Small sample based multi-modal radar active deception jamming identification method

Abstract

The application discloses a kind of based on small sample's multi-modal radar active deception jamming identification method, belong to radar active deception jamming identification technical field.The application includes: the signal processing is carried out to the radar active deception jamming signal obtained, to extract feature parameter modal dataset and time-frequency image dataset, the former is carried out Gaussian randomization expansion, the latter is expanded using handwriting character dataset Omniglot dataset, again it is divided into support set and query set;Build the prototype network based on multi-modal fusion, according to support set respectively calculate the prototype of feature parameter modal and time-frequency image modal, the prototype of feature parameter modal and time-frequency image modal is fused, and the multi-modal fusion prototype is obtained;According to the sample of query set of multi-modal fusion prototype, classification is carried out to obtain the identification model of radar active deception jamming.The application can obtain enough interference characteristics under the condition of small sample, and good radar active deception jamming identification effect is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar active deception jamming recognition, and in particular to a multi-modal radar active deception jamming recognition method based on small samples. BACKGROUND

[0002] Radar deception jamming recognition is one of the core technologies of radar anti-jamming. After distinguishing the jamming types, the radar can take corresponding anti-jamming measures to suppress such jamming, so as to achieve better anti-jamming effect. Traditional jamming recognition mostly only targets single jamming mode, and performs feature extraction in a certain dimension for jamming recognition. However, jamming signal feature extraction often depends on the subjective design of researchers, and it is difficult to design and extract deep abstract features, thereby affecting the jamming recognition performance.

[0003] In recent years, due to the excellent automatic learning and deep mining data feature ability of machine learning, many scholars have applied it to the field of radar deception jamming recognition, and used neural networks (Neural Networks, NN), convolutional neural networks (Convolutional Neural Network, CNN), Siamese networks (Siamese Network) and the like for jamming recognition, which has achieved good performance in classification accuracy and universal recognition. However, the recognition method based on machine learning or deep learning is greatly affected by the number of training samples, and requires a complete training set and sufficient samples to support the training of the network. Under the condition of limited training samples, it is difficult to capture data features and achieve the expected recognition effect. In actual application scenarios, it is difficult to capture enemy jamming information and build a complete training set, and the existing method is difficult to complete the recognition of radar active deception jamming under small samples. SUMMARY

[0004] The present application proposes a multi-modal radar active deception jamming recognition method based on small samples to effectively improve the radar active deception jamming recognition effect under small samples.

[0005] The technical scheme adopted by the present application is:

[0006] The multi-modal radar active deception jamming recognition method based on small samples comprises the following steps:

[0007] Step S1, signal processing is performed on the acquired radar active deception jamming signal, time domain, frequency domain and time-frequency domain feature parameter modal information is extracted, and feature parameter modal data set is obtained; and Short-time Fourier Transfer (STFT) is used to extract time-frequency image modal information of the radar active deception jamming signal, and time-frequency image data set is obtained;

[0008] Step S2, image enhancement and morphological closing operation noise reduction processing are performed on the time-frequency image data set obtained in step S1;

[0009] Step S3, the feature parameter modal data set obtained in step S1 is expanded in a Gaussian randomization manner, and the time-frequency image data set obtained in step S1 is expanded using a handwritten character data set Omniglot data set;

[0010] Step S4, the feature parameter modal data set expanded in step S3 is divided into a support set and a query set;

[0011] Step S5, a prototype network based on multi-modal fusion is built, prototypes of the feature parameter modal and the time-frequency image modal are calculated according to the support set respectively, feature layer fusion is performed, and a multi-modal fusion prototype is obtained;

[0012] Step S6, samples in the query set are classified according to the multi-modal fusion prototype obtained in step S5, a loss is calculated, the prototype network is optimized, and finally a recognition model of radar active deception jamming is obtained to obtain a recognition result of the radar active deception jamming signal of the target object.

