Radar interference signal semi-supervised identification method, system, equipment and medium

Abstract

The embodiment of the invention relates to the technical field of radars, in particular to a radar interference signal semi-supervised recognition method, system and device and a medium, and the method comprises the steps: carrying out the short-time Fourier transform of a radar interference signal, and constructing a training sample set containing a labeled training sample and an unlabeled training sample; pre-training the CNN network based on the labeled training sample to obtain a pre-training model; using the pre-training model to predict the unlabeled training samples, generating pseudo labels, screening the pseudo labels, and using the retained unlabeled training samples with high confidence to calculate a semi-supervised loss value; performing category perception comparison learning on the pre-training model by using the unlabeled training sample, and calculating a category perception comparison loss value; calculating a total loss value based on the semi-supervised loss value and the category perception comparison loss value, and performing fine tuning on the pre-training model to obtain an identification model; and inputting the radar interference signal to be identified into the identification model to obtain an identification result output by the identification model.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of radar technology, and in particular to a method, system, device, and medium for semi-supervised recognition of radar interference signals. Background Art

[0002] In recent years, with the rapid development of wireless communications, radar detection, and electronic countermeasures, interference issues in complex electromagnetic environments have become increasingly prominent. Radar systems play a vital role in multiple fields, including military, aerospace, and transportation. Their core tasks include target detection, tracking, and identification.

[0003] In modern adversarial environments, radars face a variety of interference signals. These interference signals severely affect the normal operation of radars through suppression or deception, making it difficult to accurately detect targets. How to effectively identify and suppress interference signals has become an important research direction in the current field of radar anti-interference. However, the currently commonly used radar interference signal recognition methods based on deep learning often require a large number of interference signal samples to train the model. However, under normal circumstances, radar interference signals in real electromagnetic environments are often difficult to obtain, and the cost of manual labeling is extremely high, which leads to a serious shortage of labeled training samples. How to use a small number of labeled samples and a large number of unlabeled samples to perform semi-supervised training on the model and improve the model's interference recognition performance under the condition of scarce labeled samples has become one of the key technical challenges in the field of radar interference recognition.

[0004] Semi-supervised learning (SSL) is a machine learning method that lies between supervised and unsupervised learning. It aims to train a model using a small number of labeled samples and a large number of unlabeled samples, thereby improving the model's generalization ability and performance. Currently, research teams at home and abroad have achieved relatively mature results in semi-supervised learning, and it is gradually being integrated with deep learning. S. Laine proposed a direct and effective semi-supervised learning method to train deep learning networks. This method uses the network output to form consistent predictions for unknown labels under various regularization and data augmentation conditions. Tarvainen et al. proposed a method that averages model weights rather than predicted labels to create a target teacher model. Miyato et al. proposed a new regularization method based on a virtual adversarial loss, which is defined as the robustness of the conditional label distribution around each input data to local perturbations. Xie et al. predicted pseudo-labels from a teacher model and used these pseudo-labels to guide a student model, ultimately enabling the student model to surpass the performance of the teacher model without requiring true labels. Zhu et al. developed a semi-supervised stacked convolutional denoising autoencoder with a label regularization term to accurately capture key features and improve classification accuracy. Luo et al. proposed a few-shot radar interference recognition method based on the time-frequency self-attention mechanism. Simulation results show that the method can still maintain an accuracy of more than 95% when there are few samples (the proportion of labeled samples is 3% to 20%). Tian et al. proposed a radar interference recognition method based on multi-resolution feature extraction and few-shot learning (FSL). This method combines wavelet transform, empirical mode decomposition and prototype network to extract features and classify interference signals. Simulation results show that the accuracy of four interference types can still be maintained above 90% on a dataset with only 8,400 samples.

[0005] Through the above analysis, it can be found that the current radar jamming signal technology still has the following problems.

[0006] First, the current radar interference signal recognition method based on deep learning mainly relies on a large amount of labeled data sets to conduct fully supervised training of the model to realize the identification of radar interference types. However, in the real environment, since radar interference signals are difficult to collect and manual labeling of interference data requires a lot of time and cost, the labeled samples available for model training are relatively scarce, and the training samples often contain a large number of unlabeled samples. The current radar interference signal recognition method based on deep learning often cannot effectively use unlabeled samples to train the model.

[0007] Second, traditional semi-supervised radar jamming signal recognition methods typically generate pseudo-labels for unlabeled samples using pseudo-labeling and consistency regularization, and then use a fixed threshold method to filter high-confidence pseudo-labels for model training. However, these methods fail to effectively utilize the differences between different types of jamming. This leads to the accumulation of errors caused by incorrect pseudo-labels during training, affecting the model's recognition accuracy. Summary of the Invention

[0008] In order to solve the above technical problems, the embodiments of the present application propose a semi-supervised identification method, system, device and medium for radar interference signals, which effectively screen out high-confidence pseudo-label samples through adaptive pseudo-label screening, reduce the error accumulation caused by erroneous pseudo-labels in the semi-supervised training process, and effectively utilize the difference information between different interference categories through category-aware comparative learning, thereby improving the confidence of the pseudo-labels output during the semi-supervised training process, thereby effectively improving the recognition accuracy and recognition versatility of radar interference signals.

[0009] In order to achieve the above-mentioned purpose, an embodiment of the present application proposes a semi-supervised recognition method for radar interference signals, the method comprising: performing a short-time Fourier transform on the radar interference signal to obtain a time-frequency diagram of the radar interference signal, and constructing a radar interference time-frequency diagram dataset as a training sample set based on the time-frequency diagram of the radar interference signal, wherein the training sample set includes a small number of labeled training samples and a large number of unlabeled training samples; pre-training a CNN network based on the labeled training samples to obtain a pre-training model; using the pre-training model to predict the unlabeled training samples, generating pseudo labels corresponding to each unlabeled training sample, and adopting an adaptive pseudo-labeling algorithm to obtain a semi-supervised recognition method for radar interference signals. The label filtering algorithm filters the pseudo labels, retains the unlabeled training samples corresponding to the pseudo labels with high confidence, and uses the retained unlabeled training samples to calculate the semi-supervised loss value; the unlabeled training samples are used to perform category-aware contrastive learning on the pre-trained model to calculate the category-aware contrastive loss value; the total loss value is calculated based on the semi-supervised loss value and the category-aware contrastive loss value, and the pre-trained model is fine-tuned based on the total loss value to obtain the trained recognition model; the recognition model is deployed to the scenario where it is needed, and the radar interference signal to be identified is input into the recognition model to obtain the recognition result output by the recognition model.

