Radar jamming countermeasure situation prediction method based on LSTM network

Abstract

The application discloses a radar jamming countermeasure situation prediction method based on an LSTM network, mainly solves the problems that the existing Bayesian network is too dependent on expert experience, the single-structure neural network is slow in processing a large amount of time sequence data, and the jamming opportunity is not considered to adjust the jamming strategy according to the working state of the radar in the actual scene. The implementation scheme is: constructing a radar working parameter and working mode data set and an active jamming style data set, and adding the radar working mode into the active jamming style data set; constructing a Transform-LSTM network and training the same by using a training set; and using the trained network to predict the radar jamming countermeasure situation. The application can process the whole sequence in parallel, improves the calculation efficiency and comprehensiveness; meanwhile, the radar working mode is included in the active jamming style data set, is more in line with the actual situation, further improves the prediction applicability of the system, and can be used for machine learning, deep learning and jamming signal analysis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of situation awareness, and particularly relates to a radar jamming countermeasure situation prediction method, which can be used for machine learning, deep learning and jamming signal analysis. BACKGROUND

[0002] The research on radar jamming situation prediction technology mainly includes two parts: radar working mode prediction and active jamming pattern prediction. Radar working mode prediction is a key technology in electronic warfare and radar reconnaissance, which involves extracting features from radar signals and using these features to predict the working state and mode of the radar, such as search, tracking or fire control mode. The research in this field mainly focuses on feature extraction, application of pattern recognition technology, and real-time data processing. In terms of feature extraction and pattern recognition, traditional feature extraction techniques focus on time-frequency analysis, statistical feature analysis and other methods. Real-time is another key research direction of radar working mode prediction. Researchers are committed to developing efficient algorithms and techniques to reduce the time required for data processing and analysis. GPU acceleration, optimized algorithm processes and other technologies are applied to improve processing speed.

[0003] The prediction of active jamming pattern mainly focuses on predicting the type and characteristics of the jamming source through the analysis of jamming signals. This technology plays a crucial role in electronic countermeasures and jamming management. Through methods such as modulation recognition and spectral analysis, researchers can extract key features from jamming signals. These features help identify the type and characteristics of the jamming source. In addition to traditional signal processing techniques, emerging signal analysis methods such as machine learning and deep learning are also applied to the analysis of jamming signals. These methods can handle more complex jamming environments, improving the accuracy and efficiency of prediction. The application of machine learning and deep learning technology is an important progress in the field of active jamming pattern prediction. By constructing deep neural networks, researchers can learn deep features from complex jamming signals, improving the accuracy of jamming pattern recognition.

[0004] A "parallel inference multi-branch situation prediction method based on XGBoost" is disclosed in a patent application document with application number CN202211735137.2 by the China Electronics Technology Group Corporation Information Science Research Institute. This method combines the XGBoost model with the situation prediction method of Bayesian network. First, the situation data is sorted by time series, preprocessed and distributed. Second, the XGBoost model is trained based on the processing results to predict the future situation probability. Then, the Bayesian network is constructed to evaluate the countermeasures, and the multi-branch situation prediction is output. Although this method uses XGBoost algorithm to combine the combat effectiveness evaluation, and the accurate prediction can be obtained through the situation probability and countermeasures, the nodes and conditional probability of the Bayesian network are determined by expert analysis, which relies too much on expert experience, and may affect the accuracy of the model due to subjective factors, resulting in inaccurate prediction.

