Multifunctional radar working mode identification method based on GAN-SA-GRU

Abstract

The invention discloses a multifunctional radar working mode recognition method based on a GAN-SA-GRU, and the method comprises the steps: building a multifunctional radar pre-data set with a label based on the pulse characteristic parameter setting of an airborne phased array radar, dividing the multifunctional radar pre-data set into a pre-training set, a pre-verification set and a pre-test set, building a multifunctional radar working mode recognition model based on the GAN-SA-GRU, and carrying out the recognition of the working mode of the multifunctional radar based on the GAN-SA-GRU. Inputting the pre-training set and the pre-verification set for training to obtain a trained multifunctional radar working mode recognition model, finally inputting the model into the pre-test set for testing, outputting a test result, and evaluating the accuracy of model recognition. According to the method, by introducing the GAN, self-training with random noise is achieved, non-ideal factors are added into the training set, data set expansion is achieved, the data set obtained through self-training is added into the GRU network for secondary training, and the data set expansion efficiency is improved. And a model which is more stable under a non-ideal condition can be obtained to realize the recognition of the working mode of the multifunctional radar.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar behavior cognition, and specifically relates to a multi-functional radar working mode recognition method based on GAN-SA-GRU. Background Art

[0002] The recognition of the working mode of a multi-functional radar is a key part of the multi-functional radar cognition technology. This technology can help accurately grasp the target situation. Through this technology, it is possible to quantitatively understand its type and operating conditions, providing crucial basis for formulating efficient response strategies and decisions for our side. The recognition and prediction of the working mode of a multi-functional radar are relatively cutting-edge directions.

[0003] In the early research field of the recognition of the working mode of a multi-functional radar, the method of correlating and matching the pulse description words obtained by the detection equipment with the features in the template database was generally adopted. With the in-depth research, more deep learning methods have also been involved. In the literature "Li Y, Zhu M, Ma Y, et al. Work modes recognition and boundary identification of MFR pulse sequences with a hierarchical seq2seq LSTM. IET radar, sonar & navigation, 2020, 14(9): 1343-1353", an end-to-end long short-term memory network was used for supervised learning, and the working mode of the multi-functional radar was recognized using the carrier frequency, pulse width, and pulse repetition interval. However, the way of providing the dataset in the literature was relatively simple and lacked practical consideration; in the literature "He C, Zhang L, Wei S, et al. Multifunction Radar Working Mode Recognition with Unsupervised Hierarchical Modeling and Functional Semantics Embedding Based LSTM. IEEE Sensors Journal, 2024", on the basis of the carrier frequency, pulse width, and pulse repetition interval, the bandwidth was added as the fourth feature parameter to complete the unsupervised hierarchical modeling and functional semantics embedding recognition of the multi-functional radar working mode based on LSTM. However, the setting of the dataset therein did not take practical consideration either and was fixed lossless content.

[0004] In summary, it is of great value to study a method for identifying the working modes of multifunctional radars under non-ideal, non-cooperative environments and incomplete data sets. In particular, a method for identifying the working modes of multifunctional radars in real error scenarios is of great significance. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a method for identifying the working modes of multifunctional radars based on GAN-SA-GRU, which can obtain a more stable model under non-ideal conditions to realize the identification of the working modes of multifunctional radars.

[0006] The technical solution adopted by the present invention is as follows: A method for identifying the working modes of multifunctional radars based on GAN-SA-GRU, the specific steps are as follows:

[0007] S1. Based on the pulse feature parameters of the airborne phased array radar, construct a pre-dataset of multifunctional radars with labels, and divide it into a pre-training set, a pre-verification set, and a pre-test set;

[0008] S2. Combine the generative adversarial network, gated recurrent unit, and self-attention mechanism to construct a model for identifying the working modes of multifunctional radars based on GAN-SA-GRU;

[0009] Among them, the model includes: a GAN network model, an SA-GRU network model; the GAN network model includes a group of generators G(·) and discriminators D(·), and is trained in an adversarial manner.