[0010] In order to achieve the above-mentioned purpose, the embodiment of the present application also proposes a semi-supervised recognition system for radar interference signals, which includes: a training sample set construction module, which is used to perform short-time Fourier transform on the radar interference signal to obtain a radar interference signal time-frequency diagram, and construct a radar interference time-frequency diagram data set as a training sample set based on the radar interference signal time-frequency diagram, wherein the training sample set contains a small number of labeled training samples and a large number of unlabeled training samples; a pre-training module, which is used to pre-train a pre-constructed CNN network based on the labeled training samples in the training sample set to obtain a pre-trained model; a semi-supervised learning module, which is used to use the pre-training model to predict the unlabeled training samples in the training sample set to generate pseudo labels corresponding to each unlabeled training sample , an adaptive pseudo-label screening algorithm is used to screen pseudo-labels, retain unlabeled training samples corresponding to pseudo-labels with high confidence, and use the retained unlabeled training samples to calculate the semi-supervised loss value; a category-aware contrastive learning module is used to perform category-aware contrastive learning on the pre-trained model using unlabeled training samples, and calculate the category-aware contrastive loss value; a fine-tuning training module is used to calculate the total loss value based on the semi-supervised loss value and the category-aware contrastive loss value, and fine-tune the pre-trained model based on the total loss value to obtain a trained recognition model; a model deployment and use module is used to deploy the recognition model to a scenario where it is needed, input the radar interference signal to be identified into the recognition model, and obtain the recognition result output by the recognition model.

[0011] In order to achieve the above-mentioned objectives, an embodiment of the present application also proposes an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a semi-supervised identification method of radar interference signals as described above.

[0012] In order to achieve the above-mentioned purpose, an embodiment of the present application further provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement a semi-supervised identification method of radar interference signals as described above.

[0013] Alternatively, let the radar interference signal be x(t), and let the window function for short-time Fourier transform be The method of performing short-time Fourier transform on the radar interference signal to obtain a radar interference signal time-frequency graph, and constructing a radar interference signal time-frequency graph dataset as a training sample set based on the radar interference signal time-frequency graph, includes:

[0014] The time-frequency diagram of x(t) is obtained by performing a short-time Fourier transform on x(t) using the following formula:

[0015]

[0016] Among them, STFT x (t,f) represents the time-frequency diagram of x(t);

[0017] Based on the time-frequency graphs of all radar interference signals obtained by short-time Fourier transform, a radar interference time-frequency graph dataset is constructed as a training sample set. Assuming that the total number of types of radar interference signals is M, under the setting of semi-supervised classification task, the training sample set is divided into labeled training sample subsets a={(a i ,y i ):i∈(1,…,B)} and the unlabeled training sample subset b={b j :j∈(1,…,μB)}, where μ is a hyperparameter used to control the ratio between labeled training samples and unlabeled training samples, B represents the number of labeled training samples, and b j represents the jth unlabeled training sample, a i represents the i-th labeled training sample, y i represents the M-dimensional one-hot classification label corresponding to the i-th labeled training sample, For a i In terms of iC =1 means a i The corresponding category is category C.

[0018] Optionally, the pre-training of a pre-built CNN network based on labeled training samples to obtain a pre-trained model includes:

[0019] Input the labeled training samples in the labeled training sample subset into the CNN network, and obtain the predicted probability of each labeled training sample output by the CNN network Based on y i and The cross entropy loss function is used to pre-train the CNN network to obtain a pre-trained model;

[0020] The cross entropy loss function is expressed as follows:

[0021]

[0022] Among them, L CrossEntropy is the cross entropy loss function value.

[0023] Optionally, the method of using a pre-trained model to predict unlabeled training samples, generating pseudo labels corresponding to each unlabeled training sample, and using an adaptive pseudo label screening algorithm to screen the pseudo labels to retain unlabeled training samples corresponding to pseudo labels with high confidence levels includes:

[0024] Perform data augmentation on the unlabeled training samples in the unlabeled training sample subset to obtain a strongly enhanced version of each unlabeled training sample Aug s (b j ) and the weakly enhanced version Aug w (b j );

[0025] Aug w (b j ) is input into the pre-training model to obtain the Aug w (b j )’s one-hot prediction vector p jw , and select p iw The category corresponding to the maximum value in b is j Pseudo labels

[0026] will p jw The probability values ​​corresponding to each category in are arranged in descending order to obtain the sorted prediction vector

[0027] calculate The ratio of the probability value of the latter item to the probability value of the previous item in the equation is obtained, which is the same as b j The corresponding ratio vector A, A is expressed by the formula:

[0028]

[0029] k=1,2,3,…,M-1;

[0030] Where ξ is the ratio factor, ξ = 1 / M;

[0031] Let the serial number corresponding to the maximum value in A be r, and the pseudo labels of the unlabeled training samples corresponding to the items in A whose serial numbers satisfy r-1≤1 are taken as pseudo labels with high confidence, and these unlabeled training samples are retained.

