A radar jamming signal recognition method based on a transformer

Abstract

This invention proposes a radar jamming signal recognition method based on Transformer, comprising: constructing a radar jamming signal time-domain dataset and processing the radar jamming signal time-domain dataset to obtain a first training dataset; fusing sampling points between samples in the first training dataset to obtain a second training dataset; concatenating the first training dataset and the second training dataset to obtain a third training dataset; inputting the third training dataset into a Transformer model for feature extraction, wherein the loss function of the Transformer model is the sum of the loss functions corresponding to the features contained in the class vector and the loss functions corresponding to the features contained in each sampling window, and obtaining the model parameters that minimize the loss of the Transformer model through an optimization algorithm; obtaining T sampling training sets by using a sampling with replacement strategy on the third training dataset, and training T Transformer models using the sampling training sets and model parameters; inputting the test dataset into the trained T Transformer models, and obtaining the recognition result through voting.

Description

Technical Field

[0001] This invention relates to the field of radar signal identification. Background Technology

[0002] With the development of electronic information technology, international warfare has gradually evolved into high-tech warfare dominated by electronic warfare, in which the normal operation of radar is crucial to victory. However, various radar jamming technologies are constantly emerging, and the existence of jamming seriously threatens the operational effectiveness of radar. Therefore, improving the anti-jamming level of radar is of great military significance, and the efficient and accurate identification of radar jamming signals is the foundation and key to radar anti-jamming technology. Existing radar jamming signal identification methods have the following main problems:

[0003] Radar jamming signal identification methods based on expert knowledge and traditional classifiers require a large amount of manpower and resources, have poor engineering practicality, and the identification results are easily affected by subjective experience factors, resulting in poor reliability.

[0004] Existing deep learning-based radar jamming signal identification methods can be mainly divided into radar jamming signal identification methods based on convolutional neural networks and recurrent neural networks. In the above two methods, the extraction of radar jamming signal features is limited by the distance between sampling points, and it is impossible to effectively extract the dependency relationship between long-distance sampling points, thus affecting the radar jamming signal identification effect.

[0005] Existing Transformer-based radar jamming signal identification methods suffer from several drawbacks. First, they only utilize features aggregated from the category vectors, failing to fully leverage the features within each radar jamming signal sampling window, thus reducing the identification rate. Second, due to the limited availability of sufficient radar jamming signal training samples in practical applications, the model may overfit, resulting in poor generalization ability and weak robustness. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art.

[0007] Therefore, the purpose of this invention is to propose a radar jamming signal identification method based on Transformer, which solves the problem of low identification accuracy caused by the inability of existing radar jamming signal identification methods to fully and effectively utilize the characteristics of radar jamming signals.

[0008] To achieve the above objectives, a first aspect of the present invention proposes a radar jamming signal identification method based on Transformer, comprising:

[0009] Construct a radar jamming signal time-domain dataset, and process the radar jamming signal time-domain dataset to obtain a first training dataset;

[0010] The sampling points of the samples in the first training dataset are fused to obtain the second training dataset;

[0011] The first training dataset and the second training dataset are concatenated to obtain the third training dataset; the third training dataset is input into the Transformer model for feature extraction. The loss function of the Transformer model is the sum of the loss function corresponding to the features contained in the class vector and the loss function corresponding to the features contained in each sampling window. Then, the model parameters that minimize the loss of the Transformer model are obtained through an optimization algorithm.

[0012] T sampled training sets are obtained by using a sampling with replacement strategy on the third training dataset. T Transformer models are trained using the sampled training sets and the model parameters. The test dataset is input into the trained T Transformer models, and the recognition result is obtained by voting.

[0013] In addition, the radar interference signal identification method based on Transformer according to the above embodiments of the present invention may also have the following additional technical features:

[0014] Furthermore, in one embodiment of the present invention, constructing the radar jamming signal time-domain dataset includes:

[0015] The simulation includes time-domain datasets of multiple types of radar jamming signals. Each type of jamming signal includes a preset number of samples. Each sample is stored as a concatenation of real and imaginary sampling points. Specifically, the first 50% of the vector corresponding to each sample is the real sampling points of the sample, and the last 50% is the imaginary sampling points of the sample.

