A synthetic pseudo-text enhanced small sample radar jamming recognition method and system

Abstract

The application discloses a kind of synthetic pseudo-text enhanced small sample radar jamming identification method and system, method includes: for radar jamming type, generate a group of k classification n sample support set;Utilize generative large language model to generate the pseudo-text corresponding to each interference type and filter to obtain pseudo-text dataset;According to support set, obtain sample image, according to the sample image and pseudo-text dataset of each interference type, generate image-text pair by the way of random sampling, constitute training set;According to training set, the parameter of contrast language-image pretraining CLIP model is fine-tuned using total loss function;The interference signal to be identified is converted into image, and the encoder of the fine-tuned CLIP model is used to extract image features and support set corresponding text features, and the similarity between image features and each category text features is calculated, the text type with maximum similarity is taken as output interference type.The application can effectively improve the interference style recognition performance under small sample condition.

Description

Technical Field

[0001] This invention relates to the field of radar interference identification, specifically to a method and system for identifying small-sample radar interference using synthetic pseudo-text enhancement. Background Technology

[0002] Electromagnetic interference (EMI) severely impacts radar detection. Common anti-interference measures include selecting appropriate waveforms and signal processing methods based on interference patterns. The performance of interference pattern recognition plays a crucial role in subsequent anti-interference efforts; therefore, interference identification has received widespread attention.

[0003] In real-world electronic warfare, radar jamming is unpredictable and inaccessible, making it difficult to collect sufficient datasets. Typically, only a few sample frames of different jamming types are available, and the recognition model is trained based on these frames. In this case, the problem of jamming pattern recognition transforms into a few-shot learning problem. The scarcity of training data severely reduces recognition accuracy. Existing few-shot classification methods mainly focus on computer vision, with limited research in the field of few-shot radar jamming recognition. Compared to images, radar signals are more diverse and complex, requiring models with better generalization capabilities. Summary of the Invention

[0004] Purpose of the invention: To address the problem of interference pattern identification under the scarcity of existing radar interference sample data, this invention proposes a method that can identify interference types even under small sample conditions, regardless of the number of samples.

[0005] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] Firstly, a method for identifying few-sample radar interference enhanced with synthetic pseudo-text includes the following steps:

[0007] For each type of radar interference, a support set of k-class n samples is generated. Using a generative large language model, based on the support set and a pre-defined list of prompt templates for querying interference types, a prompt template and temperature parameters are randomly provided to generate pseudo-text for each type of interference. The pseudo-text is then filtered to obtain a pseudo-text dataset for each type of interference.

[0008] Based on the support set, the interference signal is converted into sample image data using the time-frequency transformation method; based on the sample image data and pseudo-text dataset of each type of interference, image-text pairs are generated by random sampling, resulting in a total of k-classified image-text pair datasets, which constitute the training set.

[0009] The training set is input into the image encoder and text encoder of the contrastive language-image pre-trained CLIP model. Image feature extraction and text feature extraction are performed respectively, and the feature similarity between the image and the text is calculated. Cross-entropy loss is calculated based on the similarity, and triplet loss of image features and triplet loss of text features are calculated to form the total loss function. The model parameters are backpropagated based on the total loss function and the parameters are updated until the required number of training rounds are reached, so as to achieve fine-tuning of the CLIP model.

[0010] The interference signal to be identified is converted into time-frequency domain image data using the time-frequency transformation method. The time-frequency domain image data and the interference type category name are input into the image encoder and text encoder of the fine-tuned CLIP model, respectively. Image features and category text features are extracted, and the similarity between the image features and each category text feature is calculated. The text with the highest similarity is taken as the best matching description, and this text type is the output interference type.

[0011] Preferably, the generative large language model includes Llama-3-8b-instruct, Gemma-7b-it, Openchat-7b, and Mistral-7b-instruct, with the temperature parameter being a random floating-point number from 0 to 1.