[0032] Optionally, calculating the semi-supervised loss value using the retained unlabeled training samples includes:

[0033] The strongly enhanced version of the retained unlabeled training sample is input into the pre-training model to obtain the one-hot prediction vector p of the strongly enhanced version of the retained unlabeled training sample output by the pre-training model js ;

[0034] Based on p js and the pseudo labels corresponding to the retained unlabeled training samples Calculate and get the semi-supervised loss value L u , L u It is expressed by the formula:

[0035]

[0036] Optionally, performing category-aware contrastive learning on the pre-trained model using unlabeled training samples to calculate the category-aware contrastive loss value includes:

[0037] Perform data augmentation on the unlabeled training samples in the unlabeled training sample subset to obtain a weakly enhanced version Aug of each unlabeled training sample w (b j ) and two strongly enhanced versions Aug s1 (b j ) and Aug s2 (b j );

[0038] Aug s1 (b j ) and Aug s2 (b j ) is input into the feature extractor E of the pre-trained model to extract high-dimensional abstract features and obtain high-dimensional abstract features E[Aug s1 (b j )] and E[Aug s2 (b j )];

[0039] The projection head g of the pre-trained model is used to transform E[Aug s1 (b j )] and E[Aug s2 (b j )] is mapped to the low-dimensional embedding space to obtain the embedded feature vector z s1 and z s2 , finally based on z s1 and z s2 Calculate the feature similarity matrix S;

[0040] Among them, z s1 =g{E[Aug s1 (b j )]},z s2 =g{E[Aug s2 (b j )]};

[0041] Aug w (b j ) is input into the pre-training model to obtain the Aug w (b j )’s one-hot prediction vector p jw , and select p jw The category corresponding to the maximum value in b isj Pseudo labels

[0042] The unlabeled training samples with the same pseudo labels are regarded as positive samples, and the unlabeled training samples with different pseudo labels are regarded as negative samples, and the supervised contrast matrix W is constructed. scon ;

[0043] Set T push As the threshold for clustering and comparison in feature space, for sample pairs and If it is judged that i=j or and The confidence level v i and v j Both are greater than T push , then the sample pair and Considered as a positive sample pair, thus bringing Aug closer w (b i ) and Aug w (b j ) and the weakly enhanced version of other similar unlabeled training samples in the feature space, if i≠j and and The confidence level v i and v j There is one less than or equal to T push , then the sample pair and is regarded as a negative sample pair, thereby widening the distance between the weakly enhanced versions of the heterogeneous unlabeled training samples in the feature space, and finally W scon Converted into category-aware contrast matrix W clacon ;

[0044] Using the confidence of the pseudo label, W clacon Perform reweighting and set the confidence of the pseudo label of the i-th unlabeled training sample to be v i , the confidence of the pseudo label of the jth unlabeled training sample is v j , then calculate W clacon The reweighting factor of the element in row i and column j is Thus we get the target reweight matrix W target ;

[0045] By minimizing S and W target The cross entropy of the class perception contrast loss value L is calculated c , L c It is expressed by the formula:

[0046]

[0047] Where j≠i, τ is the temperature coefficient, Φ(i) is the confidence of the pseudo label greater than T push And the index set belongs to the same category, |Φ(i)| represents the cardinality of Φ(i), and |Φ(i)|+1 represents all positive sample pairs.

[0048] Optionally, the total loss value is calculated based on the semi-supervised loss value and the category-aware contrast loss value, and is implemented by the following formula:

[0049] L total =L u +λL c ;

[0050] Among them, L u is the semi-supervised loss value, L c is the category-aware contrast loss value, and λ is a hyperparameter used to balance L u and L c The weight, L total is the total loss value.

[0051] The semi-supervised identification method, system, device and medium of radar jamming signal proposed in this application have at least the following advantages compared with traditional semi-supervised identification methods of radar jamming signal.

[0052] First, this application designs an adaptive pseudo-label screening algorithm to adaptively screen high-confidence pseudo-labels without manually setting the threshold. It overcomes the problem of excessive reliance on threshold setting in traditional threshold-based pseudo-label screening methods, reduces the error accumulation caused by erroneous pseudo-labels during semi-supervised training, and greatly improves the accuracy of semi-supervised recognition.

[0053] Second, this application designs a category-aware contrastive learning process, which can improve the compactness of similar features during semi-supervised training, while widening the distance between heterogeneous samples, enabling the model to effectively utilize the difference information between different radar interference categories, thereby improving the confidence of the pseudo-label output during semi-supervised training, and further improving the convergence speed and recognition performance of the model during semi-supervised training.

[0054] In summary, this application studies the importance of adaptive pseudo-label screening and introducing category-aware contrastive learning in semi-supervised recognition of radar jamming signals. Any scenario involving semi-supervised recognition of radar jamming signals, especially when the training samples contain a large number of unlabeled samples, can use the semi-supervised recognition method of radar jamming signals proposed in this application for semi-supervised recognition. The semi-supervised recognition method of radar jamming signals proposed in this application has extremely high recognition accuracy and strong versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the related technologies, the following is a brief introduction to the drawings required for use in the embodiments of the present application or the description of the related technologies. Obviously, the following drawings are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. The drawings described here are only used to explain the present application and are not used to limit the present application.

[0056] Figure 1 is a flowchart of a semi-supervised identification method for radar jamming signals provided in one embodiment of the present application;

[0057] Figure 2 is a schematic diagram of a semi-supervised learning process provided in one embodiment of the present application;

[0058] Figure 3 is a schematic diagram of a category-aware contrast learning process provided in one embodiment of the present application;

[0059] Figure 4 1 is a schematic diagram of a confusion matrix of the recognition performance of a semi-supervised recognition method for radar interference signals proposed in this application under different numbers of labels, provided in one embodiment of the present application;

[0060] Figure 5 This is a comparison chart of the recognition performance of a radar jamming signal semi-supervised recognition method proposed in this application before and after removing category-aware contrastive learning, provided in one embodiment of the present application;

[0061] Figure 6 This is a comparison chart of the recognition performance of a radar jamming signal semi-supervised recognition method proposed in this application and other radar jamming signal semi-supervised recognition methods, provided in one embodiment of the present application;

[0062] Figure 7 is a structural diagram of a radar jamming signal semi-supervised recognition system provided in another embodiment of the present application;

[0063] Figure 8 It is a structural diagram of an electronic device provided in another embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, each embodiment of the present application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will appreciate that in each embodiment of the present application, many technical details are provided to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in the present application can be implemented. The division of the following embodiments is for convenience of description and should not constitute any limitation on the specific implementation of the present application. The various embodiments can be combined with each other and referenced to each other under the premise of no contradiction.

[0065] One embodiment of the present application proposes a semi-supervised identification method for radar jamming signals, which is applied to an electronic device, where the electronic device can be a terminal or a server. This embodiment and the following embodiments are described using a server as an example. The following describes the implementation details of the semi-supervised identification method for radar jamming signals proposed in this embodiment. The following content is provided for ease of understanding only and is not required for implementing this solution.