[0013] Specifically, the step S1 includes the following sub-steps:

[0014] Step S11, according to the characteristics of radar active deception jamming, nine features including time domain moment skewness, time domain moment kurtosis, time domain envelope fluctuation, normalized amplitude standard deviation, frequency domain moment skewness, frequency domain moment kurtosis, frequency domain envelope fluctuation, normalized amplitude spectrum maximum, and scale barycenter are extracted from the jamming signal for jamming recognition.

[0015] Step S12, a time-frequency image of the radar deception jamming signal is obtained by using STFT transformation, and the image is subjected to normalized gray processing;

[0016] Specifically, the step S2 includes the following sub-steps:

[0017] Step S21, the time-frequency image obtained in step S1 is subjected to image enhancement processing. In order to facilitate image denoising processing in the later stage, the image is subjected to linear gray scale transformation to expand the gray scale value between the noise and the signal to realize gray scale image enhancement. Assuming that the original gray scale range of an image f(x, y) is between [a, b], and the gray scale range of the image g(x, y) after linear transformation is between [c, d], then the gray scale value k of any pixel point in the image f(x, y) is transformed to obtain the gray scale value l of the pixel point in the image g(x, y) after transformation.

[0018]

[0019] Wherein, μ is an enhancement coefficient.

[0020] Step S22: Perform morphological closing operation on the time-frequency image obtained in step S21 to reduce noise.

[0021] Specifically, step S3 further includes the following sub-steps:

[0022] Step S31: Expand the feature parameters obtained in step S11. To accommodate the increasing number of image features, the feature parameters need to be expanded through training. The expansion method is Gaussian randomized sample expansion, that is, repeatedly adding samples with a mean of 0 and a variance of σ to each knowledge vector. c 2 Independent and identically distributed Gaussian noise is used to obtain training samples for the feature parameters. It should be noted that the Gaussian randomization process can improve robustness while expanding the training sample space and preventing overfitting of the central network.

[0023] Step S32: Expand the time-frequency image sample set obtained in step S2. For image sample expansion, a simple Gaussian expansion method for one-dimensional data cannot be used. This is mainly because although Gaussian randomization can expand the samples and enhance robustness, the essential changes in the features of two-dimensional data are minimal, and adding noise results in virtually no change. Therefore, for the expansion of the time-frequency image portion, we choose to train it together with the Omniglot dataset.

[0024] The Omniglot dataset consists of 1,623 handwritten characters from 50 different languages, with 20 different handwriting variations for each character. This results in a small sample dataset of handwritten characters, with a large number of categories (1,623) but very few samples per category (20 characters). Character recognition in this dataset relies primarily on simple line features, similar to the time-frequency images used in active radar deception jamming. This dataset can be used to achieve excellent training results.

[0025] Specifically, step S4 includes the following sub-steps:

[0026] For each category in the expanded dataset from step S3, take N respectively. s N samples are used as the support set, and N samples are selected from the remaining samples. q N is used as the query set, and N is taken during training. s :N q =1.

[0027] Specifically, step S5 further includes:

[0028] A prototype network based on multimodal fusion was built. The feature parameter set and the time-frequency image set were respectively processed by two convolutional neural networks (the first feature extraction network and the second feature extraction network) to extract and fuse feature vectors to obtain the prototype under multimodal conditions.

[0029] The first feature extraction network is used to extract feature vectors of time-frequency image modes. It includes six two-dimensional convolutional layers and converts the feature maps extracted by the six two-dimensional convolutional layers into one-dimensional vectors to obtain the prototype of the time-frequency image mode and output it. The second feature extraction network is used to extract feature vectors of feature parameter modes. It includes four two-dimensional convolutional layers and converts the feature maps extracted by the six two-dimensional convolutional layers into one-dimensional vectors to obtain the prototype of the feature parameter mode and output it. Then, the prototype of the time-frequency image mode and the prototype of the feature parameter mode are concatenated by a concatenation layer to obtain the fused feature vector of each sample, so as to obtain the multi-modal fusion prototype of samples of the same signal type.