[0012] Preferably, a generative large language model is used to generate pseudo-text corresponding to each type of interference by randomly providing a prompt template and temperature parameters based on the support set and a pre-defined list of prompt templates for interference type queries, including:

[0013] Based on the support set, a list of prompt templates for querying interference types is determined, wherein the prompt template is a prompt statement that asks for the characteristics of a specified interference type;

[0014] Based on the selected generative large language models, a list of large language models is generated;

[0015] Execute the query. For each query process, randomly select a prompt template and a large language model from the prompt template list and the large language model list based on the temperature parameter to obtain the corresponding pseudo-text.

[0016] Preferably, the pseudo-text is filtered to obtain pseudo-text data for each type of interference, including:

[0017] The query response text is split using a specified delimiter to obtain short texts. For the short texts, sentences with lengths outside the specified range, irrelevant texts, and existing texts of the same type are removed to obtain a high-quality pseudo-text dataset for each type of interference.

[0018] Preferably, the feature similarity between an image and text is represented as follows:

[0019]

[0020] represents the features of the image and text respectively, b represents the batch number, and d represents the feature dimension.

[0021] Preferably, the total loss function is expressed as:

[0022]

[0023] TripletMarginLoss represents the triplet loss, and CrossEntropyLoss represents the cross-entropy loss.

[0024] Preferably, the time-frequency transformation method employs short-time Fourier transform.

[0025] Secondly, a synthetic pseudo-text enhanced few-sample radar interference identification system includes:

[0026] The pseudo-text data generation module is used to generate a support set of k-class n samples for radar interference types. Using a generative large language model, based on the support set and a pre-set list of prompt templates for querying interference types, it randomly provides prompt templates and temperature parameters to generate pseudo-text corresponding to each interference type. The pseudo-text is then filtered to obtain a pseudo-text dataset for each interference type.

[0027] The training set generation module is used to convert interference signals into sample image data using the time-frequency transformation method based on the support set; based on the sample image data and pseudo-text dataset of each type of interference, image-text pairs are generated by random sampling, generating a total of k-classified image-text pair datasets to form the training set;

[0028] The model training module is used to input the training set into the image encoder and text encoder of the contrastive language-image pre-trained CLIP model, extract image features and text features respectively, calculate the feature similarity between the image and the text, calculate the cross-entropy loss based on the similarity, and calculate the triplet loss of image features and the triplet loss of text features to form the total loss function. The model parameters are backpropagated based on the total loss function and updated until the required number of training rounds are reached, thereby achieving fine-tuning of the CLIP model.

[0029] The interference identification module is used to convert the interference signal to be identified into time-frequency domain image data using the time-frequency transformation method. The time-frequency domain image data and the interference type category name are input into the image encoder and text encoder of the fine-tuned CLIP model, respectively. Image features and category text features are extracted, and the similarity between the image features and each category text feature is calculated. The text with the highest similarity is taken as the best matching description, and this text type is the output interference type.

[0030] Thirdly, a computer device includes: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the synthetic pseudo-text enhanced few-sample radar interference identification method as described in the first aspect of the invention.

[0031] Fourthly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the synthetic pseudo-text enhanced few-sample radar interference identification method as described above.

[0032] Beneficial Effects: Addressing the problem of interference pattern recognition under scarce existing radar interference sample data, this invention proposes a small-sample interference recognition method and system using a synthetic pseudo-text dataset. It generates pseudo-text datasets for each interference type by constructing a Generative Large Language Model (LLM), and performs time-frequency transformation on the interference signal to obtain time-frequency domain image data. Subsequently, the text and image data are processed by corresponding encoders to obtain image-to-text datasets. After fine-tuning the CLIP model, a reasoning-ready model is obtained. The signal to be recognized is converted into time-frequency domain image data, and features are extracted by the encoder along with the text for each potential category. The text with the highest similarity is calculated and selected to achieve interference type recognition. This invention achieves effective interference recognition even under conditions of scarce radar interference samples by generating pseudo-text datasets instead of scarce radar interference sample datasets. Simulation results show that this method achieves an average recognition accuracy of 93.32% with only 5 samples for each interference type. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method of the present invention;