[0016] Furthermore, in one embodiment of the present invention, processing the radar jamming signal time-domain dataset to obtain the first training dataset includes:

[0017] The radar jamming signal time-domain dataset is divided into three non-overlapping sets according to a predetermined ratio: a training dataset, a validation dataset, and a test dataset.

[0018] Data preprocessing is performed on each sample in the time-domain dataset, including dividing the sampling points in the sample into sampling windows according to a predetermined window size.

[0019] Furthermore, in one embodiment of the present invention, after obtaining the second training dataset, the method further includes:

[0020] Calculate the correlation between the first training dataset and the second training dataset, and use the correlation coefficient as the smoothing coefficient in the dynamic label smoothing operation to perform dynamic label smoothing operation on the label of each sample in the second training dataset to obtain the second label.

[0021] Furthermore, in one embodiment of the present invention, after obtaining the second tag, the method further includes:

[0022] The labels corresponding to the third training dataset are obtained by concatenating the labels corresponding to the first training dataset with the second labels.

[0023] Further, in one embodiment of the present invention, the third training dataset is input into a Transformer model for feature extraction. The loss function of the Transformer model is the sum of the loss functions corresponding to the features contained in the class vector and the loss functions corresponding to the features contained in each sampling window. Then, the model parameters that minimize the network loss are obtained through an optimization algorithm, including:

[0024] The Transformer model is used to extract features from the third training dataset. The extracted features are connected to a fully connected layer to produce a linear output and generate a loss function. When calculating the loss function, the loss function corresponding to the category vector is added to the loss function corresponding to each sampling window to form the final loss function of the model.

[0025] The gradient of the final loss function with respect to each parameter in the model is calculated using the chain rule of derivatives during backpropagation. The network parameters are then continuously updated based on the gradient using an optimization algorithm to find the model parameters that minimize the loss of the Transformer model.

[0026] Furthermore, in one embodiment of the present invention, the step of inputting the test dataset into T trained Transformer models and obtaining the recognition result through voting includes:

[0027] The test dataset is input in parallel into T trained Transformer models to obtain T radar interference signal identification results. The identification results of each individual learner are voted on and output using a relative majority voting method to obtain the final identification result.

[0028] To achieve the above objectives, a first aspect of the present invention provides a radar jamming signal identification device based on Transformer, comprising:

[0029] A construction module is used to construct a radar jamming signal time-domain dataset and process the radar jamming signal time-domain dataset to obtain a first training dataset.

[0030] The fusion module is used to fuse the sampling points between samples in the first training dataset to obtain the second training dataset;

[0031] The optimization module is used to concatenate the first training dataset and the second training dataset to obtain a third training dataset; input the third training dataset into the Transformer model for feature extraction, wherein the loss function corresponding to the Transformer model is the sum of the loss function corresponding to the features contained in the class vector and the loss function corresponding to the features contained in each sampling window; and then obtain the model parameters that minimize the loss of the Transformer model through the optimization algorithm.

[0032] The recognition module is used to obtain T sampled training sets by using a sampling with replacement strategy on the third training dataset, train T Transformer models using the sampled training sets and the model parameters, input the test dataset into the trained T Transformer models, and obtain the recognition result by voting.

[0033] To achieve the above objectives, a third aspect of the present invention provides a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the radar jamming signal identification method based on Transformer as described above.

[0034] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the Transformer-based radar jamming signal identification method as described above.

[0035] The radar jamming signal identification method based on Transformer in this invention first mitigates overfitting by using soft-label data augmentation. Then, it fully utilizes features contained in the category vectors and features within the sampling window. Finally, by constructing a homogeneous ensemble through sampling with replacement on the training set, the accuracy of radar jamming signal identification is further improved by leveraging the differences between individual learners. Compared to existing radar jamming signal identification methods, this proposed method enhances the accuracy of radar jamming signal identification, overcomes the limitation of existing methods in obtaining long-distance sampling point dependencies, and improves the model's generalization ability. Attached Figure Description

[0036] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0037] Figure 1 This is a flowchart illustrating a radar interference signal identification method based on Transformer, provided in an embodiment of the present invention.

