A radar anti-active jamming method based on a generative adversarial network

Abstract

The application relates to a radar active jamming interference suppression method based on a generative adversarial network and relates to the technical field of radar anti-jamming.The application is aimed at solving the problems of high complexity of a traditional radar anti-jamming method, poor interference suppression effect when facing a complex jamming scene, even missing of target echo information, and invalidation of anti-jamming capability.The application inputs a radar echo real signal into a trained generative adversarial network to obtain a target echo signal after removing interference.The application utilizes nonlinear fitting capability to generate a signal with similar data distribution of the target echo signal, removes the interference signal from the radar echo signal in a scene where radar detection exists active jamming interference, and restores the target echo signal.The application does not need to perform multi-domain processing on the received echo signal, can directly realize interference suppression on the echo signal, restores the target echo information, and can cope with a complex and changeable electromagnetic interference environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar anti-jamming. BACKGROUND

[0002] In recent years, with the continuous development of modern electronic technology and intelligent technology, the electromagnetic environment in electronic warfare is becoming more and more complex, and the working performance and survivability of radar systems are facing more and more severe threats and challenges. Therefore, the radar anti-jamming problem has gradually become an extremely important research content in the field of radar technology.

[0003] The existing radar signal processing technology mainly uses multi-domain representation of target echo and jamming for anti-jamming, and often needs to perform time domain processing, frequency domain processing, polarization domain processing and digital filtering on the radar echo signal, so as to realize the suppression of jamming signals and the extraction of detection information. These methods not only have high complexity, but also have poor jamming suppression effect in complex jamming scenarios, and even the target echo information is missing, and the anti-jamming ability is invalid. SUMMARY

[0004] The present application is to solve the problem of high complexity of traditional radar anti-jamming method, poor jamming suppression effect when facing complex jamming scene, even target echo information missing, and anti-jamming ability invalid, and provides a radar anti-active suppression jamming method based on generative adversarial network.

[0005] A radar anti-active suppression jamming method based on generative adversarial network inputs the radar echo real signal into the trained generative adversarial network to obtain the target echo signal after removing the jamming;

[0006] The generative adversarial network comprises a generator and a discriminator,

[0007] The generator is a U-net structure, the generator comprises five encoding blocks, four decoding blocks and one full connection layer, and the discriminator comprises three encoding blocks and one full connection layer,

[0008] The encoding block comprises a convolution layer, a LeakyReLU activation function layer and a maximum pooling layer, and the decoding block comprises an up-sampling layer, a convolution layer and a LeakyReLU activation function layer.

[0009] Further, the training method of the above generative adversarial network is as follows:

[0010] Step one: build the model structure of the generative adversarial network;

[0011] Step two: convert the radar echo simulation signal in the training data set into a matrix S r As the input of the generator, the output data G(S r), the matrix S t and the output data G(S r ) as the input of the discriminator, obtaining the output data D(S t ) and D[G(S r )] of the discriminator,

[0012] The output data G(S r ) represents the result of learning the target echo simulation signal, and the output data D(S t ) and D[G(S r )] respectively represent the probability of the input data being S t and G(S r );

[0013] Step three: substitute D(S t ) and D[G(S r )] obtained in step two into the following loss function:

[0014]

[0015] According to the result of the loss function, the weights of the generator and the discriminator are optimized;

[0016] Step four: judge whether all samples in the training data set are traversed, if yes, the training is completed, otherwise return to step two;

[0017] In the above formula, E represents expectation, P data (·) represents the distribution of data, and V(·) represents the loss function.

[0018] Further, after the training in the above step four is completed, the following steps are further included:

[0019] Step five: the matrix S transforming the radar echo simulation signal in the test data set is taken as the input of the trained generator, and its output data G(S

[0020] Step six: the output data G(S ) and the matrix S transforming the target echo simulation signal in the test data set are taken into the following formula, and the root mean square error value RMSE k of the kth iteration is obtained:

[0021]

[0022] In the above formula, ||·|| represents the matrix norm;

[0023] Step seven: judging whether the curve formed by the root mean square error values obtained in the previous k iterations is convergent, if yes, the test is completed, otherwise, k=k+1 and returning to step two;

[0024] In the above formula, m is the number of samples in the test data set, i=1, 2,..., m, and k=1, 2,....