[0034] Figure 2 Time-frequency domain plot of the COMB interference pattern;

[0035] Figure 3 Feature maps of the text dataset before and after fine-tuning;

[0036] Figure 4 The graph shows the test results in the form of a confusion matrix. Detailed Implementation

[0037] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0038] This invention is applicable to situations where interference samples are scarce, enabling the identification of interference patterns affecting radar signals. The overall flow of the identification method is shown in the attached figure. Figure 1 As shown, the specific steps include:

[0039] Step 1: Synthesize the pseudo-text dataset

[0040] This invention utilizes Generative Large Language Models (LLM) to generate a pseudo-text dataset for each type of jamming sample. s(t) represents the transmitted waveform. It is assumed that the jammer employs DRFM (Digital RF Memory) jamming technology, capable of intercepting and modulating radar signals. f c (·) represents the modulation function used by the DRFM jammer, where c = 1, 2, ..., C represents the type of interference. The received signal subjected to interference can be represented as:

[0041]

[0042] Where τ represents time delay and n represents Gaussian white noise.

[0043] Different types of interference manifest in different ways. Table 1 shows six classic DRFM interference formulas: C&I (Chopper & Intercept), COMB (Comb Spectrum Jamming), NCJ (Noise Convolution Jamming), FM (Frequency Modulation), SMSP (Smeared Spectrum Jamming), and ISRJ (Interrupted Sampling Repeater Jamming).

[0044] Table 1 Interference Formula

[0045]

[0046] For typical few-shot learning tasks, only a few samples are available, and this set serves as the support set. A few-shot learning task with a support set is called a kways and nshots task, where k represents the number of classes and n represents the number of samples in each class. This invention selects a set containing 6 types of perturbations, with 5 samples in each class, as the support set for training a 6ways 5shots task.

[0047] The support set contains six types of interference, as shown in Table 1. Interference is generated using the formulas in Table 1, and the parameters for generating the interference signal are shown in Table 2. Each class has five samples, making the support set 6-way 5-shots.

[0048] Table 2 Interference Parameter Settings

[0049]

[0050] This invention selects Llama-3-8b-instruct, Gemma-7b-it, Openchat-7b, and Mistral-7b-instruct as the LLMs used, with a temperature parameter of a random floating-point number between 0 and 1. For each type of interference query process, a prompt template and a large language model are randomly selected from the prompt template list and the large language model list based on the temperature parameter to obtain the corresponding pseudo-text, thus obtaining a text description for each type of interference. Then, the response text is split and filtered, removing sentences that are too long or too short, inaccurate descriptions, and irrelevant text.

[0051] Specifically, this invention first designs a set of prompt word templates, such as "Do you know something about {}", "Give me several features {}have", "Give me features of spectrum of {}", etc. Then, random sampling is performed on the interference type category names and the prompt word template set to form complete prompt words. These prompt words are then provided to a randomly selected Large Language Model (LLM). Based on the LLM's response, pseudo-text data is generated. Next, the response text generated by the LLM model is split and cleaned to obtain high-quality pseudo-text data. The splitting operation includes segmenting the text using delimiters such as commas, periods, etc., to obtain short texts. The cleaning operation includes removing texts with fewer than 10 or more occurrences, texts with similar data already in the dataset, and obviously irrelevant texts.

[0052] Examples of high-quality pseudo-text datasets for each class are shown in Table 3.

[0053] Table 3 Examples of Pseudotext Data

[0054]

[0055] It should be understood that the pseudo-text data generated in this embodiment is English text, but this is only for illustrative purposes and does not mean that only English can be generated. In practical applications, pseudo-text in Chinese or other languages ​​can be generated based on the prompt template and the large language model.