[0038] Figure 2 The time-domain waveform diagram of the blocking interference provided in the embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram of the fusion process of blocking interference and frequency sweeping interference sampling points provided in an embodiment of the present invention.

[0040] Figure 4 This is a schematic diagram of a model for a radar interference signal identification method based on Transformer provided in an embodiment of the present invention.

[0041] Figure 5 This is a schematic diagram of homogeneous integration provided for an embodiment of the present invention.

[0042] Figure 6 This is a schematic flowchart of a radar interference signal identification device based on Transformer, provided for an embodiment of the present invention. Detailed Implementation

[0043] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0044] The radar jamming signal identification method based on Transformer according to embodiments of the present invention is described below with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart illustrating a radar interference signal identification method based on Transformer provided in an embodiment of the present invention.

[0046] like Figure 1 As shown, the radar jamming signal identification method based on Transformer includes the following steps:

[0047] S101: Construct a time-domain dataset of radar jamming signals and process the time-domain dataset of radar jamming signals to obtain the first training dataset;

[0048] S102: Perform sampling point fusion between samples in the first training dataset to obtain the second training dataset;

[0049] S103: Concatenate the first training dataset with the second training dataset to obtain the third training dataset; input the third training dataset into the Transformer model for feature extraction. The loss function of the Transformer model is the sum of the loss function corresponding to the features contained in the class vector and the loss function corresponding to the features contained in each sampling window. Then, the model parameters that minimize the loss of the Transformer model are obtained through the optimization algorithm.

[0050] S104: T sampled training sets are obtained by using a sampling with replacement strategy on the third training dataset. T Transformer models are trained using the sampled training sets and model parameters. The test dataset is input into the trained T Transformer models, and the recognition result is obtained by voting.

[0051] Furthermore, in one embodiment of the present invention, constructing a radar jamming signal time-domain dataset includes:

[0052] A time-domain dataset including multiple types of radar jamming signals was simulated. Each type of jamming signal includes a preset number of samples. Each sample is stored as a concatenation of real and imaginary sampling points. Specifically, the first 50% of the vector corresponding to each sample is the real sampling points of the sample, and the last 50% is the imaginary sampling points of the sample.

[0053] Specifically, this invention first analyzes the composition and current status of the complex electromagnetic environment in which the radar system operates. It then simulates a time-domain dataset containing 12 types of radar jamming signals, simulating an actual battlefield environment. Each type of jamming signal has 500 samples. Each sample is stored as a concatenation of real and imaginary sampling points; that is, the first 50% of the vector corresponding to each sample represents the real sampling points, and the last 50% represents the imaginary sampling points. This time-domain dataset covers pure noise, suppression jamming (suppression jamming is further divided into aiming jamming, frequency sweeping jamming, and jamming), deception jamming (deception jamming is further divided into range deception jamming and dense false target jamming), novel jamming (novel jamming is further divided into intermittent sampling and forwarding jamming and smart noise jamming), chaff jamming, and composite jamming (the composite method is additive composite, and composite jamming is further divided into dense false target jamming + smart noise jamming, chaff jamming + intermittent sampling and forwarding jamming, and range deception jamming + frequency sweeping jamming). Figure 2 The time-domain waveform of the blocking interference is shown in the constructed time-domain dataset.

[0054] Furthermore, in one embodiment of the present invention, processing the radar jamming signal time-domain dataset to obtain a first training dataset includes:

[0055] The radar jamming signal time-domain dataset is divided into three non-overlapping sets according to a predetermined ratio: the training dataset, the validation dataset, and the test dataset.

[0056] Data preprocessing is performed on each sample in the time-domain dataset, including dividing the sampling points in the sample into sampling windows according to a predetermined window size.

[0057] Specifically, the radar jamming signal time-domain dataset is first randomly divided into three disjoint sets in a 3:1:1 ratio: the training dataset, the validation dataset, and the test dataset. The training dataset is denoted as X (X={x1,x2,...,x...). n}, x n Let n be the number of samples in the training set X, and let Y be the label of the training dataset (Y = {y1, y2, ..., y3}). n}, y n (The labels are the corresponding to the samples in the training set X). Then, data preprocessing is performed on each sample in the time-domain dataset, that is, the sampling points in the sample are divided into sampling windows according to a certain window size. The total length of the sampling points is s, and the window size is w. Then the number of sampling windows N is N = s / w.