[0025] Further, before the above step one, a data sample set is first constructed, and the specific method is as follows:

[0026] Adjusting the time delay and the Doppler frequency shift to generate a plurality of different target echo simulation signals, respectively applying an active suppression jamming signal on each target echo simulation signal to obtain a radar echo simulation signal corresponding thereto,

[0027] Converting each target echo simulation signal and radar echo simulation signal into a matrix with the same dimension,

[0028] Taking all the matrices as a data sample set, and dividing the data in the data sample set into a training data set and a test data set.

[0029] Further, the specific method for converting the target echo simulation signal and the radar echo simulation signal into a matrix with the same dimension is as follows:

[0030] Sampling the target echo simulation signal at N points, and constructing a 1-row N-column matrix from the data obtained by the N sampling points, transforming the 1-row N-column matrix into an a-row b-column matrix, and taking the a-row b-column matrix as the matrix S t converted by the target echo simulation signal;

[0031] Sampling the radar echo simulation signal at N points, and constructing a 1-row N-column matrix from the data obtained by the N sampling points, transforming the 1-row N-column matrix into an a-row b-column matrix, and taking the a-row b-column matrix as the matrix S r converted by the radar echo simulation signal;

[0032] Wherein a*b=N, a and b are both integers greater than 1 and less than N, and N is the number of sampling points.

[0033] The beneficial effects of the present application are as follows:

[0034] The application adopts a generative adversarial network (GAN) as a network model, utilizes the powerful nonlinear fitting capability of the GAN, can generate a signal with similar data distribution of a target echo signal, removes an interference signal from the radar echo signal in a radar detection scene with active suppression interference, and restores the target echo signal. The generator adopts a U-net structure, is simple in structure, low in complexity, and high in efficiency. Compared with the prior art, the application does not need to perform multi-domain processing on the received echo signal, can directly implement interference suppression on the echo signal, restores target echo information, and can cope with a complex and changeable electromagnetic interference environment. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a radar detection scene schematic diagram with active interference;

[0036] Figure 2 is a structure schematic diagram of a generator in the GAN network of the application;

[0037] Figure 3 is a structure schematic diagram of a discriminator in the GAN network of the application;

[0038] Figure 4 is a training and testing flowchart of the GAN;

[0039] Figure 5 is a root mean square error (RMSE) convergence curve diagram;

[0040] Figure 6 is a data distribution image generated by the GAN network, wherein (a) represents (b) represents (c) represents DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application. It should be noted that, in the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0042] The radar detection scene is as follows Figure 1As shown, in the radar detection, the radar transmitting signal contacts the target to generate a target echo signal, and in the process of returning of the target echo signal, the active suppression jamming signal generated by the jammer is mixed into the target echo signal, so that the target echo signal has the jamming signal, and finally is received by the radar receiver. In order to eliminate the active suppression jamming in the radar echo signal, the following specific embodiment is given.

[0043] Reference Figures 2 to 6 The embodiment is specifically explained, and a radar anti-active suppression jamming method based on a generative adversarial network is described in the embodiment, and the method is specifically as follows:

[0044] First, a data sample set is constructed, and the specific method is as follows:

[0045] Adjust the time delay and the Doppler frequency shift to generate a plurality of different target echo simulation signals, and apply an active suppression jamming signal to each target echo simulation signal to obtain a corresponding radar echo simulation signal.

[0046] Each target echo simulation signal and radar echo simulation signal is converted into a matrix with the same dimension, and specifically:

[0047] The target echo simulation signal is sampled at 8192 points, and the data obtained by 8192 sampling points is formed into a 1x8192 matrix, the 1x8192 matrix is transformed into a 256x32 matrix, and the 256x32 matrix is taken as the matrix S t .