[0066] The specific process of the semi-supervised identification method of radar jamming signals proposed in this embodiment can be as follows: Figure 1 As shown, including:

[0067] Step 11: Perform short-time Fourier transform on the radar interference signal to obtain a radar interference signal time-frequency diagram, and construct a radar interference signal time-frequency diagram dataset as a training sample set based on the radar interference signal time-frequency diagram. The training sample set contains a small number of labeled training samples and a large number of unlabeled training samples.

[0068] In its implementation, the server first prepares for training the recognition model by constructing a training sample set. The server collects a large number of radar interference signals, performs a short-time Fourier transform on them, and generates radar interference signal time-frequency maps. The server then constructs a radar interference time-frequency map dataset based on these time-frequency maps as the training sample set. It's important to note that the training sample set includes a small number of labeled training samples and a large number of unlabeled training samples.

[0069] In one example, let the radar interference signal be x(t), and let the window function for short-time Fourier transform be The server performs a short-time Fourier transform on x(t) using the following formula to obtain the time-frequency graph of x(t):

[0070]

[0071] Among them, STFT x (t,f) represents the time-frequency diagram of x(t).

[0072] Next, the server needs to construct a radar interference time-frequency map dataset as a training sample set based on all radar interference signal time-frequency maps obtained by short-time Fourier transform. Assuming that the total number of radar interference signal types in the training sample set is M, under the setting of semi-supervised classification task, the training sample set is divided into labeled training sample subsets a = {(a i ,y i ):i∈(1,…,B)} and the unlabeled training sample subset b={b j :j∈(1,…,μB)}, where μ is a hyperparameter used to control the ratio between labeled training samples and unlabeled training samples. Generally speaking, μ is an integer greater than 1, B represents the number of labeled training samples, and b j represents the jth unlabeled training sample, a i represents the i-th labeled training sample, a i 、b j Will be used as the input of the subsequent model, y i represents the M-dimensional one-hot classification label corresponding to the i-th labeled training sample, For a i In terms of iC =1 means a i The corresponding category is category C.

[0073] Step 12: Pre-train the CNN network based on the labeled training samples to obtain a pre-trained model.

[0074] In the specific implementation, the server extracts all labeled training samples from the training sample set, and uses the labeled training samples to pre-train the pre-built CNN network to obtain a pre-trained model.

[0075] In one example, the server inputs the labeled training samples in the labeled training sample subset into the constructed CNN network, and obtains the predicted probability of each labeled training sample output by the CNN network. Based on y i and The cross entropy loss function is used to pre-train the CNN network to obtain a pre-trained model.

[0076] The cross entropy loss function can be expressed as:

[0077]

[0078] Among them, L CrossEntropy is the cross entropy loss function value.

[0079] In step 13, the pre-trained model is used to predict the unlabeled training samples, and pseudo labels corresponding to each unlabeled training sample are generated. The pseudo labels are screened using an adaptive pseudo label screening algorithm, and the unlabeled training samples corresponding to the pseudo labels with high confidence are retained. The semi-supervised loss value is calculated using the retained unlabeled training samples.

[0080] In practice, the pre-trained model must undergo semi-supervised learning and category-aware contrastive learning before it can become a mature recognition model. Semi-supervised learning is the first step. The server uses the pre-trained model to predict all unlabeled training samples in the training sample set, generating pseudo-labels for each unlabeled training sample. An adaptive pseudo-label screening algorithm is then used to filter the pseudo-labels, retaining the unlabeled training samples corresponding to the pseudo-labels with high confidence. The retained unlabeled training samples are then used to calculate the semi-supervised loss value to achieve semi-supervised learning.

[0081] In an example, the server first needs to perform data augmentation on the unlabeled training samples in the unlabeled training sample subset to obtain a strongly enhanced version Aug of each unlabeled training sample. s (b j ) and the weakly enhanced version Aug w (b j ).

[0082] Then, Aug w (b j ) is input into the pre-training model f, and the output of the pre-training model is obtained. w (b j )’s one-hot prediction vector p jw , p jw =f[Aug w (b j )], and select p jw The category corresponding to the maximum value in b is j Pseudo labels

[0083] Next, p jw The probability values ​​corresponding to each category in are arranged in descending order to obtain the sorted prediction vector Calculate in sequence The ratio of the probability value of the latter item to the probability value of the previous item in the equation is obtained, which is the same as b j The corresponding ratio vector A, A is expressed by the formula:

[0084]

[0085] k=1,2,3,…,M-1;

[0086] Where ξ is the ratio factor, ξ = 1 / M.

[0087] Afterwards, the serial number corresponding to the maximum value in A is recorded as r. From this, we can see that the probability change of the ratio vector A at r is the largest, and the probability change before r is relatively stable. Therefore, the server takes the pseudo labels of the unlabeled training samples corresponding to the items in A whose serial numbers satisfy r-1≤1 as pseudo labels with high confidence, and retains these unlabeled training samples.

[0088] In one example, after obtaining the reserved unlabeled training samples, the server needs to input the strongly enhanced version of the reserved unlabeled training samples into the pre-training model to obtain the one-hot prediction vector p of the strongly enhanced version of the reserved unlabeled training samples output by the pre-training model. js , and finally based on p js and the pseudo labels corresponding to the retained unlabeled training samples Calculate and get the semi-supervised loss value L u , L u It is expressed by the formula:

[0089]

[0090] In one example, the semi-supervised learning process of the model can be as follows Figure 2 shown.

[0091] In step 14, the pre-trained model is subjected to category-aware contrastive learning using unlabeled training samples, and the category-aware contrastive loss value is calculated.

[0092] In the specific implementation, category-aware contrastive learning is performed simultaneously with unsupervised learning. The server uses all unlabeled training samples in the training sample set to perform category-aware contrastive learning on the pre-trained model, thereby calculating the category-aware contrastive loss value.

[0093] In an example, the server first needs to perform data augmentation on the unlabeled training samples in the unlabeled training sample subset to obtain a weakly enhanced version Aug of each unlabeled training sample. w (b j ) and two strongly enhanced versions Aug s1 (b j ) and Aug s2 (b j ).