[0030] Let S be a small set of labeled datasets, represented as:

[0031]

[0032] Where x1 is the D1-dimensional feature parameter set input data, x2 is the D2-dimensional image set input data, and y represents the label of the data, i.e., the category corresponding to x. Using S... k Let K represent the data set with category k, and the total number of categories be denoted as K.

[0033] The network f acts as a mapping function, mapping the original D-dimensional features to a P-dimensional space.

[0034]

[0035] For the i-th object, the feature parameters and image features extracted from the single-modal model are denoted as f1 and f2, respectively. i , The two features are combined to obtain the fused feature.

[0036]

[0037] Select S k The set, by averaging all P-dimensional vectors within it, yields the k-class multimodal fusion prototype c. k :

[0038]

[0039] Specifically, step S6 further includes:

[0040] After obtaining the prototype of each class, the probability of each sample corresponding to each class is calculated using the query set:

[0041]

[0042] Where d(·) represents the weighted squared Euclidean distance, its expression is:

[0043]

[0044]

[0045]

[0046] in, and Representing different modal components in the prototype, the weighted average of the basic Euclidean distance can effectively improve the confusion caused by the failure of a single feature under low signal-to-noise ratio.

[0047] Then, the loss can be calculated using a general classification loss function, and the parameters can be optimized. The main optimization objective during training is to find a suitable mapping function f. φ That is, the appropriate f1 and f2 make the prototype C k This allows for a better representation of the features of each image category. Assuming the true class of a sample is k', the loss function can be defined as:

[0048] J(φ)=-log(p φ (y=k'|x))

[0049] Here, log represents the logarithmic operation.

[0050] The technical solution provided by this invention brings at least the following beneficial effects:

[0051] This invention uses noise reduction processing of time-frequency images and expands the original dataset using the Omniglot dataset. It also employs a prototype network and multimodal fusion approach to identify radar active deception interference. This allows for the acquisition of sufficient interference features under small sample conditions, achieving good radar active deception interference identification results under such conditions. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 The flowchart illustrates the processing of a multi-modal radar active deception interference identification method based on small samples, as provided in this embodiment of the invention.

[0054] Figure 2 This is a schematic diagram of the network structure of the multimodal recognition network based on small samples according to the present invention.

[0055] Figure 3 This is the confusion matrix diagram used in the embodiments of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0057] like Figure 1 As shown in the figure, the multi-modal radar active deception interference identification method based on small samples provided in this embodiment of the invention specifically includes the following steps:

[0058] Step S1: Process the acquired radar active deception jamming signal to obtain the characteristic parameter mode information in the time domain, frequency domain, and time-frequency domain, and use Short-time Fourier Transform (STFT) to obtain the time-frequency image mode information of the signal.

[0059] As one possible implementation, step S1 specifically includes the following steps:

[0060] Step S11: Based on the characteristics of active radar deception jamming, use the jamming signal to extract ten features in total, including time-domain moment skewness, time-domain moment kurtosis, time-domain envelope variability, normalized amplitude standard deviation, frequency-domain moment skewness, frequency-domain moment kurtosis, frequency-domain envelope variability, normalized amplitude spectrum maximum value, and scale centroid, to identify the jamming.

[0061] 1) Time-domain moment skewness:

[0062]

[0063] In this context, the signal amplitude vector is denoted by X, where X = [x(1), x(2), ..., x(N)], x(i) is the signal amplitude at time i (i = 1, 2, ..., N), N represents the number of sampling points, i.e., the signal length, μ and σ are the mean and standard deviation of X, respectively, and E(·) represents the calculation of the first moment. The time-domain moment skewness represents the degree of asymmetry in the signal distribution in the time domain.

[0064] 2) Time-domain moment kurtosis:

[0065]

[0066] The kurtosis of the time-domain moment indicates the steepness of the signal's time-domain distribution.

[0067] 3) Temporal envelope variability:

[0068]

[0069] Envelope undulation characterizes the degree of change in the signal envelope.