[0005] In his thesis "Research on electromagnetic situation visualization and prediction method" (Master's thesis, Hangzhou University of Electronic Science and Technology, June 2020), Shi Hongxin introduces a situation prediction method based on GRU-BP neural network. The steps of this method include: extracting features from original historical data to construct training set and performing normalization; initializing GRU network to obtain accurate GRU model through training; using GRU output as BP network input, and obtaining GRU-BP combined model through iterative training; inputting test data into GRU-BP model, and calculating model accuracy by comparing with samples. This method mainly analyzes historical trajectory data and improves prediction accuracy by combining BP and GRU networks. However, it does not fully consider various factors affecting the situation change of both sides, such as equipment working mode, which limits the applicability of the prediction technology and has poor reliability. SUMMARY

[0006] The purpose of the present application is to overcome the shortcomings of the prior art and provide a radar jamming countermeasure situation prediction method based on LSTM network to improve the applicability and reliability of the system and further improve the accuracy of radar jamming countermeasure situation prediction.

[0007] The technical idea for achieving the object of the present application is that: by combining the feature extraction of the Transformer model and the sequence data processing capability of the LSTM network, a Transformer-LSTM network model is constructed to realize the rapid parallel processing and complex pattern learning of data, overcome the limitations of the structure of a single network, and further improve the prediction accuracy of the network. By comprehensively considering the influence of various factors such as radar working parameters, working modes, and jammer jamming patterns on the situation change, the problem of incomplete analysis of the situation prediction result caused by single analysis of influencing factors in the existing situation prediction method is solved, and the comprehensiveness and reliability of the system prediction are improved; at the same time, by including the radar working mode in the data set of the active jamming pattern, it is more in line with the actual scene, and the applicability of the situation prediction technology is enhanced.

[0008] According to the above idea, the technical scheme of the present application includes the following steps:

[0009] (1) predicting the radar working parameters and working modes;

[0010] (1a) constructing the training set and test set of the radar working parameters and working modes:

[0011] Randomly generate thousands of radar pulse signal samples, select five characteristic parameters of carrier frequency, pulse width, pulse repetition interval, pulse amplitude and instantaneous bandwidth from each sample, and normalize and label the characteristic parameters of each sample;

[0012] All normalized samples and their corresponding state labels are combined to form a radar working parameter and working mode prediction data set, 70% of the samples are used as the training set, and 30% of the samples are used as the test set;

[0013] (1b) constructing a Transformer-LSTM network:

[0014] Selecting an LSTM layer and configuring it with 15 input feature dimensions and 32 hidden units;

[0015] Selecting a Transformer encoder layer and configuring it with 32 feature dimensions and 4 attention heads;

[0016] The sequence input layer, LSTM layer, Transformer encoder layer, flattening layer, fully connected layer, and regression output layer are connected in sequence to form a Transformer-LSTM network;

[0017] (1c) training the Transformer-LSTM network:

[0018] Divide the training set data into groups of 15 consecutive time points each and input them into the Transformer-LSTM network model.

[0019] The initial learning rate of 0.001 and the maximum number of iterations are set, the iterative update of the network parameters is carried out through the Adam optimizer, until the maximum number of iterations is reached, and the trained Transformer-LSTM network is obtained;

[0020] (1d) input the test set into the trained Transformer-LSTM network model, and perform inverse normalization processing on the output result to obtain the prediction result of the radar working parameter and the working mode;

[0021] (2) predicting the active jamming pattern;

[0022] (2a) randomly generating thousands of jamming signal samples, selecting signal center frequency, signal duration and jamming signal bandwidth as three characteristic parameters, constructing an active jamming pattern prediction data set, and dividing it into a training set and a test set according to a ratio of 7:3;

[0023] (2b) using a reinforcement learning algorithm based on Q-Learning to add the radar working mode to the active jamming pattern data set;

[0024] (2c) input the active jamming pattern training set into the Transformer-LSTM network, and train the network by using a method similar to step 1c);

[0025] (2d) input the active jamming pattern test set into the trained Transformer-LSTM network model, and perform inverse normalization processing on the output result to obtain the prediction result of the active jamming pattern.