[0010] S3. Input the pre-training set obtained in step S1 into the GAN network model in step S2 for training, obtain a generated training set according to random noise, and combine the generated training set and the pre-training set in step S1 to form a training data set;

[0011] S4. Import the training data set constructed in step S3 and the pre-verification set in step S1 into the SA-GRU network model for training until the maximum number of training rounds is reached, and obtain a trained model for identifying the working modes of multifunctional radars;

[0012] S5. Add different deviation levels to the data in the pre-test set in step S1 to construct a noisy test set, and then input the pre-test set and the noisy test set into the trained model for identifying the working modes of multifunctional radars in step S4 for testing, output the test results and evaluate the accuracy of model identification.

[0013] Further, the specific steps of step S1 are as follows:

[0014] First, according to different working modes and corresponding characteristic parameters, use Matlab software to generate different pulse sequences of multifunctional radar working modes Generate the labeled sample Data of the multifunctional radar working mode pulse sequence by one-hot vector encoding to label the corresponding true label y sequence =[X N , y].

[0015] Among them, feature represents the feature parameters, including: RF, PRF, PW; represents the true value of the parameter feature of the i p th pulse, N represents the total number of pulses, [·] T represents the transpose operation. y represents the corresponding labeled label of the multifunctional radar working mode pulse sequence.

[0016] Perform several Monte Carlo simulations on the above process according to the actual situation to obtain several labeled samples Data of the multifunctional radar working mode pulse sequence generated under ideal conditions sequence , and finally form the pre-dataset Data pre , and divide it into the pre-training set Data pre-train , the pre-validation set Data pre-val and the pre-test set Data pre-test according to a certain proportion, and the division proportion is determined according to the actual situation.

[0017] Among them, the data in the pre-dataset Data pre needs to be normalized and then processed. Z-score normalization is used, and the parameter value for normalization is

[0018] Among them, α m represents the mth parameter value of the working mode pulse sequence, α m ' represents the normalization result, represents the average value and standard deviation of this parameter in this sequence, and M represents the length of this sequence.

[0019] Furthermore, the specific steps of step S3 are as follows:

[0020] Input the pre-training set Data obtained in step S1 pre-train into the GAN network model in step S2 for training, and obtain the generated training set according to the random noise z.

[0021] In the GAN network model, the generator G(·) generates samples G(Z) that approximate the true distribution by adding random noise Z. If the generated samples G(Z) can deceive the discriminator D(·), then the loss function of the generator G(·) is L G = g 1 L G1 + g 2 LG2 。

[0022] Among them, g 1 and g 2 respectively represent the weight factors of L G1 and L G2 where g 1 + g 2 = 1; represents the generator loss function of the GAN, represents taking the expected value, z ~ p Z (z) indicates that the random noise variable z is sampled from the distribution p Z (z) of the random noise Z, log(·) represents taking the logarithm, G(z) represents the generated data output by the generator, and D(G(z)) represents the probability of the discriminator judging the authenticity of G(z); represents minimizing the similarity loss between the generated data, ||·|| 2 represents taking the L2 norm, z 1 , z 2 ~ p Z (z) indicates that the random noise variables z 1 and z 2 are sampled from the distribution p Z (z) of the random noise Z, and G(z 1 ) and G(z 2 ) represent the generated data generated by the generator according to the random noise variables z 1 and z 2 through adversarial generation.

[0023] The discriminator D(·) continuously conducts adversarial training on real samples and generated samples until the two reach equilibrium. The loss function of the discriminator is L D = d 1 L D1 + d 2 L D2 .

[0024] Among them, d 1 and d 2 respectively represent the weight factors of L D1 and L D2 where d 1 + d 2 = 1; represents the discriminator loss in the GAN, maximizing the discrimination probability of real samples and minimizing the discrimination probability of generated samples. x ~ p X (x) indicates that the real sample data x is sampled from the distribution p X (x) of the real sample X, and D(x) represents the probability of the discriminator judging the authenticity of x; Denotes the contrast loss, which forces the similarity between real samples to be higher than that between real and generated samples, x 1 , x 2 ~p X (x) represents the real sample x 1 and x 2 Sampled from the distribution p X (x) of the real sample X, D(x 1 ), D(x 2 ) respectively represent the probability of the discriminator judging the authenticity of x 1 , x 2 for.