[0070] 4) Normalized amplitude standard deviation:

[0071]

[0072] Where, N L Indicates signal length. x cn Here, x(i) represents the zero-center normalized amplitude, and x(i) represents the signal amplitude. Zero-center normalization can eliminate the influence of signal gain on characteristic parameters. The extracted characteristic parameters can better reflect the actual change law of the signal and can measure different interference signals on the same scale.

[0073] 5) Frequency domain moment skewness:

[0074]

[0075] Where, μ ω σ ω The signal power spectrum X ω The mean and standard deviation. Similar to the time-domain moment skewness, the frequency-domain moment skewness coefficient reflects the distribution of the power spectrum.

[0076] 6) Frequency domain moment kurtosis:

[0077]

[0078] Frequency domain moment kurtosis reflects the steepness of the signal power spectrum distribution; the steeper the distribution, the larger the value of the frequency domain moment kurtosis.

[0079] 7) Frequency domain envelope ripple:

[0080]

[0081] The spectral envelope undulation characterizes the degree of change in the signal power spectrum envelope; the greater the degree of change, the larger the value of the spectral envelope undulation.

[0082] 8) The maximum value of the normalized amplitude spectrum:

[0083]

[0084] Here, max(·) represents taking the maximum value, and DFT(·) represents the Discrete Fourier Transform. The maximum value of the normalized amplitude spectrum can be used to characterize the degree of fluctuation of the signal.

[0085] 9) Center of gravity:

[0086] The scale centroid of the time-frequency image of the interference is defined as:

[0087] C = [x c ,y c ]

[0088]

[0089]

[0090] Where A(i,j) represents the pixel of the time-frequency image, and M and N represent the number of rows and columns of the time-frequency image, respectively.

[0091] In this embodiment of the invention, the above nine feature parameter modes are used, wherein the scale centroid includes two dimensions: x c and y c The remaining features are one-dimensional, with a total of 10 feature parameters used for interference identification.

[0092] Step S12: Obtain the time-frequency image of the radar deception jamming signal based on STFT transformation, and perform normalized grayscale processing on the image.

[0093] Step S2: Perform image enhancement and morphological closing operation noise reduction on the time-frequency images obtained in step S1 to obtain an expanded time-frequency image sample set.

[0094] As one possible implementation, step S2 specifically involves:

[0095] Step S21: Perform image enhancement processing on the time-frequency image obtained in step S1.

[0096] To facilitate subsequent image denoising, a linear grayscale transformation is first performed on the image to widen the grayscale value gap between noise and signal, thereby enhancing the grayscale image. Assuming the original grayscale range of an image f(x,y) is between [a,b], and the grayscale range of the transformed image g(x,y) is between [c,d], then the grayscale value k of any pixel in image f(x,y) can be represented by the grayscale value l of the transformed pixel in image g(x,y) as follows:

[0097]

[0098] Where μ is the enhancement coefficient.

[0099] Step S22: Perform morphological closing operation on the enhanced time-frequency image from step S21 to reduce noise.

[0100] The closing operation is performed by first expanding and then eroding. The expansion of set A by set B can be represented as follows: Its definition is:

[0101]

[0102] Where A is called the input image, B is called the structuring element, and A cThis is the complement of A. In digital image processing, set B is also known as a structuring element.

[0103] Set A is corrupted by set B, denoted as Its definition is:

[0104]

[0105] Where A is called the input image, and B is called the structuring element. It consists of all points x that are translated x from B but are still contained within A.

[0106] The closing operation of an image can be represented as:

[0107]

[0108] Step S3: Expand the dataset of feature parameter modes using Gaussian randomization, and expand the time-frequency image dataset of image modes using the Omniglot dataset.

[0109] As one possible implementation, step S3 specifically includes:

[0110] Step S31: Expand the feature parameters obtained in step S11.