[0026] Compared with the prior art, the present application has the following advantages:

[0027] Firstly, the Transformer-LSTM network model constructed by combining LSTM and Transformer model in the present application can utilize the advantage of LSTM in processing local dependency in sequence, and the advantage of Transformer in effectively capturing long-distance dependency through self-attention mechanism, which helps to learn more complex patterns in sequence and improves the calculation efficiency; at the same time, the multi-head attention of Transformer allows the model to learn information from different representation subspaces, which can enrich the understanding of sequence data by the model, thereby improving the prediction accuracy of the network.

[0028] Secondly, the present application realizes the prediction of radar working parameters, working states and jammer jamming modes by selecting five radar signal characteristic parameters and three jamming signal characteristic parameters, overcomes the problem of imperfect analysis of situation prediction results caused by only considering single influencing factor in the existing situation prediction method, and enhances the comprehensiveness and reliability of the situation prediction technology.

[0029] Thirdly, the present application considers that the jammer selects a suitable jamming strategy according to the working state of the radar, and innovatively integrates the radar working mode into the data set of active jamming modes, which is more in line with the actual scene and further improves the prediction applicability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The implementation block diagram of the present application. DETAILED DESCRIPTION

[0031] In order to enable the personnel in the art to better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, other embodiments obtained by the personnel in the art without creative labor should belong to the protection scope of the present application.

[0032] It should be noted that the step numbers in the specification and claims of the present application are only for clearly describing the embodiments of the present application, facilitating understanding, and the sequence of the numbers is not limited.

[0033] Referring to Figure 1 The implementation of the present embodiment includes two parts of predicting radar working parameters and working modes and predicting active jamming modes.

[0034] First, predicting radar working parameters and working modes

[0035] Step 1, constructing the training set and test set of radar working parameters and working modes.

[0036] 1.1) Selecting five characteristic parameters of carrier frequency, pulse width, pulse repetition interval, pulse amplitude and instantaneous bandwidth to form a radar parameter set:

[0037] Suppose that the working state of the radar is randomly switched among searching, tracking and guiding, and MATLAB simulation is used to generate a plurality of groups of pulse signals, each group of pulses containing a plurality of pulse strings of three working states;

[0038] Randomly selecting a pulse characteristic parameter under a working mode as a working parameter under each working state;

[0039] 1.2) Normalization of feature parameters of each sample, the formula is as follows:

[0040]

[0041] wherein, represents the b-th feature parameter in the a-th sample of the p-th working equipment in the normalized sample set, p represents the serial number of the working equipment, p = 1 represents the radar, and p = 2 represents the jammer; when p = 1, b represents the serial number of the radar feature parameter, and when p = 2, b represents the serial number of the jammer feature parameter; represents the maximum value of the b-th feature parameter of the p-th working equipment in the sample set, represents the minimum value of the b-th feature parameter of the p-th working equipment in the sample set;

[0042] The value of b represents different feature parameters of the radar and the jammer according to the serial number value of the working equipment p. When p = 1, b represents the serial number of the radar feature parameter, and different values of b represent different parameters of the radar feature:

[0043] b = 1 represents the repetition interval of the radar pulse in the sample set,

[0044] b = 2 represents the pulse width of the radar pulse in the sample set,

[0045] b = 3 represents the pulse amplitude of the radar pulse in the sample set,

[0046] b = 4 represents the carrier frequency of the radar pulse in the sample set,

[0047] b = 5 represents the bandwidth of the radar pulse in the sample set;

[0048] When p = 2, b represents the serial number of the jammer feature parameter, and different values of b represent different parameters of the jammer signal feature:

[0049] b = 1 represents the center frequency of the jamming signal in the sample set,

[0050] b = 2 represents the duration of the jamming signal in the sample set,

[0051] b = 3 represents the bandwidth of the jamming signal in the sample set;

[0052] 1.3) Labeling the normalized samples, i.e., labeling the corresponding radar working state of the samples;

[0053] 1.4) Grouping all the normalized samples and their corresponding state labels to form a radar working parameter and working mode prediction data set, and taking 70% of the samples as a training set and 30% of the samples as a test set.