[0056] Step 2: Fine-tuning the recognition model

[0057] This invention uses a Contrastive Language-Image Pre-Training (CLIP) model as the recognition model. For the CLIP model, the dataset is in the form of image-text pairs. Image-text pairs are generated through random sampling based on the support set and pseudo-text dataset for each type of interference. To convert the interference signal into an image, this invention uses Short Time Fourier Transform (STFT) to obtain the time-frequency domain image data of the interference signal. The training set contains 35 samples, including 5 undisturbed samples and 5 samples for each of the 6 interference styles. (Appendix) Figure 2 An example diagram of a COMB interference pattern is provided.

[0058] The backbone image encoder of the CLIP model chosen in this invention is Vit-B / 16. During fine-tuning, following the CLIP forward pass, both image and text data pass through the FSRRM model, including an image encoder and a text encoder, for image feature extraction and text feature extraction, respectively. Then, the feature similarity between the image and text is calculated. During training, CLIP learns embedding vectors by maximizing the similarity with the corresponding image and text descriptions while minimizing the similarity with other mismatched pairs. Through contrastive learning, images and text are mapped to the same embedding space. The loss function is calculated, and the model parameters are backpropagated and updated until the required number of training epochs are reached. This process is called model fine-tuning. This represents the features of a small batch of images and text. 'b' represents the batch number, and 'd' represents the feature dimension. The similarity between the image and text is:

[0059]

[0060] To achieve better model training results, this invention adds two loss terms to the CLIP model loss function: the triplet loss for image features and the triplet loss for text features. The total loss function can be represented by the sum of the following three parts: the triplet loss for image features, the triplet loss for text features, and the loss calculated by using the similarity matrix to calculate the log-likelihood ratio between the image and text, and then calculating the cross-entropy loss, which can be expressed as:

[0061]

[0062] CrossEntropyLoss reflects the alignment between images and text, while the two TripletMarginLoss values ​​reflect the clustering and distinguishability between different categories of images and different categories of text, respectively.

[0063]

[0064] The above formula is the formula for calculating the Triplet Margin Loss function, where x a For anchor point samples, x p For positive samples, x n The negative samples are represented by α, which represents the tolerance. This loss function improves the clustering effect of feature vectors by minimizing the distance between samples of the same class and maximizing the distance between samples of different classes.

[0065] Figure 3 Feature maps of the text dataset before and after fine-tuning are presented. It can be seen that by fine-tuning the CLIP model on the pseudo-text-image dataset, text data of the same class are more clustered in the latent space, while text data of different classes are more distinguishable. Therefore, the fine-tuning process enables the model to identify radar jamming types by calculating and comparing the similarity between the time-frequency domain image of the jamming and the descriptive text cues for each potential class.

[0066] Step 3: Interference Identification

[0067] Based on the fine-tuned model, interference in the input is identified:

[0068] First, the interference signal to be identified is converted into time-frequency domain image data via STFT. Then, together with the category name of each potential interference type, it is passed through an encoder. The image encoder and text encoder calculate the similarity between the time-frequency domain image data to be identified and each potential category, respectively. The similarity values ​​are compared, and the text with the highest similarity is selected as the best matching description. This text type is the output interference type.

[0069] The present invention performs recognition tests on a test dataset containing 2000 samples in each category. Figure 4 The test results are presented in the form of a confusion matrix. As can be seen from the figure, this invention achieves high recognition accuracy for the five types: NCJ, FM, ISRJ, C&I, and COMB. However, the recognition accuracy for SMSP is 72%, slightly lower than the other interference categories. This is because... Figure 3 As can be seen, compared with other classes, the characteristics of SMSP are more dispersed and closer to those of undisturbed classes.