[0094] Next, Aug s1 (b j ) and Aug s2 (b j) is input into the feature extractor E of the pre-trained model to extract high-dimensional abstract features and obtain high-dimensional abstract features E[Aug s1 (b j )] and E[Aug s2 (b j )], and then E[Aug s1 (b j )] and E[Aug s2 (b j )] is mapped to the low-dimensional embedding space to obtain the embedded feature vector z s1 and z s2 Among them, z s1 =g{E[Aug s1 (b j )]},z s2 =g{E[Aug s2 (b j )]}.

[0095] Based on z s1 and z s2 The feature similarity matrix S can be calculated.

[0096] The i-th embedded feature vector z in the feature similarity matrix S i and the jth embedded feature vector z j The similarity is denoted as s ij , s ij It is expressed by the formula:

[0097] s ij =exp[(z i z j ) / τ];

[0098] Where τ is the temperature coefficient.

[0099] After that, Aug w (b j ) is input into the pre-training model to obtain the Aug w (b j )’s one-hot prediction vector p jw , and select p jw The category corresponding to the maximum value in b is j Pseudo labels The unlabeled training samples with the same pseudo labels are regarded as positive samples, and the unlabeled training samples with different pseudo labels are regarded as negative samples, and the supervised contrast matrix W is constructed. scon ,

[0100] W sconis defined as follows:

[0101]

[0102] in, W scon The element at row i and column j of .

[0103] Set T push As the threshold for clustering and comparison in feature space, for sample pairs and If it is judged that i=j or and The confidence level v i and v j Both are greater than T push , then the sample pair and Considered as a positive sample pair, thus bringing Aug closer w (b i ) and Aug w (b j ) and the weakly enhanced version of other similar unlabeled training samples in the feature space, if i≠j and and The confidence level v i and v j There is one less than or equal to T push , then the sample pair and is regarded as a negative sample pair, thereby widening the distance between the weakly enhanced versions of the heterogeneous unlabeled training samples in the feature space, and finally W scon Converted into category-aware contrast matrix W clacon .set up W clacon The element in row i and column j of It is used to indicate whether the i-th sample and the j-th sample constitute a positive sample pair, which is defined as follows:

[0104]

[0105] Among them, v i is the confidence of the pseudo label of the i-th unlabeled training sample, v j is the confidence of the pseudo label of the jth unlabeled training sample.

[0106] The category-aware contrastive learning process of the model can be described as Figure 3 As shown, when W is obtained clacon After that, the server needs to use the confidence of the pseudo label to clacon Reweighting, server v i and v j Based on the calculation of Wclacon The reweighting factor of the element in row i and column j is Thus we get the target reweight matrix W target .

[0107] W target The element in row i and column j of is defined as follows:

[0108]

[0109] Finally, the server minimizes S and W target The cross entropy of the class perception contrast loss value L is calculated c , L c It is expressed by the formula:

[0110]

[0111] Where j≠i, τ is the temperature coefficient, Φ(i) is the confidence of the pseudo label greater than T push And the index set belongs to the same category, |Φ(i)| represents the cardinality of Φ(i), and |Φ(i)|+1 represents all positive sample pairs.

[0112] In step 15, a total loss value is calculated based on the semi-supervised loss value and the category-aware contrast loss value, and the pre-trained model is fine-tuned based on the total loss value to obtain a trained recognition model.

[0113] In the specific implementation, after obtaining the semi-supervised loss value and the category-aware contrast loss value, the server can calculate the total loss value based on the semi-supervised loss value and the category-aware contrast loss value, and fine-tune the pre-trained model based on the total loss value to obtain a trained recognition model.

[0114] In an example, the server calculates the total loss value based on the semi-supervised loss value and the category-aware contrast loss value, which can be achieved by the following formula:

[0115] L total =L u +λL c ;

[0116] Among them, L u is the semi-supervised loss value, L c is the category-aware contrast loss value, and λ is a hyperparameter used to balance L u and L c The weight, L total is the total loss value.

[0117] Step 16: deploy the recognition model to a scenario where it is needed, input the radar interference signal to be identified into the recognition model, and obtain the recognition result output by the recognition model.

[0118] In a specific implementation, after obtaining the recognition model, the server can deploy the recognition model to the scene where it is needed. When using the recognition model, the radar interference signal to be identified is input into the recognition model to obtain the recognition result output by the recognition model.

[0119] For the radar interference signal x to be identified * For example, the recognition result output by the recognition model f is recorded as y pred ,y pred It is expressed by the formula: pred =argmax j∈{1,2,…,M} [f(x * ) j ].

[0120] The radar jamming signal semi-supervised recognition method, system, device, and medium proposed in this embodiment have at least the following advantages over traditional radar jamming signal semi-supervised recognition methods.

[0121] First, this embodiment designs an adaptive pseudo-label screening algorithm to adaptively screen high-confidence pseudo-labels without manually setting the threshold. This overcomes the problem of excessive reliance on threshold setting in traditional threshold-based pseudo-label screening methods, reduces the error accumulation caused by erroneous pseudo-labels during semi-supervised training, and greatly improves the accuracy of semi-supervised recognition.

[0122] Second, this embodiment designs a category-aware contrastive learning process, which can improve the compactness of similar features during semi-supervised training, while increasing the distance between heterogeneous samples, enabling the model to effectively utilize the difference information between different radar interference categories, thereby improving the confidence of the pseudo-label output during semi-supervised training, and further improving the convergence speed and recognition performance of the model during semi-supervised training.

[0123] In summary, this embodiment studies the importance of adaptive pseudo-label screening and introducing category-aware contrastive learning in semi-supervised recognition of radar jamming signals. Any scenario involving semi-supervised recognition of radar jamming signals, especially when the training samples contain a large number of unlabeled samples, can use the semi-supervised recognition method of radar jamming signals proposed in this embodiment for semi-supervised recognition. The semi-supervised recognition method of radar jamming signals proposed in this embodiment has extremely high recognition accuracy and strong versatility.

[0124] The steps of the various methods described above are divided for clarity of description only. In actual implementation, they can be combined into a single step, or some steps can be broken down into multiple steps. As long as they contain the same logical relationships, they are all within the scope of protection of this application. Insignificant modifications or designs added to the algorithm or process that do not change the core design of the algorithm or process are also within the scope of protection of this application.