[0054] Step 2, construct a Transformer-LSTM network.

[0055] 2.1) Configure an LSTM layer with 15 input feature dimensions and 32 hidden units to extract the time-dependent features of the sequence;

[0056] 2.2) Set a Transformer encoder layer with 32 feature dimensions and 4 attention heads, divide the feature dimensions into 4 groups every 8, and cross-connect the 4 groups with the attention heads to form a self-attention mechanism and a feedforward neural network. The self-attention mechanism is used to calculate the attention of each position in the input sequence with all other positions. The feedforward neural network is used to perform nonlinear transformation of the input through multiple layers of perception mechanism;

[0057] 2.3) Set a flattening layer to convert the data stream from multi-dimensional to one-dimensional;

[0058] 2.4) Set a fully connected layer to map the data stream to a single output;

[0059] 2.5) Connect the sequence input layer, LSTM layer, Transformer encoder layer, flattening layer, fully connected layer, and regression output layer in sequence to form a Transformer-LSTM network.

[0060] Step 3, train the Transformer-LSTM network.

[0061] 3.1) Divide the training set data into groups of 15 consecutive time points each and input them into the Transformer-LSTM network model. Set the initial learning rate to 0.001 and the maximum number of iterations;

[0062] 3.2) Update the network parameters iteratively through the Adam optimizer, with the following update steps:

[0063] 3.2.1) Let g t be the gradient of the t-th iteration parameter, and β1 and β2 be the decay coefficients of the two exponential weighted average values. Calculate the first moment estimate m t and the second moment estimate v t :

[0064]

[0065] 3.2.2) Calculate the first moment estimate t and the second moment estimate t after correcting the gradient deviation and the second moment estimate

[0066]

[0067]

[0068] wherein, and denotes the attenuation coefficient of the tth iteration;

[0069] 3.2.3) Let θ t be the updated parameter, and the parameter before updating θ t+1 is obtained according to the result of step 3.2.2).

[0070]

[0071] wherein, η is the learning rate, and ε is a constant value tending to 0, used to avoid division by zero;

[0072] 3.3) Repeat step 3.2) to iteratively update the network parameters until the maximum number of iterations is reached, save the training parameters, and obtain the trained Transformer-LSTM radar operating parameter and operating mode prediction network.

[0073] Step 4, use the trained Transformer-LSTM radar operating parameter and operating mode prediction network to obtain the prediction results of radar operating parameters and operating modes.

[0074] Input the radar operating parameter and operating mode test set into the trained Transformer-LSTM network model, and perform inverse normalization processing on the output results according to the following formula to obtain the prediction results of radar operating parameters and operating modes:

[0075]

[0076] wherein, denotes the bth feature parameter in the ath sample of the pth operating equipment in the inverse normalized test sample set, and the meanings of other symbols are the same as those in the normalization formula in step 1.2).

[0077] Second, predict the active jamming pattern

[0078] Step 5, construct the training set and test set of the active jamming pattern.

[0079] 5.1) Select three feature parameters: signal center frequency, signal duration, and jamming signal bandwidth;

[0080] 5.2) Use MATLAB simulation software to generate jamming signal time domain and frequency domain waveforms under different feature parameters, and the jamming patterns include noise amplitude modulation jamming, noise frequency modulation jamming, noise phase modulation jamming, and sweep frequency jamming, and six kinds of deception jamming, such as distance multi-false target jamming and intermittent sampling and retransmission jamming.

[0081] 5.3) Normalize the feature parameters of each sample, the method is the same as step 1.2);

[0082] 5.4) Label the normalized samples, that is, label the active jamming pattern corresponding to the sample, and then group all the normalized samples and their corresponding state labels into an active jamming pattern data set.

[0083] Step 6, using a reinforcement learning algorithm based on Q-Learning to add radar operating mode to the active jamming pattern data set.