[0025] Combining the loss functions of the generator G(·) and the discriminator D(·), the loss function of the GAN network model is obtained

[0026] Among them, Represents the double optimization problem on the generator G(·) and the discriminator D(·), that is, alternately optimizing the generator G(·) and the discriminator D(·) to minimize L G and L D .

[0027] When the loss function L GAN converges, it is considered that the GAN network model converges at this time, and the generated training set Data gen-train is output, and the pre-training set Data pre-train and the generated training set Data gen-train are combined to form the training data set Data train .

[0028] Furthermore, the specific steps of step S4 are as follows:

[0029] Input the training data set Data train and the pre-verification set Data pre-val into the SA-GRU network model for training. The SA-GRU inserts a self-attention layer after each time step of the GRU, calculates the context vector through self-attention, and inputs the concatenated result into the fully connected layer for classification. The hidden state output by the GRU. Use the Adam optimizer to update the network parameters until the preset maximum number of training rounds is reached, and obtain a trained multi-functional radar working mode recognition model based on GAN-SA-GRU.

[0030] Among them, the SA-GRU network model uses the cross-entropy loss function with L2 regularization as the loss function, that is, L SA-GRU = L CE + L R .

[0031] Among them, N m represents the total number of samples, m represents the working mode label of the nth sample recognized by the GAN-SA-GRU network, represents the nth m true working mode label of the sample; ω 2 represents the L2 regularization weight vector, and λ represents the L2 regularization factor.

[0032] Furthermore, the specific steps of step S5 are as follows:

[0033] Add data with different deviation levels DL to the pre-test set Data pre-test in step S1 to construct a noisy test set, and obtain Data n-test .

[0034] Among them, the deviation level feature represents the specified feature parameter, and P feature and respectively represent the value of the mth dimension parameter of the pulse sequence in the ideal electromagnetic environment and the value under the corresponding error deviation.

[0035] Input the pre-test set Data pre-test and the noisy test set Data n-test into the multi-functional radar working mode recognition model trained in step S4 for testing, obtain the test working mode label, and calculate the recognition accuracy

[0036] Among them, N test represents the number of samples in the test set, and respectively represent the true working mode label and the test working mode label of the nth test sample. When Otherwise,

[0037] Advantages of the present invention: The method of the present invention first constructs a multi-functional radar pre-dataset with labels based on the pulse characteristic parameters of an airborne phased array radar, and divides it into a pre-training set, a pre-validation set, and a pre-test set. Then, a multi-functional radar working mode recognition model based on GAN-SA-GRU is constructed. The pre-training set and the pre-validation set are input for training to obtain a trained multi-functional radar working mode recognition model. Finally, the pre-test set is input for testing, and the test results are output and the recognition accuracy of the model is evaluated. By introducing GAN, the method of the present invention realizes self-training with random noise, adds non-ideal factors to the training set, realizes the expansion of the dataset, and adds the self-trained dataset to the GRU network for secondary training, so as to obtain a more stable model under non-ideal conditions to realize the recognition of the working mode of the multi-functional radar. This is beneficial to improving the reliability of the recognition results of the working mode of the multi-functional radar under non-ideal, non-cooperative environments and incomplete dataset conditions, and provides support for cognitive technology or interference technology. Description of the Drawings

[0038] Figure 1 It is a flowchart of a method for recognizing the working mode of a multi-functional radar based on GAN-SA-GRU of the present invention.

[0039] Figure 2 It is a schematic diagram of the GAN network model structure in an embodiment of the present invention.

[0040] Figure 3 It is a schematic diagram of the technical approach of the SA-GRU network model in an embodiment of the present invention.

[0041] Figure 4 It is a comparison chart of the working mode recognition accuracy under different loss levels in an embodiment of the present invention. Detailed Embodiments

[0042] The method of the present invention will be further described below in conjunction with the drawings and embodiments.

[0043] As Figure 1 shown, a flowchart of a method for recognizing the working mode of a multi-functional radar based on GAN-SA-GRU (Generative Adversarial Network-Self-Attention-Gated Recurrent Unit) of the present invention is as follows:

[0044] S1. Based on the pulse characteristic parameters of an airborne phased array radar, construct a multi-functional radar pre-dataset with labels, and divide it into a pre-training set, a pre-validation set, and a pre-test set;

[0045] S2. Combine the generative adversarial network, gated recurrent unit, and self-attention mechanism to construct a multi-functional radar working mode recognition model based on GAN-SA-GRU;

[0046] Among them, the model includes: a GAN network model and an SA-GRU network model; the GAN network model includes a set of generators G(·) and discriminators D(·), and is trained in an adversarial manner.