[0058] Furthermore, in one embodiment of the present invention, after obtaining the second training dataset, the method further includes:

[0059] Calculate the correlation between the first training dataset and the second training dataset, and use the correlation coefficient as the smoothing coefficient in the dynamic label smoothing operation to perform dynamic label smoothing operation on the label of each sample in the second training dataset to obtain the second label.

[0060] Furthermore, in one embodiment of the present invention, after obtaining the second tag, the method further includes:

[0061] The labels for the third training dataset are obtained by concatenating the labels corresponding to the first training dataset with the second labels.

[0062] Specifically, the obtained training dataset X is used to perform cross-sample analysis (sample x... A and sample x B The sampling points of the sampled data are fused, and the new training set formed after the sampling point fusion is denoted as X1(X1={x1',x2',...,x...). n '}, x n 'The samples in the training set X1 after sampling point fusion' are denoted as Y1 (Y1 = {y1', y2', ..., y...). n '},y n 'The label is the sample corresponding to the training set X1.' Figure 3As shown, the sampling point fusion process between samples is illustrated using blocking interference and frequency sweeping interference as examples. During sampling point fusion, a random sampling point fusion region B is selected. For each sample, its original label is one-hot encoded, meaning that the label vector in each sample has a target class probability of 1 only, while other non-target class probabilities are 0. Therefore, the label corresponding to the new sample after sampling point fusion should be y' = λy. A +(1-λ)y B , where λ is the ratio of the sampling point length l(B) of the fusion region to the total sampling point length s in the sample.

[0063] The correlation between the original training set X before sampling point fusion and the new dataset X1 formed after sampling point fusion is calculated, and the obtained correlation coefficient is used as the smoothing coefficient ε in the dynamic label smoothing operation. The dynamic label smoothing operation operates on the label of each sample in the training set X1. Based on the sampling point fusion in step three, the dynamic label smoothing operation changes the target class probability to 1 minus the smoothing coefficient, while the probabilities of other non-target classes follow a normal distribution. It is worth noting that the sum of the target class probability and the non-target class probability should be 1. The label after the dynamic label smoothing operation is denoted as Y2 (Y2 = {y1”, y2”, ..., y n ”},y n "The labels are the samples in the training set X1".

[0064] Furthermore, in one embodiment of the present invention, a third training dataset is input into a Transformer model for feature extraction. The loss function corresponding to the Transformer model is the sum of the loss functions corresponding to the features contained in the class vector and the loss functions corresponding to the features contained in each sampling window. Then, the model parameters that minimize the network loss are obtained through an optimization algorithm, including:

[0065] The Transformer model is used to extract features from the third training dataset. The extracted features are connected to a fully connected layer to produce a linear output and generate a loss function. When calculating the loss function, the loss function corresponding to the class vector is added to the loss function corresponding to each sampling window to form the final loss function of the model.

[0066] The gradient of the final loss function with respect to each parameter in the model is calculated using the chain rule of derivatives during backpropagation. The network parameters are then continuously updated based on the gradient using an optimization algorithm to find the model parameters that minimize the loss of the Transformer model.

[0067] Specifically, such as Figure 4 As shown, the training set X2 (X2={x1,x2,...,x...) is formed by concatenating the original training set X with the samples in the training set X1 obtained after step three. n,x1',x2',...,x n '}, x n For samples in the training set X before sampling point fusion, x n The samples in the training set X1 after sampling point fusion are input into the Transformer model, and their corresponding labels are Y and Y2, which are the final labels Y3 formed by concatenating them (Y3 = {y1, y2, ..., y2}). n ,y1”,y2”,...,y n ”},y n Let y be the label corresponding to the sample in the training set X before sampling point fusion. n "The labels are those corresponding to the samples in the training set X1 after dynamic label smoothing following sampling point fusion." The Transformer model consists of five stacked identical Transformer blocks. Each Transformer block comprises a multi-head attention mechanism, Dropout, residuals, Layer Normalization, and a fully connected layer. After extracting features from the time-domain data of radar interference signals using the Transformer, the features output from the class vector and the features output from each sampling window are connected to a fully connected layer for linear output and to generate a loss function. When calculating the loss function, the loss function L0 corresponding to the class vector is added to the loss function Lm (m = 1, ..., N, as described in step two, where N is the number of sampling windows and N = s / w) corresponding to each sampling window to form the final loss function of the model, which is denoted as L. Then, during backpropagation, the gradient of the loss function L with respect to each parameter in the model is calculated using the derivative chain rule. The network parameters are continuously updated based on the gradient using an optimization algorithm to find the parameters that minimize the network loss.