[0084] 6.1) Define the operating state of the radar as the state of the learning environment, and define the possible jamming patterns as the actions available to the learning agent;

[0085] 6.2) According to the characteristics of the jamming pattern selection affected by the radar operating state, use the current Q table in the Q-Learning algorithm and the epsilon-greedy strategy to determine whether to adopt a new jamming pattern or maintain the current jamming pattern, and realize the optimization of the selection of the jamming pattern;

[0086] 6.3) According to the reward matrix in the Q-Learning algorithm, iteratively update the Q table, and finally generate a Q table reflecting the optimal jamming pattern selection strategy, and then input the radar operating mode data into the updated Q table, and output the active jamming pattern, that is, realize the addition of the radar operating mode to the active jamming pattern data set;

[0087] 6.4) Divide the active jamming pattern data set with radar operating mode into an active jamming pattern training set and an active jamming pattern test set in a ratio of 7:3.

[0088] Step 7, input the active jamming pattern training set into the Transformer-LSTM network, and train the network using a method similar to step 3 to obtain a trained Transformer-LSTM active jamming pattern prediction network model.

[0089] Step 8, input the active jamming pattern test set into the trained Transformer-LSTM active jamming pattern prediction network model, and perform reverse normalization processing on the output results to obtain the prediction results of the active jamming pattern.

[0090] At this point, the prediction results of the radar operating parameters and operating mode, and the prediction results of the active jamming pattern are obtained, and the radar jamming countermeasure situation prediction is completed.

[0091] The effect of the present application will be further described in combination with a simulation experiment:

[0092] 1. Simulation experiment conditions:

[0093] The hardware platform of the simulation experiment is: the processor is 12th Gen Intel(R) Core(TM) i7-12700H CPU, the main frequency is 2.3GHz, and the memory is 16GB.

[0094] The software platform of the simulation experiment is: Windows 11 operating system, MATLAB 2022b, PyCharm 2023.2.5, CUDA 12.1, python 3.10.9, torch 2.0.0, and torchvision 0.15.1.

[0095] The radar pulse parameter settings are: the pulse repetition interval range is 30-150us, the modulation mode is sliding, jitter, and staggered, the pulse width range is 0.03-3.0us, the pulse amplitude range is 0.5-3.0, the carrier frequency range is 2.0-2.5GHz, and the signal bandwidth range is 50-100MHz. The jammer interference types include noise amplitude modulation, noise frequency modulation, noise phase modulation, sweep frequency interference, range multi-false target interference, and intermittent sampling and forwarding interference.

[0096] The training set used in the simulation experiment is the data set generated by steps 1 and 6 of the specific embodiments of the application, including 10000 samples of radar and jammer 2 types of working equipment and corresponding state labels, divided into training set and test set in the ratio of 7:3, the training set is used to train the Transformer-LSTM network, and the test set is used to evaluate the prediction effect of the application.

[0097] 2. Simulation experiment evaluation index:

[0098] In order to illustrate the effect of the application, the root mean square error RMSE, mean square error MSE, mean absolute error MAE, and mean absolute percentage error MAPE are used as evaluation indexes to evaluate the prediction effect of the application method and the existing technology LSTM network prediction algorithm. The calculation formulas of each index are as follows:

[0099] Root mean square error This index is an important index for measuring the difference between the classification recognition result and the actual category. The smaller the RMSE, the more accurate the prediction model can reflect the experimental data. Wherein, y i represents the true value of the test set sample, represents the predicted value of the test set sample, and l represents the total number of samples in the test set.

[0100] Mean square error This index reflects the difference between the estimated value and the estimated value, which is obtained by calculating the square difference between the true value and the predicted value and then averaging.

[0101] Mean Absolute Error This index is the average of the absolute difference between the predicted value and the true value, used to evaluate the error size of the prediction result. The smaller the mean absolute error, the closer the prediction result is to the true value, and the better the prediction result of the network model.