[0047] S3. Input the pre-training set obtained in step S1 into the GAN network model in step S2 for training, obtain a generated training set according to random noise, and combine the generated training set and the pre-training set in step S1 to form a training data set;

[0048] S4. Import the training data set constructed in step S3 and the pre-validation set in step S1 into the SA-GRU network model for training until the maximum number of training rounds is reached to obtain a trained multi-functional radar working mode recognition model;

[0049] S5. Add different deviation levels to the data in the pre-test set in step S1 to construct a noisy test set, and then input the pre-test set and the noisy test set into the trained multi-functional radar working mode recognition model in step S4 for testing, output the test results, and evaluate the recognition accuracy of the model.

[0050] In this embodiment, step S1 is specifically as follows:

[0051] A radar pulse is the basic unit of a radar system, which represents the basic time unit of a radar signal and can be represented by multiple parameter feature dimensions. A radar working mode is composed of a series of radar pulses. There is no clear one-to-one relationship between the working mode and the pulses, but the working mode can be regarded as the result of combining multiple radar pulses. First, according to different working modes and corresponding characteristic parameters, use Matlab software to generate different multi-functional radar working mode pulse sequences Generate a multi-functional radar working mode pulse sequence annotation sample Data sequence =[X N , y] by one-hot vector encoding to label the corresponding true label y.

[0052] Among them, feature represents characteristic parameters, including: RF, PRF, PW; represents the true value of the parameter feature of the i p -th pulse, N represents the total number of pulses, and [·] T represents the transpose operation. y represents the corresponding multi-functional radar working mode pulse sequence annotation label.

[0053] Perform the Monte Carlo simulation on the above process several times according to the actual situation to obtain a number of labeled samples Data of the multifunctional radar working mode pulse sequences generated under ideal conditions sequence , and finally form the pre-dataset Data pre , and divide it into the pre-training set Data pre-train , the pre-verification set Data pre-val and the pre-test set Data pre-test according to a certain ratio, and the division ratio is determined according to the actual situation.

[0054] Among them, the data in the pre-dataset Data pre needs to be normalized before processing. Z-score normalization is used, and the parameter values for normalization are

[0055] Among them, α m represents the m-th parameter value of the working mode pulse sequence, and α m ' represents the normalization result, represents the average value and standard deviation of this parameter in this sequence, and M represents the length of this sequence.

[0056] In this embodiment, the pulse parameter range of the airborne multifunctional radar working mode is shown in Table 1. 1000 groups of working mode pulse data are generated in Matlab, that is, Data pre is 400 groups, and their pulse lengths are all 200. Divide them into the pre-training set Data pre-train , the pre-verification set Data pre-val and the pre-test set Data pre-test according to the ratio of 3:1:1, and obtain Data pre-train 600 groups, Data pre-val 200 groups, Data pre-test 200 groups.

[0057] Table 1

[0058]

[0059] In this embodiment, the specific steps of step S2 are as follows:

[0060] A multi-functional radar working mode recognition method based on GAN-SA-GRU is constructed by combining the GAN network model and the SA-GRU network model. The GAN network model is used to expand the dataset by combining random noise with a small amount of labeled data, and the expanded dataset is imported into the SA-GRU network for training to obtain better training results. The advantage of the GAN network lies in extracting implicit features from the input data and capturing the inherent patterns of the target. The GAN network includes a group of generators G(·) and discriminators D(·), and the two are trained in an adversarial manner. As a variant of the RNN network, GRU not only has excellent capabilities for processing sequence data, but is also simpler and more efficient, solving problems such as gradient explosion and gradient disappearance in existing RNNs. SA-GRU inserts a self-attention layer after the existing GRU, calculates the context vector through self-attention, and the two are concatenated and then input into the fully connected layer for classification.