[0111] To accommodate the increasing number of image features, the feature parameters need to be augmented during training. This augmentation is achieved through Gaussian randomized sample augmentation, which involves repeatedly adding samples with a mean of 0 and a variance of σ to each knowledge vector. c 2 Independent and identically distributed Gaussian noise is used to obtain training samples for the feature parameters. It should be noted that the Gaussian randomization process can improve robustness while expanding the training sample space and preventing overfitting of the central network.

[0112] Step S32: Expand the time-frequency image sample set obtained in step S2.

[0113] For image sample augmentation, a simple Gaussian augmentation method for one-dimensional data cannot be used. This is mainly because although Gaussian randomization can augment the samples and enhance robustness, the essential changes in the features of two-dimensional data are minimal, and adding noise results in virtually no change. Therefore, for the augmentation of the time-frequency image portion, we choose to train the Omniglot dataset together with the time-frequency images.

[0114] The Omniglot dataset consists of 1,623 handwritten characters from 50 different languages, with 20 different handwriting variations for each character. This results in a small sample dataset of handwritten characters, with a large number of categories (1,623) but very few samples per category (20 characters). Character recognition in this dataset relies primarily on simple line features, similar to the time-frequency images used in active radar deception jamming. This dataset can be used to achieve excellent training results.

[0115] Step S4: Divide the expanded dataset (feature parameter dataset and time-frequency image dataset) from step S3 into a support set and a query set;

[0116] As one possible implementation, step S4 specifically involves taking N values ​​for each category in the expanded dataset from step S3. s (Preset value, empirical value) samples are used as the support set, and N samples are selected from the remaining samples. q N is used as the query set, and N is taken during training. s :N q =1.

[0117] Step S5: Build a prototype network based on multimodal fusion. Calculate the prototypes of feature parameter modes and time-frequency image modes according to the support set, and perform feature layer fusion to obtain the multimodal fusion prototype.

[0118] The prototype network includes a first feature extraction network f1 (based on a convolutional neural network, used to extract the prototype of the time-frequency image modality, i.e., the feature vector of the time-frequency image) and a second feature extraction network f2 (used to extract the prototype of the feature parameter modality, i.e., the feature vector of the feature parameter), and a feature fusion layer, used to fuse the feature vectors output by the first feature extraction network f1 and the second feature extraction network f2 to output a multimodal prototype, such as... Figure 2 As shown.

[0119] As one possible implementation, step S5 specifically includes:

[0120] A network f1 is established to be responsible for the prototype calculation of time-frequency image modes. It contains an input layer and six Conv2d layers (two-dimensional convolutional blocks). The input data matrix of network f1 is N1×1×112×112, and the output matrix is ​​N1×M1, where N1 is the number of samples, M1 is the number of feature points, and 112×112 represents the resolution of the time-frequency image. Ten categories are selected from all categories for training each time. In this embodiment, M1 is set to 64.

[0121] A network f2 is established to handle the prototype calculation of the feature parameter modalities. It contains an input layer and four Conv2d layers. The input data matrix of network f2 is N2×1×1×10, where 10 represents the dimension of the feature parameters. The output matrix is ​​N2×M2, where N2 is the number of samples and M2 is the number of feature points. Ten categories are selected from all categories for training each time; in this embodiment, M2 is also set to 64. It should be noted that the values ​​of M1 and M2 can be the same or different. That is, when network f1 extracts the feature vector of the time-frequency image, it converts it into a one-dimensional vector before outputting it. Similarly, when network f2 extracts the feature vector of the feature parameters, it also converts it into a one-dimensional vector before outputting it. This achieves the concatenation of the feature vectors of the two modalities.

[0122] Let S be a small set of labeled datasets, represented as:

[0123]

[0124] Where x1 is the D1-dimensional feature parameter set input data, x2 is the D2-dimensional time-frequency image set input data, and y represents the label of the data, i.e., the category corresponding to x. Using S... k Let K represent the data set with category k, and the total number of categories be denoted as K.

[0125] Network f acts as a mapping function, mapping the original D-dimensional features to a P-dimensional space:

[0126] For the i-th object, the feature parameters and image features extracted from the single-modal model are denoted as f1 and f2, respectively. i , The two features are combined to obtain the fused feature.