[0068] Furthermore, in one embodiment of the present invention, the test dataset is input into T trained Transformer models, and the recognition result is obtained through voting, including:

[0069] The test dataset is input into T trained Transformer models in parallel to obtain T radar interference signal recognition results. The recognition results of each individual learner are voted on and output using a relative majority voting method to obtain the final recognition result.

[0070] like Figure 5 As shown, the present invention proposes specific operational steps to improve the accuracy of radar interference signal identification by constructing a homogeneous ensemble through sampling the training set with replacement, and then utilizing the differences between individual learners.

[0071] Specifically, this invention utilizes the differences between individual learners to construct "good but different" individual learners, further improving the accuracy of radar interference signal identification based on Transformer. A sampling with replacement strategy is applied to the training set X2 generated in step four, resulting in T independent sampled training sets XX1,...,XXT. These sampled training sets are then used to train T Transformer models. Because the T sampled training sets obtained through sampling with replacement have certain differences, the trained T individual learners also have certain differences. This invention improves the robustness of the system by constructing a homogeneous ensemble.

[0072] The test dataset is input into T trained Transformer models in parallel to obtain T radar interference signal recognition results. The recognition results of each individual learner are voted on and output using a relative majority voting method to obtain the final recognition result.

[0073] The radar jamming signal identification method based on Transformer in this invention first mitigates overfitting by using soft-label data augmentation. Then, it fully utilizes features contained in the category vectors and features within the sampling window. Finally, by constructing a homogeneous ensemble through sampling with replacement on the training set, the accuracy of radar jamming signal identification is further improved by leveraging the differences between individual learners. Compared to existing radar jamming signal identification methods, this proposed method enhances the accuracy of radar jamming signal identification, overcomes the limitation of existing methods in obtaining long-distance sampling point dependencies, and improves the model's generalization ability.

[0074] Figure 6 This is a schematic diagram of a radar interference signal identification device based on Transformer, provided as an embodiment of the present invention.

[0075] like Figure 6 As shown, the Transformer-based radar jamming signal identification device includes: a construction module 100, a fusion module 200, an optimization module 300, and an identification module 400, wherein...

[0076] The module is used to build a time-domain dataset of radar jamming signals and process the time-domain dataset of radar jamming signals to obtain the first training dataset.

[0077] The fusion module is used to fuse the sampling points between samples in the first training dataset to obtain the second training dataset.

[0078] The optimization module is used to concatenate the first training dataset with the second training dataset to obtain the third training dataset; the third training dataset is input into the Transformer model for feature extraction. The loss function of the Transformer model is the sum of the loss function corresponding to the features contained in the class vector and the loss function corresponding to the features contained in each sampling window. Then, the optimization algorithm is used to obtain the model parameters that minimize the loss of the Transformer model.

[0079] The recognition module is used to obtain T sampled training sets by using a sampling strategy with replacement on the third training dataset, and to train T Transformer models using the sampled training sets and model parameters; the test dataset is input into the trained T Transformer models, and the recognition result is obtained by voting.

[0080] To achieve the above objectives, a third aspect of the present invention provides a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the radar jamming signal identification method based on Transformer as described above.

[0081] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the Transformer-based radar jamming signal identification method as described above.