[0061] The schematic diagram of the GAN model in this embodiment is as Figure 2 shown. In the generator G(·) of the GAN network, the encoder uses three layers of temporal convolutional network (TCN), the channel is set to [128, 128], the kernel size is 4 / 4 / 2, and Tconv1 and Tconv 2 are combined with the LeakyReLU activation function and the batch normalization layer; the decoder uses three layers of transposed convolution, the kernel sizes are 2 / 4 / 4 respectively, ConvTran1d 1 and ConvTran1d 2 are combined with the LeakyReLU activation function and the batch normalization layer, and ConvTran1d 3 uses the Tanh activation function.

[0062] The discriminator D(·) of the GAN network adopts a structure of three layers of TCN plus two layers of fully connected. The convolution and decoder settings are the same. The second layer of the fully connected layer is combined with the Sigmoid activation function, and the output dimension is 1.

[0063] The schematic diagram of the technical approach of the SA-GRU model in this embodiment is as Figure 3 shown. In the SA-GRU network, a double-layer bidirectional GRU network is used, which contains 128 hidden layer units. The double-layer GRU is used because radar signals usually contain multi-level temporal features and can improve the feature abstraction ability by stacking layers, forming a complement with the self-attention mechanism; the bidirectional GRU is used because the radar working mode may imply a lag effect, and the mode switching requires cumulative observations over multiple pulse periods, and can enhance the model's sensitivity to the mode switching boundary of the working mode and improve the classification robustness. The hidden state sequence output by the GRU is H = [h 1 , h 2 ,..., h T , and the dimension is T×d h .

[0064] Among them, T represents the sequence length, and d h represents the dimension of the hidden state, that is, the number of GRU hidden units, and h t (t = 1, 2,..., T) represents the hidden state value.

[0065] The attention mechanism includes four steps: linear mapping, correlation calculation, weight assignment, and weighted summation.

[0066] Through linear mapping, the hidden state sequence is H, and the query (Query) matrix Q = HW Q , the key (Key) matrix K = HW K , the value (Value) matrix V = HW V , W Q , W K , W V are their respective projection matrices, and W Q and W K have dimensions of d h ×d k , d k represents the dimension of the key vector, and W V has dimensions of d h ×d v , d v represents the value vector dimension; the correlation matrix is where T represents the sequence length; the similarity matrix is normalized by Softmax to generate the attention weight matrix The dimension of A is T×T, and Softmax(·) represents the Softmax function operation to normalize each row, so that represents the attention probability of the i A th time step to the j A th time step; the value matrix V is weighted and summed using the attention weight A, and the output context vector matrix C = AV is obtained. C aggregates the information of all time steps in the sequence and has dimensions of T×d v ,

[0067] In this embodiment, the specific steps of step S3 are as follows:

[0068] The pre-training set Data pre-train obtained in step S1 is input into the GAN network model in step S2 for training, and a generated training set is obtained according to the random noise z.

[0069] In the GAN network model, the generator G(·) generates samples G(Z) that approximate the real distribution by adding random noise Z. The generated samples G(Z) can deceive the discriminator D(·), and the loss function of the generator G(·) is LG = g 1 L G1 + g 2 L G2 。

[0070] Among them, g 1 and g 2 respectively represent the weight factors of L G1 and L G2 The weight factor, g 1 + g 2 = 1; represents the generator loss function of GAN, represents taking the expected value, z ~ p Z (z) indicates that the random noise variable z is sampled from the distribution p Z (z) of the random noise Z, log(·) represents taking the logarithm, G(z) represents the generated data output by the generator, and D(G(z)) represents the probability of the discriminator judging the authenticity of G(z); represents minimizing the similarity loss between generated data, ||·|| 2 represents taking the L2 norm, z 1 , z 2 ~ p Z (z) indicates that the random noise variable z 1 and z 2 are sampled from the distribution p Z (z) of the random noise Z, G(z 1 ) and G(z 2 ) represent the generated data generated by the generator according to the random noise variables z 1 and z 2 through adversarial generation.