[0127] Select S k The set, by averaging all P-dimensional vectors within it, yields the k-class multimodal fusion prototype c. k :

[0128] Step S6: Based on the multimodal fusion prototype obtained in step S5, classify the samples in the query set, calculate the loss, optimize the network, and finally obtain the identification model of radar active deception jamming to obtain the identification result of radar active deception jamming signal of the target object.

[0129] As one possible implementation, step S6 also includes:

[0130] After obtaining the prototype of each class, the probability of each sample corresponding to each class is calculated using the query set:

[0131]

[0132] Where d(·) represents the weighted squared Euclidean distance, its expression is:

[0133]

[0134]

[0135]

[0136] in, and Representing different modal components in the prototype, the weighted average of the basic Euclidean distance can effectively improve the confusion caused by the failure of a single feature under low signal-to-noise ratio.

[0137] Then, the loss can be calculated using a general classification loss function, and the parameters can be optimized. The main optimization objective during training is to find a suitable mapping function f. φ That is, the appropriate f1 and f2 make the prototype C k This allows for a better representation of the features of each image category. Assuming the true class of a sample is k', then the loss function can be defined as follows:

[0138] J(φ)=-log(p φ (y=k'|x))

[0139] Here, log represents the logarithmic operation.

[0140] In this invention, based on the identification model of radar active deception jamming, interference identification of the current radar active deception jamming signal to be identified is performed. First, its time-frequency image is acquired, and then the time-frequency image is processed by image enhancement and morphological closing operation noise reduction, and then input into the identification model to obtain its image features; and the feature parameters of the radar active deception jamming signal to be identified are input into the identification model to obtain its feature parameters; the image features and feature parameters are spliced ​​to obtain the multimodal fusion prototype to be identified, and it is matched with the multimodal fusion prototypes in the jamming signal library. The Euclidean distance between the multimodal fusion prototype to be identified and each multimodal fusion prototype in the library is calculated. If the minimum Euclidean distance meets the specified condition, the jamming signal type of the signal to be identified is determined based on the jamming signal type of the minimum Euclidean distance; otherwise, it indicates that the current jamming signal to be identified does not match the library, and it is saved. When a certain number of signals are met, they are clustered and new signal categories are determined based on the clustering structure and stored in the library.

[0141] In this invention example, ten active radar deception jamming signals were obtained through simulation, including Range Multi-false-target Jamming (RMTJ), Velocity Multi-false-target Jamming (VMTJ), Range-Velocity Multi-false-target Jamming (RVMTJ), Chopping and Interleaving Jamming (CIJ), Interrupted Sampling Repeater Jamming (ISRJ), Smeared Spectrum Jamming (SSJ), Convolutional Smart Noise Jamming (CSNJ), Range Gate Pull-off Jamming (RGPJ), Velocity Gate Pull-off Jamming (VGPJ), and Range-Velocity Gate Pull-off Jamming (RVGPJ). The parameter settings are shown in Table 1.

[0142] In the specific implementation process, the processor was set up with an Nvidia GeForce RTX 2060 GPU; the software platform was a simulation experimental platform using PyCharm.

[0143] Table 1 Parameter Settings

[0144]

[0145]

[0146] The optimal confusion matrix for multimodal recognition networks based on small samples is as follows: Figure 3 As shown. Observation Figure 3 It can be seen that the multimodal approach achieved good recognition results for each type of interference, with an overall recognition rate of 97.35%.