[0102] Mean Absolute Percentage Error This index represents a relative measure, obtained by calculating the ratio of the difference between the true value and the predicted value to the true value and then averaging it.

[0103] 3. Simulation content and its result analysis:

[0104] In the simulation experiment, the LSTM time series prediction algorithm proposed by Meng Chengzhen in his master's degree thesis "Parallelization Research on Time Series Prediction Algorithm Based on LSTM" (University of Electronic Science and Technology, 2020.) is used as a comparative existing technology.

[0105] Simulation 1, under the above simulation conditions, the radar operating parameters and operating modes are predicted by the present invention and the existing LSTM prediction network algorithm respectively, and the evaluation indexes are calculated, the results are shown in Table 1.

[0106] Table 1 Comparison of radar pulse parameter prediction evaluation indexes

[0107]

[0108] From Table 1, for the prediction problem of radar operating parameters and operating states, the root mean square error RMSE, mean square error MSE, mean absolute error MAE, and mean absolute percentage error MAPE of the prediction using the Transformer-LSTM network are all lower than those of the existing LSTM network prediction method, indicating that the present invention has better accuracy in predicting radar operating parameters and operating states.

[0109] Simulation 2, under the above simulation conditions, the active jamming pattern is predicted by the present invention and the existing LSTM network respectively, and the size of each evaluation index is calculated, the results are shown in Table 2:

[0110] Table 2 Comparison of jammer jamming pattern prediction evaluation indexes

[0111] Evaluation metrics Existing LSTM network Inventive Transformer-LSTM network Performance gain / % RMSE 0.9973 0.9131 8.4428 MSE 0.9946 0.8337 16.1774 MAE 0.6459 0.5144 20.3592 MAPE 0.3038 0.2513 17.2811

[0112] From table 2, for the prediction problem of active jamming pattern, the four evaluation indexes of root mean square error RMSE, mean square error MSE, mean absolute error MAE and mean absolute percentage error MAPE of the prediction by the application using the Transform-LSTM network are all lower than those of the existing LSTM network prediction method, indicating that the application can more accurately predict the active jamming pattern.

[0113] In summary, the application can accurately predict radar working parameters, working states and active jamming patterns, and can play a good effect in the prediction of enemy radar working parameters, working states and active jamming patterns in actual battlefield scenarios.

Claims

1. A radar jamming countermeasure situation prediction method based on LSTM network, characterized in that, Includes the following steps: (1) Predicting radar operating parameters and operating modes; (1a) Construct training and testing sets for radar operating parameters and modes: Thousands of radar pulse signal samples are randomly generated. Five characteristic parameters are selected from each sample: carrier frequency, pulse width, pulse repetition interval, pulse amplitude, and instantaneous bandwidth. The characteristic parameters of each sample are then normalized and labeled. All normalized samples and their corresponding state labels are combined to form a radar operating parameter and operating mode prediction dataset. 70% of the samples are used as the training set and 30% of the samples are used as the test set. (1b) Constructing the Transformer-LSTM network: An LSTM layer was selected and configured with 15 input feature dimensions and 32 hidden units. The Transformer encoder layer was selected and configured with 32 feature dimensions and 4 attention heads; The sequence input layer, LSTM layer, Transformer encoder layer, flattening layer, fully connected layer, and regression output layer are connected in series to form a Transformer-LSTM network. (1c) Training the Transformer-LSTM network: The training set data is divided into groups of 15 consecutive time points and input into the Transformer-LSTM network model; Set an initial learning rate of 0.001 and a maximum number of iterations, and iteratively update the network parameters using the Adam optimizer until the maximum number of iterations is reached to obtain a trained Transformer-LSTM network. (1d) Input the test set into the trained Transformer-LSTM network model, and perform inverse normalization on its output to obtain the prediction results of radar operating parameters and operating modes; (2) Predict the patterns of active interference; (2a) Randomly generate thousands of interference signal samples, select three feature parameters: signal center frequency, signal duration, and interference signal bandwidth, and construct an active interference pattern dataset; (2b) The radar operating modes were added to the active jamming pattern dataset using a Q-Learning-based reinforcement learning algorithm, and the dataset was divided into a training set and a test set in a 7:3 ratio. (2c) Input the active interference pattern training set into the Transformer-LSTM network and train the network using the method in step 1c); (2d) Input the active interference pattern test set into the trained Transformer-LSTM network model, and perform inverse normalization on its output to obtain the prediction results of the active interference pattern.