[0071] The discriminator D(·) continuously conducts adversarial training on real samples and generated samples until the two reach equilibrium. The loss function of the discriminator is L D = d 1 L D1 + d 2 L D2 。

[0072] Among them, d 1 and d 2 respectively represent the weight factors of L D1 and L D2 The weight factor, d 1 + d 2 = 1; represents the discriminator loss in GAN, maximizing the discrimination probability of real samples and minimizing the discrimination probability of generated samples, x ~ p X (x) indicates that the real sample data x is from the distribution p XSampled from \(p_{data}(x)\), \(D(x)\) represents the probability of the discriminator judging the authenticity of \(x\). Denotes the contrastive loss, which enforces higher similarity among real samples than between real and generated samples, \(x\) 1 , \(x\) 2 ~\(p\) X \(p_{data}(x)\) represents the real sample \(x\) 1 and \(x\) 2 Sampled from the distribution \(p_{data}(x)\) of real samples \(X\), \(D(x\) X )、\(D(x\) 1 ) respectively represent the probability of the discriminator judging the authenticity of \(x\) 2 、\(x\) 1 、\(x\) 2 .

[0073] Combining the loss functions of the generator \(G(\cdot)\) and the discriminator \(D(\cdot)\), the loss function of the GAN network model is obtained

[0074] Among them, Denotes the double optimization problem on the generator \(G(\cdot)\) and the discriminator \(D(\cdot)\), that is, alternately optimizing the generator \(G(\cdot)\) and the discriminator \(D(\cdot)\) to minimize \(L\) G and \(L\) D .

[0075] When the loss function \(L\) GAN converges, it is considered that the GAN network model converges at this time, and the generated training set Data gen-train is output, and the pre-training set Data pre-train and the generated training set Data gen-train are combined to form the training data set Data train .

[0076] In this embodiment, for the random noise \(Z\) used for adversarial generation, the input random noise feature represents the specified feature parameters, including: RF, PRF, PW, Denotes that the specified feature parameters satisfy the error with a mean of 0 and a variance of , and \(I\) represents the identity matrix. Considering the comprehensive influence, in this embodiment, the noise intensity in is set to [0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5], a total of 9 cases. The pre-training set Data pre-train After being augmented with different random noises, the generated data set Data gen-train can be obtained, which contains 5400 groups of working mode pulse data, and the pulse length is 200 for all. Combining Data pre-trainA training dataset Data containing 6000 groups of working mode pulse data can be obtained train 。

[0077] In this embodiment, the step S4 is specifically as follows:

[0078] Input the training dataset Data train and the pre-verification set Data pre-val into the SA-GRU network model for training. The self-attention layer is inserted after each time step of the GRU. The context vector is obtained through self-attention calculation. After splicing the two, it is input into the fully connected layer for classification. The hidden state output by the GRU. Use the Adam optimizer to update the network parameters until the preset maximum number of training epochs is reached, and a trained multi-functional radar working mode recognition model based on GAN-SA-GRU is obtained.

[0079] Among them, the SA-GRU network model uses the cross-entropy loss function with L2 regularization as the loss function, that is, L SA-GRU =L CE +L R 。

[0080] Among them, N m represents the total number of samples, represents the working mode label of the nth m sample recognized by the GAN-SA-GRU network, represents the nth m sample's true working mode label; ω 2 represents the L2 regularization weight vector, and λ represents the L2 regularization factor.

[0081] In this embodiment, the GRU network part uses a bidirectional double-layer GRU network, which contains 128 hidden layer units. The network training is set as follows: the maximum number of network training epochs is set to 100, Dropout is set to 0.3, Batchsize is set to 64, and the learning rate is a dynamic learning rate.

[0082] In this embodiment, the step S5 is specifically as follows:

[0083] Add data with different deviation levels DL to the pre-test set Data pre-test in step S1 to construct a noisy test set, and obtain Data n-test 。

[0084] Among them, the deviation level feature represents the specified feature parameter, P feature and respectively represent the value of the m - dimensional parameter of the pulse sequence in the ideal electromagnetic environment and the value under the corresponding error deviation.

[0085] In this embodiment, DL ranges from 10% to 50%, that is, DL = [0.1, 0.2, 0.3, 0.4, 0.5]. At each deviation level, 200 are obtained, and a total of Data n-test A total of 1000.