[0147] This invention expands the time-frequency image dataset by transferring the Omniglot dataset, and integrates the features of the feature parameters and the features of the time-frequency images to achieve feature complementarity. Finally, it trains multimodal features with a prototype network to realize multimodal radar active deception interference identification based on a small sample size.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0149] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A multi-mode radar active deception interference identification method based on small samples, characterized in that, Includes the following steps: Step S1: Perform signal processing on the acquired radar active deception jamming signal, extract the time-domain, frequency-domain and time-frequency domain feature parameter mode information to obtain the feature parameter mode dataset; and use short-time Fourier transform to extract the time-frequency image mode information of the radar active deception jamming signal to obtain the time-frequency image dataset. Step S2: Perform image enhancement and morphological closing operation noise reduction on the time-frequency image dataset obtained in step S1. Step S3: Expand the feature parameter modality dataset obtained in step S1 using Gaussian randomization, and expand the time-frequency image dataset obtained in step S1 using the handwritten character dataset Omniglot dataset. Step S4: Divide the feature parameter modality dataset processed in step S3 into a support set and a query set; Step S5: Build a prototype network based on multimodal fusion. Calculate the prototypes of feature parameter modes and time-frequency image modes according to the support set. Perform feature fusion on the prototypes of feature parameter modes and time-frequency image modes to obtain the multimodal fusion prototype. Step S6: Classify the samples in the query set according to the multimodal fusion prototype, calculate the classification loss to optimize the network parameters of the prototype network, and finally obtain the identification model of radar active deception jamming to obtain the identification result of radar active deception jamming signal of the target object. In step S5, the prototype network includes a first feature extraction network, a second feature extraction network, and a splicing layer; The first feature extraction network is used to extract the feature vector of the time-frequency image mode. It includes six two-dimensional convolutional layers and converts the feature maps extracted by the six two-dimensional convolutional layers into one-dimensional vectors to obtain the prototype of the time-frequency image mode and output it. The second feature extraction network is used to extract the feature vector of the feature parameter mode. It includes four two-dimensional convolutional layers and converts the feature map extracted by the six two-dimensional convolutional layers into a one-dimensional vector to obtain the prototype of the feature parameter mode and output it. Then, the prototypes of the time-frequency image modes and the feature parameter modes are spliced ​​together by the splicing layer to obtain the fusion feature vectors of each sample, so as to obtain the multi-modal fusion prototypes of samples of the same signal type.

2. The method as described in claim 1, characterized in that, In step S1, the extracted time-domain, frequency-domain, and time-frequency domain feature parameter modal information includes: time-domain moment skewness, time-domain moment kurtosis, time-domain envelope variability, normalized amplitude standard deviation, frequency-domain moment skewness, frequency-domain moment kurtosis, frequency-domain envelope variability, normalized amplitude spectrum maximum value, and scale centroid.

3. The method as described in claim 1, characterized in that, The extraction of time-frequency image modal information of radar active deception jamming signals using short-time Fourier transform specifically includes: extracting time-frequency images of radar active deception jamming signals using short-time Fourier transform, and performing normalized grayscale processing on the extracted video images.

4. The method as described in claim 1, characterized in that, In step S2, the image enhancement processing of the time-frequency image dataset obtained in step S1 is specifically as follows: Define image The original grayscale range is in The image after linear transformation The grayscale range is in Between, the image grayscale value of any pixel The image obtained after transformation For pixel grayscale values It can be represented as: in, This is the enhancement coefficient.

5. The method as described in claim 1, characterized in that, In step S4, the ratio of the number of support sets to the number of query sets is 1.

6. The method according to any one of claims 1 to 5, characterized in that, In step S5, the multimodal fusion prototype is specifically as follows: definition Indicates the signal category as The data set, the total number of categories is denoted as ; and definition Indicates the first The fusion feature vector of each sample; For data sets The fused feature vector of all samples The average value is Multimodal fusion prototype of categories ; Calculate the probability of each sample corresponding to each class using the query set: in, Represents a sample of the query set The fused feature vector, Indicates the currently identified category. Indicates the true category, A multimodal fusion prototype representing the true category. The weighted squared Euclidean distance is expressed as follows: in, and This represents the intermediate computational cost for different modes. , These represent samples of the query set. The eigenvectors of the time-frequency image mode and the feature parameter mode, and They represent Multimodal fusion prototype of categories The time-frequency image modal components and feature parameter modal components; Furthermore, in step S6, the classification loss function is: .

Images