[0125] In one embodiment, in order to evaluate the performance of a semi-supervised identification method for radar jamming signals proposed in this application (hereinafter referred to as OURS), we conducted relevant simulation verification.

[0126] The specific simulation conditions and parameters are as follows:

[0127] The experiment was conducted using a computer with a 16-core AMD EPYC 7502 processor, an NVIDIA RTX 4090 GPU, and 96GB of memory. The operating system was Ubuntu 20.04, the software environment was Python 3.10, and the deep learning framework Pytorch 2.3.0 was used to build and train the CNN network. During training, the learning rate was set to 0.001, the batch size was 64, the number of epochs was 50, and the stochastic gradient descent optimizer was used with a momentum factor of 0.9. The hyperparameter λ was set to 0.8, and the category perception threshold T was set to 0. push Set to 0.8.

[0128] In the simulation experiment, eight types of radar jamming signals, including noise amplitude modulation jammer (AM), noise frequency modulation jammer (FM), comb spectrum jammer (COMB), spectrum dispersion jammer (SMSP), slice reconstruction jammer (CI), intermittent sampling forwarding jammer (IS), densely copied false target jammer (MT) and suppression-deception composite jammer, were selected as the main research objects.

[0129] The sampling frequency was set to 2.4 GHz. The interference-to-noise ratio (JNR) range for the training set was -10 dB to 18 dB, and the JNR range for the test set was -10 dB to 18 dB, with intervals of 2 dB. The noise type was Gaussian white noise. The sample set was divided into a training set and a test set. For each radar jammer in the training set, 100 time-frequency samples were generated at each JNR for training, resulting in a total of 8 × 15 × 100 = 12,000 samples. For each jammer in the test set, 20 samples were generated at each JNR for testing, resulting in a total of 8 × 15 × 20 = 2,400 samples.

[0130] Figure 4 The recognition performance (confusion matrix) of OURS under different number of labels is shown. Figure 4As can be seen in the figure, as the number of labels increases, the color of the diagonal line of the confusion matrix gradually darkens, and the recognition model's accuracy for the eight radar interference signals increases accordingly. When the number of labels is small, only 80, the model can achieve recognition accuracy of over 80% for the other seven interference types except composite interference. When the number of labels increases to 480, the model's recognition accuracy for all eight interference signals reaches over 96%.

[0131] Figure 5 The loss function curves and validation set recognition rate curves of the model training process before and after removing category-aware contrastive learning are shown. The experimental results show that OURS can achieve better recognition performance under the same training rounds compared with the model without category-aware contrastive learning. This is because the category-aware contrastive learning process designed by OURS makes the distribution of similar interference signal features more clustered, the distribution of inter-class features more dispersed, and the decision boundary clearer during the model training process. As a result, the recognition model outputs higher pseudo-label confidence for unlabeled training samples during training, promoting more training samples with high-confidence pseudo-labels to participate in training, accelerating model convergence while also improving model recognition performance.

[0132] Figure 6 The recognition performance comparison between OURS and other semi-supervised recognition methods is shown. The comparison methods are fixed threshold PL-CCL, SSGAN and Fixmatch. The fixed threshold PL-CCL method replaces the adaptive threshold pseudo-label screening method with the fixed threshold pseudo-label screening method based on OURS, that is, only pseudo-labels with a confidence level greater than the threshold are screened for training. In order to ensure the fairness of the comparison experiment, the CNN network structure in the PL-CCL and Fixmatch methods is set to be consistent with that in this article. The recognition accuracy of the four methods is tested when the number of labels is 80, 120, 360, and 480, respectively. The results are as follows: Figure 6 As shown, from Figure 6As can be seen from the figure, OURS outperforms the other three methods in terms of the number of labels. The fixed-threshold PL-CCL method requires manual threshold setting to filter pseudo-labels. Setting the threshold too low will result in too many falsely predicted pseudo-labels participating in training, misleading the model's optimization direction. Setting the threshold too high will result in only a small number of high-confidence pseudo-labels above the threshold participating in training. Over-reliance on a small number of high-confidence pseudo-labels may cause the model to overfit these samples, affecting recognition performance and resulting in this method's recognition performance being far lower than OURS. The SSGAN method performs lower than OURS when the number of labels is low. As the number of labels increases, the performance of this method approaches that of OURS. This is because when the number of labels is too small, the SSGAN discriminator may overfit the data. When the amount of labeled data increases, this method can fully utilize the labeled sample information to generate pseudo-samples through the generator, thereby improving the model's recognition performance. OURS solves the problem of sensitive threshold setting by combining adaptive pseudo-label screening and category-aware contrastive learning. At the same time, by introducing category-aware contrastive learning, the model can learn more discriminative features from unlabeled samples, so that it can still maintain high confidence output for unlabeled samples even with a small number of labels. Therefore, OURS can still maintain high recognition performance when the number of labels is small.

[0133] Another embodiment of the present application proposes a semi-supervised recognition system for radar interference signals. The following is a detailed description of the details of the semi-supervised recognition system for radar interference signals proposed in this embodiment. The following content is only for the convenience of understanding the implementation details and is not necessary for the implementation of this embodiment. Figure 7 This is a structural diagram of a radar interference signal semi-supervised recognition system proposed in this embodiment, which specifically includes: a training sample set construction module 21, a pre-training module 22, a semi-supervised learning module 23, a category-aware contrast learning module 24, a fine-tuning training module 25 and a model deployment and use module 26.

[0134] The training sample set construction module 21 is used to perform short-time Fourier transform on the radar interference signal to obtain the radar interference signal time-frequency diagram, and construct a radar interference time-frequency diagram dataset as a training sample set based on the radar interference signal time-frequency diagram. The training sample set contains a small number of labeled training samples and a large number of unlabeled training samples.

[0135] The pre-training module 22 is used to pre-train the pre-built CNN network based on the labeled training samples in the training sample set to obtain a pre-training model.