[0086] Input the pre - test set Data pre-test and the noisy test set Data n-test into the multi - functional radar working mode recognition model trained in step S4 for testing, obtain the test working mode label (recognition result), and calculate the recognition accuracy

[0087] Among them, N test represents the number of samples in the test set, and respectively represent the true working mode label and the test working mode label of the n test th sample. When Otherwise,

[0088] The specific recognition effect of this embodiment is as shown in the GAN - SA - GRU curve in Figure 4 . It can be seen the improvement of GAN - SA - GRU compared with the GRU, CNN - GRU, and SA - GRU models. After calculation, under different random noise deviation levels, the GAN - SA - GRU model of the method of the present invention has a greater improvement compared with CNN - GRU, SA - GRU, and GRU. From the perspective of accuracy, under the optimal conditions, GAN - SA - GRU has shown the highest benchmark value, which is 1.53%, 5.29%, and 9.94% higher than CNN - GRU, SA - GRU, and GRU respectively; under the most severe conditions, the advantage reaches a break - level gap, with improvements of 41.38%, 115.79%, and 272.73% respectively. In addition, the attenuation amplitude of the accuracy of GAN - SA - GRU is significantly lower than that of CNN - GRU, SA - GRU, and GRU; the full - scene stability is significantly higher than that of CNN - GRU, SA - GRU, and GRU.

[0089] In summary, the method of the present invention realizes self-training with random noise by introducing GAN, adds non-ideal factors to the training set to expand the data set, and adds the data set obtained by self-training into the SA-GRU network for secondary training, so as to obtain a more stable model under non-ideal conditions to realize the recognition of the working modes of multifunctional radars. This is conducive to improving the reliability of the recognition results of the working modes of multifunctional radars under non-ideal, non-cooperative environments and incomplete data sets, and provides support for cognitive technology or interference technology.

[0090] Those of ordinary skill in the art will realize that the embodiments described herein are for helping readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention according to these technical revelations disclosed by the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A multifunctional radar working mode recognition method based on GAN-SA-GRU, the specific steps are as follows: S1. Based on the pulse characteristic parameter setting of the airborne phased array radar, a multi-function radar pre-data set with labels is constructed and divided into a pre-training set, a pre-verification set, and a pre-test set; S2. Combine the generative adversarial network, gated recurrent unit and self-attention mechanism to build a multi-function radar working mode recognition model based on GAN-SA-GRU; in, The model includes: a GAN network model and a SA-GRU network model; the GAN network model includes a set of generators G(·) and discriminators D(·), and is trained in an adversarial manner; S3, inputting the pre-training set obtained in step S1 into the GAN network model in step S2 for training, obtaining a generated training set according to random noise, and combining the generated training set with the pre-training set in step S1 to form a training data set; S4, importing the training data set constructed in step S3 and the pre-verification set in step S1 into the SA-GRU network model for training until the maximum training round is reached, thereby obtaining a trained multi-function radar working mode recognition model; S5. Add different deviation levels to the data in the pre-test set in step S1 to construct a noisy test set, and then input the pre-test set and the noisy test set into the multi-function radar working mode recognition model trained in step S4 for testing, output the test results and evaluate the accuracy of model recognition.

2. A multifunctional radar working mode recognition method based on GAN-SA-GRU according to claim 1, characterized in that: The step S1 is specifically as follows: Firstly, according to different working modes and corresponding characteristic parameters, different multi-function radar working mode pulse sequences are generated using Matlab software. Generate the multi-function radar working mode pulse sequence labeled sample Data by annotating the corresponding real label y through one-hot vector encoding sequence =[X N ,y]; Among them, feature represents feature parameters, including: RF, PRF, PW; Indicates the i p The true value of the parameter feature of the pulse, N represents the total number of pulses, [·] T Represents the transpose operation; y Indicates the corresponding multi-function radar working mode pulse sequence label; According to the actual situation, the above process is simulated several times by Monte Carlo simulation to obtain the multi-function radar working mode pulse sequence labeling sample Data generated under several ideal conditions. sequence , and finally form the pre-dataset Data pre , pre-training set Data according to the proportion pre-train , Pre-validation set Data pre-val And the pre-test set Data pre-test The division ratio shall be determined according to the actual situation; Among them, for the pre-dataset Data pre The data needs to be normalized before processing, and Z-score normalization is used. The normalized parameter value is Among them, α m Indicates the mth parameter value of the working mode pulse sequence, α m ' represents the normalized result, represents the mean and standard deviation of the parameter in the sequence, and M represents the length of the sequence.

