samples with strong relevance to the category prototype receive the minimum weighted distance, thereby improving the accuracy of category discrimination.

8. The radar active interference small sample identification method according to claim 7, characterized in that: The loss function is the cross-entropy loss function, which updates all parameters of the multi-scale asymmetric feature embedding network, the improved prototype network, and the distance-aware module through backpropagation until the model converges.

Images