[0136] The semi-supervised learning module 23 is used to use the pre-trained model to predict the unlabeled training samples in the training sample set, generate pseudo labels corresponding to each unlabeled training sample, use an adaptive pseudo label screening algorithm to screen the pseudo labels, retain the unlabeled training samples corresponding to the pseudo labels with high confidence, and use the retained unlabeled training samples to calculate the semi-supervised loss value.

[0137] The category-aware contrastive learning module 24 is used to perform category-aware contrastive learning on the pre-trained model using unlabeled training samples to calculate the category-aware contrastive loss value.

[0138] The fine-tuning training module 25 is used to calculate the total loss value based on the semi-supervised loss value and the category-aware contrast loss value, and fine-tune the pre-trained model based on the total loss value to obtain a trained recognition model.

[0139] The model deployment and use module 26 is used to deploy the recognition model to the scene where it is needed, input the radar interference signal to be identified into the recognition model, and obtain the recognition result output by the recognition model.

[0140] It is not difficult to find that this embodiment is a system embodiment corresponding to the above-mentioned method embodiment, and this embodiment can be implemented in conjunction with the above-mentioned method embodiment. The relevant technical details and technical effects mentioned in the above-mentioned embodiments are still valid in this embodiment, and to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above-mentioned embodiments.

[0141] It should be noted that the modules mentioned in this embodiment are all logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovation of this application, this embodiment does not include units that are not closely related to solving the technical problem proposed by this application. However, this does not mean that other units do not exist in this embodiment.

[0142] Another embodiment of the present application provides an electronic device, the structure of which is as follows: Figure 8 As shown, it includes: at least one processor 31; and a memory 32 communicatively connected to the at least one processor 31; wherein the memory 32 stores instructions that can be executed by the at least one processor 31, and the instructions are executed by the at least one processor 31 so that the at least one processor 31 can execute a radar interference signal semi-supervised identification method as described in the above-mentioned method embodiments.

[0143] The memory and processor are connected using a bus, which includes any number of interconnected buses and bridges. The bus connects various circuits of one or more processors and memories. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits. These are all well known in the art and therefore will not be described further in this article. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. Data processed by the processor is transmitted on a wireless medium via an antenna. Furthermore, the antenna also receives data and transmits it to the processor.

[0144] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0145] Another embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement a semi-supervised identification method for radar interference signals as described in the above method embodiments.

[0146] That is, those skilled in the art will understand that all or part of the steps in the above-described embodiments can be implemented by instructing the relevant hardware through a program, which is stored in a storage medium and includes a number of instructions for causing a device (such as a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application. Storage media include: U disk, mobile hard disk, read-only memory, random access memory, magnetic disk or optical disk, etc., various media that can store program code.

[0147] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and that in actual applications, various modifications may be made to the embodiments in form and detail without departing from the spirit and scope of the present application. Those skilled in the art will appreciate that improvements and modifications may be made without departing from the principles of the present application, and such improvements and modifications are also considered to be within the scope of protection of the present application.

Claims

1. A semi-supervised recognition method for radar jamming signals, characterized in that: include: Perform short-time Fourier transform on the radar jamming signal to obtain the radar jamming signal time-frequency diagram. Based on the radar jamming signal time-frequency diagram, a radar jamming time-frequency diagram dataset is constructed as a training sample set. The training sample set contains a small number of labeled training samples and a large number of unlabeled training samples. Pre-train the pre-built CNN network based on labeled training samples to obtain a pre-trained model; The pre-trained model is used to predict the unlabeled training samples, and pseudo labels corresponding to each unlabeled training sample are generated. The pseudo labels are filtered using an adaptive pseudo label screening algorithm, and the unlabeled training samples corresponding to the pseudo labels with high confidence are retained. The semi-supervised loss value is calculated using the retained unlabeled training samples. Use unlabeled training samples to perform category-aware contrastive learning on the pre-trained model and calculate the category-aware contrastive loss value; The total loss value is calculated based on the semi-supervised loss value and the category-aware contrast loss value, and the pre-trained model is fine-tuned based on the total loss value to obtain the trained recognition model; Deploy the recognition model to the required scenario, input the radar interference signal to be identified into the recognition model, and obtain the recognition result output by the recognition model.

2. A semi-supervised identification method for radar interference signals according to claim 1, characterized in that The radar interference signal is x(t), and the window function used for short-time Fourier transform is The method of performing short-time Fourier transform on the radar interference signal to obtain a radar interference signal time-frequency graph, and constructing a radar interference signal time-frequency graph dataset as a training sample set based on the radar interference signal time-frequency graph, includes: The time-frequency diagram of x(t) is obtained by performing a short-time Fourier transform on x(t) using the following formula: Among them, STFT x (t,f) represents the time-frequency diagram of x(t); Based on the time-frequency graphs of all radar interference signals obtained by short-time Fourier transform, a radar interference time-frequency graph dataset is constructed as a training sample set. Assuming that the total number of types of radar interference signals is M, under the setting of semi-supervised classification task, the training sample set is divided into labeled training sample subsets a={(a i ,y i ):i∈(1,…,B)} and the unlabeled training sample subset b={b j :j∈(1,…,μB)}, where μ is a hyperparameter used to control the ratio between labeled training samples and unlabeled training samples, B represents the number of labeled training samples, and b j represents the jth unlabeled training sample, a i represents the i-th labeled training sample, y i represents the M-dimensional one-hot classification label corresponding to the i-th labeled training sample, For a i In terms of iC =1 means a i The corresponding category is category C.

3. A semi-supervised identification method for radar jamming signals according to claim 2, characterized in that: The CNN network is pre-trained based on labeled training samples to obtain a pre-trained model, including: Input the labeled training samples in the labeled training sample subset into the CNN network, and obtain the predicted probability of each labeled training sample output by the CNN network Based on y i and The cross entropy loss function is used to pre-train the CNN network to obtain a pre-trained model; The cross entropy loss function is expressed as follows: Among them, L CrossEntropy is the cross entropy loss function value.