[0070] This invention selects two typical few-shot learning methods—a matching network and a prototype network—as control groups for comparison. As shown in Table 4, the proposed method demonstrates better recognition performance in the NCJ, FM, C&I, and COMB classes. Although the proposed method does not show any advantage over the control methods in the SMSP and ISRJ classes, with accuracy rates 2.73% and 0.9% lower than the matching network, respectively, on average, the recognition accuracy for each interference type is 93.32%, representing improvements of 4.15% and 5.94% compared to the prototype network and the matching network, respectively.

[0071] Table 4 Accuracy Comparison Results

[0072]

[0073] This invention identifies query signals by selecting a set of 6 ways and 5 shots as the support set; text descriptions are selected as supporting data to help establish consistency between interference and category. First, consistency between the signal and the text description is established; then, identification is proven by calculating the similarity between the input signal and the potential text description; finally, the best-matching text description should reflect the interference type of the input signal. Simulation results show that with only 5 samples for each interference type, the average recognition accuracy reaches 93.32%. Therefore, it can be said that this invention achieves effective interference identification even under conditions of scarce radar interference samples. Furthermore, this invention utilizes the semantic generalization ability of existing pre-trained models and employs text enhancement for fine-tuning, achieving radar interference pattern recognition under small sample conditions; compared with traditional methods, it has strong portability, and the pre-trained model is replaceable; and it uses LLM generation to replace manual data collection, exhibiting good simplicity and efficiency.

[0074] According to another embodiment of the present invention, a few-sample radar interference identification system with synthetic pseudo-text enhancement is provided, comprising:

[0075] The pseudo-text data generation module is used to generate a support set of k-class n samples for radar interference types. Using a generative large language model, based on the support set and a pre-set list of prompt templates for querying interference types, it randomly provides prompt templates and temperature parameters to generate pseudo-text corresponding to each interference type. The pseudo-text is then filtered to obtain a pseudo-text dataset for each interference type.

[0076] The training set generation module is used to convert interference signals into sample image data using the time-frequency transformation method based on the support set; based on the sample image data and pseudo-text dataset of each type of interference, image-text pairs are generated by random sampling, generating a total of k-classified image-text pair datasets to form the training set;

[0077] The model training module is used to input the training set into the image encoder and text encoder of the contrastive language-image pre-trained CLIP model, extract image features and text features respectively, calculate the feature similarity between the image and the text, calculate the cross-entropy loss based on the similarity, and calculate the triplet loss of image features and the triplet loss of text features to form the total loss function. The model parameters are backpropagated based on the total loss function and updated until the required number of training rounds are reached, thereby achieving fine-tuning of the CLIP model.

[0078] The interference identification module is used to convert the interference signal to be identified into time-frequency domain image data using the time-frequency transformation method. The time-frequency domain image data and the interference type category name are input into the image encoder and text encoder of the fine-tuned CLIP model, respectively. Image features and category text features are extracted, and the similarity between the image features and each category text feature is calculated. The text with the highest similarity is taken as the best matching description, and this text type is the output interference type.

[0079] It should be understood that the synthetic pseudo-text enhanced small-sample radar interference identification system provided in this embodiment can implement all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0080] The present invention also provides a computer device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the synthetic pseudo-text enhanced few-sample radar interference identification method as described above.

[0081] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the synthetic pseudo-text enhanced few-sample radar interference identification method as described above.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A synthetic pseudo-text enhanced small sample radar jamming recognition method, characterized in that, The method comprises the following steps: According to the support set, the interference signal is converted into sample image data by using a time-frequency transformation method; according to the sample image data and the pseudo-text data set of each interference type, an image-text pair is generated by random sampling, a total of k classification image-text pair data sets are generated, and a training set is formed; The training set is input into the image encoder and the text encoder of the contrast language-image pre-training CLIP model, image feature extraction and text feature extraction are respectively performed, the feature similarity between the image and the text is calculated, the cross-entropy loss is calculated according to the similarity, the triplet loss of the image feature and the triplet loss of the text feature are calculated, a total loss function is formed, the model parameters are updated by back propagation according to the total loss function until the required training number of rounds is reached, and the CLIP model is fine-tuned; The interference signal to be recognized is converted into time-frequency domain image data by using a time-frequency transformation method, the time-frequency domain image data and the interference type category name are input into the image encoder and the text encoder of the fine-tuned CLIP model respectively, the image feature and the category text feature are extracted respectively, the similarity between the image feature and each category text feature is calculated, the text with the maximum similarity is taken as the best matching description, and the text type is the output interference type. The generative large language model includes Llama-3-8b-instruct, Gemma-7b-it, Openchat-7b, and Mistral-7b-instruct, and the temperature parameter is a random floating point number between 0 and 1.