[0082] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0083] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0084] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for radar jamming signal recognition based on Transformer, characterized in that, Includes the following steps: Construct a radar jamming signal time-domain dataset, and process the radar jamming signal time-domain dataset to obtain a first training dataset; The sampling points of the samples in the first training dataset are fused to obtain the second training dataset; The first training dataset and the second training dataset are concatenated to obtain the third training dataset; the third training dataset is input into the Transformer model for feature extraction. The loss function of the Transformer model is the sum of the loss function corresponding to the features contained in the class vector and the loss function corresponding to the features contained in each sampling window. Then, the model parameters that minimize the loss of the Transformer model are obtained through an optimization algorithm. T sampled training sets are obtained by using a sampling with replacement strategy on the third training dataset. T Transformer models are trained using the sampled training sets and the model parameters. The test dataset is input into the trained T Transformer models, and the recognition result is obtained by voting.

2. The method according to claim 1, wherein constructing the radar jamming signal time-domain dataset comprises: The simulation includes time-domain datasets of multiple types of radar jamming signals. Each type of jamming signal includes a preset number of samples. Each sample is stored as a concatenation of real and imaginary sampling points. Specifically, the first 50% of the vector corresponding to each sample is the real sampling points of the sample, and the last 50% is the imaginary sampling points of the sample.

3. The method according to claim 1, wherein processing the radar jamming signal time-domain dataset to obtain the first training dataset includes: The radar jamming signal time-domain dataset is divided into three non-overlapping sets according to a predetermined ratio: a training dataset, a validation dataset, and a test dataset. Data preprocessing is performed on each sample in the time-domain dataset, including dividing the sampling points in the sample into sampling windows according to a predetermined window size.

4. The method according to claim 1, further comprising, after obtaining the second training dataset: Calculate the correlation between the first training dataset and the second training dataset, and use the correlation coefficient as the smoothing coefficient in the dynamic label smoothing operation to perform dynamic label smoothing operation on the label of each sample in the second training dataset to obtain the second label.

5. The method according to claim 1 or 4, further comprising, after obtaining the second label: The labels corresponding to the third training dataset are obtained by concatenating the labels corresponding to the first training dataset with the second labels.

6. The method according to claim 1, wherein the step of inputting the third training dataset into the Transformer model for feature extraction, wherein the loss function corresponding to the Transformer model is the sum of the loss functions corresponding to the features contained in the class vector and the loss functions corresponding to the features contained in each sampling window, and then obtaining the model parameters that minimize the loss of the Transformer model through an optimization algorithm, includes: The Transformer model is used to extract features from the third training dataset. The extracted features are connected to a fully connected layer to produce a linear output and generate a loss function. When calculating the loss function, the loss function corresponding to the category vector is added to the loss function corresponding to each sampling window to form the final loss function of the model. The gradient of the final loss function with respect to each parameter in the model is calculated using the chain rule of derivatives during backpropagation. The network parameters are then continuously updated based on the gradient using an optimization algorithm to find the model parameters that minimize the loss of the Transformer model.

7. The method according to claim 1, wherein inputting the test dataset into T trained Transformer models and obtaining the recognition result through voting includes: The test dataset is input in parallel into T trained Transformer models to obtain T radar interference signal identification results. The identification results of each individual learner are voted on and output using a relative majority voting method to obtain the final identification result.

8. A radar jamming signal identification device based on Transformer, characterized in that, Includes the following modules: A construction module is used to construct a radar jamming signal time-domain dataset and process the radar jamming signal time-domain dataset to obtain a first training dataset. The fusion module is used to fuse the sampling points between samples in the first training dataset to obtain the second training dataset; The optimization module is used to concatenate the first training dataset and the second training dataset to obtain a third training dataset; input the third training dataset into the Transformer model for feature extraction, wherein the loss function corresponding to the Transformer model is the sum of the loss function corresponding to the features contained in the class vector and the loss function corresponding to the features contained in each sampling window; and then obtain the model parameters that minimize the loss of the Transformer model through the optimization algorithm. The recognition module is used to obtain T sampled training sets by using a sampling with replacement strategy on the third training dataset, train T Transformer models using the sampled training sets and the model parameters, input the test dataset into the trained T Transformer models, and obtain the recognition result by voting.

9. A computer device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the radar jamming signal identification method based on any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the radar jamming signal identification method based on any one of claims 1-6.

Images