4. A semi-supervised identification method for radar jamming signals according to claim 3, characterized in that: The method uses the pre-trained model to predict the unlabeled training samples, generates pseudo labels corresponding to each unlabeled training sample, and uses an adaptive pseudo label screening algorithm to screen the pseudo labels, retaining the unlabeled training samples corresponding to the pseudo labels with high confidence, including: Perform data augmentation on the unlabeled training samples in the unlabeled training sample subset to obtain a strongly enhanced version of each unlabeled training sample Aug s (b j ) and the weakly enhanced version Aug w (b j ); Aug w (b j ) is input into the pre-training model to obtain the Aug w (b j )’s one-hot prediction vector p jw , and select p iw The category corresponding to the maximum value in b is j Pseudo labels will p jw The probability values ​​corresponding to each category in are arranged in descending order to obtain the sorted prediction vector calculate The ratio of the probability value of the latter item to the probability value of the previous item in the equation is obtained, which is the same as b j The corresponding ratio vector A, A is expressed by the formula: k=1,2,3,…,M-1; Where ξ is the ratio factor, ξ = 1 / M; Let the serial number corresponding to the maximum value in A be r, and the pseudo labels of the unlabeled training samples corresponding to the items in A whose serial numbers satisfy r-1≤1 are taken as pseudo labels with high confidence, and these unlabeled training samples are retained.

5. A semi-supervised identification method for radar jamming signals according to claim 4, characterized in that: The semi-supervised loss value is calculated using the retained unlabeled training samples, including: The strongly enhanced version of the retained unlabeled training sample is input into the pre-training model to obtain the one-hot prediction vector p of the strongly enhanced version of the retained unlabeled training sample output by the pre-training model js ; Based on p js and the pseudo labels corresponding to the retained unlabeled training samples Calculate and get the semi-supervised loss value L u , L u It is expressed by the formula:

6. A semi-supervised identification method for radar jamming signals according to claim 3, characterized in that: The method of performing category-aware contrastive learning on the pre-trained model using unlabeled training samples and calculating the category-aware contrastive loss value includes: Perform data augmentation on the unlabeled training samples in the unlabeled training sample subset to obtain a weakly enhanced version Aug of each unlabeled training sample w (b j ) and two strongly enhanced versions Aug s1 (b j ) and Aug s2 (b j ); Aug s1 (jb j ) and Aug s2 (b j ) is input into the feature extractor E of the pre-trained model to extract high-dimensional abstract features and obtain high-dimensional abstract features E[Aug s1 (b j )] and E[Aug s2 (b j )]; The projection head g of the pre-trained model is used to transform E[Aug s1 (b j )] and E[Aug s2 (b j )] is mapped to the low-dimensional embedding space to obtain the embedded feature vector z s1 and z s2 , finally based on z s1 and z s2 Calculate the feature similarity matrix S; Among them, z s1 =g{E[Aug s1 (b j )]},z s2 =g{E[Aug s2 (b j )]}; Aug w (b j ) is input into the pre-training model to obtain the Aug w (b j )’s one-hot prediction vector p jw , and select p jw The category corresponding to the maximum value in b is j Pseudo labels The unlabeled training samples with the same pseudo labels are regarded as positive samples, and the unlabeled training samples with different pseudo labels are regarded as negative samples, and the supervised contrast matrix W is constructed. scon ; Set T push As the threshold for clustering and comparison in feature space, for sample pairs and If it is judged that i=j or and The confidence level v i and v j Both are greater than T push , then the sample pair and Considered as a positive sample pair, thus bringing Aug closer w (b i ) and Aug w (b j ) and the weakly enhanced version of other similar unlabeled training samples in the feature space, if i≠j and and The confidence level v i and v j There is one less than or equal to T push , then the sample pair and is regarded as a negative sample pair, thereby widening the distance between the weakly enhanced versions of the heterogeneous unlabeled training samples in the feature space, and finally W scon Converted into category-aware contrast matrix W clacon ; Using the confidence of the pseudo label, W clacon Perform reweighting and set the confidence of the pseudo label of the i-th unlabeled training sample to be v i , the confidence of the pseudo label of the jth unlabeled training sample is v j , then calculate W clacon The reweighting factor of the element in row i and column j is Thus we get the target reweight matrix W target ; By minimizing S and W target The cross entropy of the class perception contrast loss value L is calculated c , L c It is expressed by the formula: Where j≠i, τ is the temperature coefficient, Φ(i) is the confidence of the pseudo label greater than T push And the index set belongs to the same category, |Φ(i)| represents the cardinality of Φ(i), and |Φ(i)|+1 represents all positive sample pairs.

7. A semi-supervised identification method for radar jamming signals according to any one of claims 1 to 6, characterized in that: The total loss value is calculated based on the semi-supervised loss value and the category-aware contrast loss value, and is implemented by the following formula: THE total =L u +λL c ; Among them, L u is the semi-supervised loss value, L c is the category-aware contrast loss value, and λ is a hyperparameter used to balance L u and L c The weight, L total is the total loss value.

8. A radar jamming signal semi-supervised recognition system, characterized in that: include: The training sample set construction module is used to perform short-time Fourier transform on the radar interference signal to obtain the radar interference signal time-frequency diagram, and construct a radar interference time-frequency diagram dataset based on the radar interference signal time-frequency diagram as the training sample set. The training sample set contains a small number of labeled training samples and a large number of unlabeled training samples; The pre-training module is used to pre-train the pre-built CNN network based on the labeled training samples in the training sample set to obtain a pre-trained model; The semi-supervised learning module is used to use the pre-trained model to predict the unlabeled training samples in the training sample set, generate pseudo labels corresponding to each unlabeled training sample, use an adaptive pseudo label screening algorithm to filter the pseudo labels, retain the unlabeled training samples corresponding to the pseudo labels with high confidence, and use the retained unlabeled training samples to calculate the semi-supervised loss value; The category-aware contrastive learning module is used to perform category-aware contrastive learning on the pre-trained model using unlabeled training samples and calculate the category-aware contrastive loss value; The fine-tuning training module is used to calculate the total loss value based on the semi-supervised loss value and the category-aware contrast loss value, and fine-tune the pre-trained model based on the total loss value to obtain the trained recognition model; The model deployment and use module is used to deploy the recognition model to the required scenario, input the radar interference signal to be identified into the recognition model, and obtain the recognition result output by the recognition model.

9. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the radar interference signal semi-supervised recognition method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it can implement a radar interference signal semi-supervised recognition method according to any one of claims 1 to 7.

Images