2. The method according to claim 1, characterized in that, In steps 1a) and 1d), the feature parameters of each sample are normalized using the following formula: ； in, Represents the th sample in the normalized sample set. The first piece of work equipment In the nth sample One feature parameter, Indicates the serial number of the work equipment. Indicates radar, Indicates a jammer; when hour, Indicates the index of the radar characteristic parameter, when hour, The serial number representing the characteristic parameters of the jammer; Indicates the first sample in the sample set The first piece of work equipment The maximum value of each feature parameter Indicates the first sample in the sample set The first piece of work equipment The minimum value of each feature parameter.

3. The method according to claim 2, characterized in that, The feature parameters Its value depends on the work equipment. The serial numbers represent different characteristic parameters of the radar and the jammer, respectively: when hour, The index represents the radar characteristic parameter; different values ​​represent different parameters of the radar characteristic. Indicates the repetition interval of radar pulses in the sample set. This represents the pulse width of the radar pulses in the sample set. This represents the amplitude of radar pulses in the sample set. This indicates the carrier frequency of the radar pulses in the sample set. Indicates the instantaneous bandwidth of radar pulses in the sample set; when hour, The index represents the sequence number of the jammer's characteristic parameters; different values ​​represent different parameters characteristic of the jamming signal. This represents the center frequency of the interference signal in the sample set. Indicates the duration of the interference signal in the sample set. This represents the bandwidth of the interference signal in the sample set.

4. The method according to claim 1, characterized in that, The Transformer encoder layer selected in step 1b) includes multiple cross-connected self-attention mechanisms and a feedforward neural network. The self-attention mechanism is used to perform attention calculations on each position of the input sequence with all other positions. The feedforward neural network is used to perform non-linear transformations on the output of the attention mechanism through a multilayer perceptron mechanism.

5. The method according to claim 1, characterized in that, In step 1c), the network parameters are iteratively updated using the Adam optimizer, as follows: 1c1) Let Let be the gradient of the parameters in the t-th iteration. and Given the decay coefficients of the two exponentially weighted averages, calculate the first moment estimates of the gradients for each. and second moment estimates : ； 1c2) Based on the first moment estimate and second moment estimates Calculate the first moment estimate after gradient bias correction. and second moment estimates : ； in, and This represents the decay coefficient in the t-th iteration; 1c3) Given the parameters before the update, obtain the updated parameters based on the results of step 1c2). : ； in, It's the learning rate. It is a constant whose value tends to 0, used to avoid division by zero.

6. The method according to claim 1, characterized in that, In step 2b), a Q-Learning-based reinforcement learning algorithm is used to add the radar operating modes to the active jamming pattern dataset, as follows: The radar's operating state is defined as the state of the learning environment, and the possible jamming patterns are defined as the actions that the learning subject can choose. Based on the characteristic that the selection of jamming patterns is affected by the radar's operating state, the current Q-table and ε-greedy strategy in the Q-Learning algorithm are used to determine whether to adopt a new jamming pattern or maintain the current jamming pattern, thereby optimizing the selection of jamming patterns. The Q-table is iteratively updated based on the reward matrix in the Q-Learning algorithm, and finally a Q-table reflecting the optimal jamming pattern selection strategy is generated. The radar operating mode data is input into the updated Q-table, and the active jamming pattern data is output, thus realizing the addition of the radar operating mode to the active jamming pattern dataset.

Images