3. The multifunctional radar working mode recognition method based on GAN-SA-GRU according to claim 1 is characterized in that: The step S3 is specifically as follows: The pre-training set Data obtained in step S1 pre-train Input into the GAN network model in step S2 for training, and generate a training set based on the random noise z; In the GAN network model, the generator G(·) generates samples G(Z) close to the real distribution by adding random noise Z. The generated samples G(Z) can deceive the discriminator D(·). The loss function of the generator G(·) is L G =g1L G1 +g2L G2 ; Among them, g1 and g2 represent L G1 and L G2 Weight factor, g1+g2=1; represents the generator loss function of GAN, Indicates taking the expected value, z~p Z (z) represents the random noise variable z from the distribution p of random noise Z Z (z), log(·) represents the logarithm, G(z) represents the generated data output by the generator, and D(G(z)) represents the probability of the discriminator judging the authenticity of G(z); represents minimizing the similarity loss between generated data, ||·||2 represents taking the L2 norm, z1,z2~p Z (z) represents the random noise variables z1 and z2 from the distribution p of random noise Z Z (z), G(z1) and G(z2) represent the generated data generated by the generator according to the random noise variables z1 and z2; The discriminator D(·) continuously conducts adversarial training on real samples and generated samples until the two reach a balance. The loss function of the discriminator is L D =d1L D1 +d2L D2 ; Where d1 and d2 represent L D1 and L D2 Weight factor, d1+d2=1; represents the discriminator loss in GAN, maximizing the discrimination probability of real samples and minimizing the discrimination probability of generated samples, x~p X (x) represents the real sample data x from the real sample X distribution p X (x), D(x) represents the probability of the discriminator making a true or false judgment on x; represents the contrast loss, forcing the similarity between real samples to be higher than that between real and generated samples, x1,x2~p X (x) represents the distribution p of the real samples x1 and x2 from the real sample X X (x), D(x1) and D(x2) represent the true and false judgment probabilities of the discriminator for x1 and x2 respectively; Combining the loss functions of the generator G(·) and the discriminator D(·), we get the loss function of the GAN network model in, It represents a dual optimization problem on the generator G(·) and the discriminator D(·), that is, the generator G(·) and the discriminator D(·) are optimized alternately to minimize L G and L D ; When the loss function L GAN Convergence, then it is considered that the GAN network model converges at this time, and the output generates the training set Data gen-train , the pre-training set Data pre-train And generate training set Data gen-train Combined to form the training data set Data train .

4. The multifunctional radar working mode recognition method based on GAN-SA-GRU according to claim 1 is characterized in that: The step S4 is specifically as follows: The training data set Data train And the pre-validation set Data pre-val The data is input into the SA-GRU network model for training. SA-GRU inserts a self-attention layer after each time step of GRU. The context vector is obtained through self-attention calculation. The two are concatenated and input into the fully connected layer for classification. The hidden state of GRU output; The Adam optimizer is used to update the network parameters until the preset maximum number of training rounds is reached, and a trained multi-function radar working mode recognition model based on GAN-SA-GRU is obtained; The SA-GRU network model adopts the cross entropy loss function with L2 regularization as the loss function, that is, L SA-GRU =L CE +L R ; in, N m represents the total number of samples, Indicates the nth m The working mode labels of the samples, Indicates the nth m The true working mode labels of samples; ω2 represents the L2 regularization weight vector, and λ represents the L2 regularization factor.

5. The multifunctional radar working mode recognition method based on GAN-SA-GRU according to claim 1 is characterized in that: The step S5 is specifically as follows: The pre-test set Data in step S1 pre-test Add different deviation levels of DL to the data in to construct a noisy test set and get Data n-test ; Among them, the deviation level feature indicates the specified feature parameter, P feature and They represent the value of the m-th dimension parameter of the pulse sequence under an ideal electromagnetic environment and under the corresponding error deviation respectively; The pre-test set Data pre-test And noisy test set Data n-test Input the trained multi-function radar working mode recognition model in step S4 for testing, obtain the test working mode label, and calculate the recognition accuracy Among them, N test represents the number of test set samples, and Respectively represent the nth test The real working mode label and the test working mode label of the sample are otherwise,

Images