2. The method of claim 1, wherein, According to the support set and the preset prompt template list for interference type query, the generative large language model randomly gives a prompt template and a temperature parameter, generates pseudo-text corresponding to each interference type, and filters the pseudo-text to obtain pseudo-text data set of each interference type.

3. The method of claim 1, wherein, According to the support set, the prompt template list for interference type query is determined, and the prompt template is a prompt sentence for inquiring the characteristics of a specified interference type; According to the selected generative large language model, a large language model list is formed; For each query process, a prompt template and a large language model are randomly selected from the prompt template list and the large language model list according to the temperature parameter to obtain corresponding pseudo-text. The pseudo-text is filtered to obtain pseudo-text data of each interference type, including:

4. The method of claim 1, wherein, The text of the query response is split by a specified separator to obtain short texts; for the short texts, sentences with lengths not in a specified range, irrelevant texts, and texts of the same type are removed to obtain high-quality pseudo-text data sets of each interference type. The feature similarity between the image and the text is represented as:

5. The method of claim 1, wherein, The total loss function is represented as: b denotes batch number, d denotes feature dimension.

6. The method of claim 5, wherein, TripletMarginLoss represents the triplet loss, and CrossEntropyLoss represents the cross-entropy loss. ​ 7. The method of claim 1, wherein, The time-frequency transformation method adopts a short-time Fourier transform.

8. A synthetic pseudo-text enhanced small sample radar jamming recognition system, characterized in that, The method comprises the steps of: The pseudo-text data generation module is configured to generate a support set of k categories and n samples for each radar jamming type, use a generative large language model to generate a pseudo-text corresponding to each jamming type according to the support set and a preset prompt template list for jamming type query, and filter the pseudo-text to obtain a pseudo-text data set of each jamming type; The training set generation module is configured to convert the jamming signals into sample image data by using a time-frequency transformation method according to the support set, generate image-text pairs by random sampling according to the sample image data and the pseudo-text data set of each jamming type, generate a total of k category image-text pair data sets, and form a training set; The model training module is configured to input the training set into an image encoder and a text encoder of a contrastive language-image pre-training CLIP model, respectively perform image feature extraction and text feature extraction, calculate the feature similarity between the image and the text, calculate the cross-entropy loss according to the similarity, calculate the three-tuple loss of the image feature and the three-tuple loss of the text feature, form a total loss function, perform back propagation on the model parameters according to the total loss function, and update the parameters until a required number of training rounds is reached, so as to fine-tune the CLIP model; The interference recognition module is configured to convert the to-be-recognized jamming signals into time-frequency domain image data by using a time-frequency transformation method, input the time-frequency domain image data and the jamming type category name into the image encoder and the text encoder of the fine-tuned CLIP model respectively, extract the image feature and the category text feature respectively, calculate the similarity between the image feature and each category text feature, take the text with the maximum similarity as the best matching description, and output the text type as the jamming type.

9. A computer device, comprising: The method comprises the steps of: one or more processors; memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs are executed by the processor to implement the steps of the small sample radar jamming recognition enhanced by the synthesized pseudo-text as claimed in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the small sample radar jamming recognition method enhanced by the synthesized pseudo-text as claimed in any one of claims 1-7.

Images