[0048] The radar echo simulation signal is sampled at 8192 points, and the data obtained by 8192 sampling points is formed into a 1x8192 matrix, the 1x8192 matrix is transformed into a 256x32 matrix, and the 256x32 matrix is taken as the matrix S r .

[0049] All the matrices are collectively taken as a data sample set, and the data in the data sample set is divided into a training data set and a test data set.

[0050] The above converts the two kinds of echo simulation signals into matrices with the same dimension, which can enhance the correlation of the internal data of the signals, make the data distribution characteristics more obvious, be beneficial to the feature extraction of the data in the network, and thus improve the network performance.

[0051] Then, the model is trained and tested by using the data sample set, and the specific method includes the following steps:

[0052] Step 1: Construct the model structure of the generative adversarial network.

[0053] The generative adversarial network includes a generator and a discriminator.

[0054] The generator is a U-net structure, and the generator comprises five encoding blocks, four decoding blocks and one full connection layer; the discriminator comprises three encoding blocks and one full connection layer.

[0055] The encoding block comprises a convolution layer, a LeakyReLU activation function layer and a maximum pooling layer; the decoding block comprises an up-sampling layer, a convolution layer and a LeakyReLU activation function layer.

[0056] The generator can prevent model collapse and overfitting, and is conducive to network convergence. In the above structure, LeakyReLU is used as an activation function, and maximum pooling is to take the point with the maximum value in the local acceptance domain, and the advantage is that it can learn the edge and texture structure of the image. The up-sampling layer is realized by the deconvolution operation. The full connection layer is realized by the convolution operation. The structure uses skip connection, and the feature maps of the encoding block and the decoding block are combined together in a hierarchical connection manner.

[0057] Step two: in order to enable the generator to learn the data distribution of the matrix S t , the matrix S r converted by the radar echo simulation signal in the training data set is taken as the input of the generator, and the output data G(S r ) of the generator is obtained.

[0058] The matrix S t converted by the target echo simulation signal in the training data set and the output data G(S r ) are taken as the input of the discriminator, and the output data D(S t ) and D[G(S r )] of the discriminator are obtained.

[0059] The output data G(S r ) represents the result of learning the target echo simulation signal, and the output data D(S t ) and D[G(S r )] respectively represent the probability of the input data being S t and G(S r ).

[0060] Step three: substitute D(S t ) and D[G(S r )] obtained in step two into the following loss function:

[0061]

[0062] According to the result of the loss function, the weights of the generator and the discriminator are optimized.

[0063] In the above process, the objective of the discriminator is to distinguish true from false, and D[·] is a probability output, so the expectation D(S t ) is close to 1, while the expectation D[G(S r )] is close to 0, so maximizing the loss function of formula is the objective of the discriminator. The objective of the generator is to generate as realistic data as possible to confuse the discriminator, that is, to make the data D[G(S r )] generated by the generator close to 1, so the objective of the generator is to minimize the second term of the loss function . In the training process, the generator generates data G(S t ) close to S r , and the discriminator judges the true or false of the data G(S t ) generated by the generator according to the data distribution of S r . Through the mutual game of the generator and the discriminator, the network weights are continuously optimized, so that the generator generates signals more and more similar to the target echo signal, and the interference suppression performance is continuously improved.

[0064] Step four: judge whether all samples in the training data set are traversed, if yes, complete the training, and then execute step five; otherwise, return to step two.

[0065] In the above formula, E represents expectation, P data (·) represents the distribution of data, and V(·) represents the loss function.

[0066] Step five: convert the radar echo simulation signal in the test data set into a matrix as the input of the trained generator, and obtain the output data

[0067] Step six: bring the output data and the matrix converted from the target echo simulation signal in the test data set into the following formula, and obtain the root mean square error value RMSE of the kth iteration k :

[0068]

[0069] In the above formula, ||·|| represents the matrix norm.

[0070] Step seven: judge whether the curve formed by the root mean square error values obtained in the first k iterations converges, and the curve converges, which means that the generated data and have similar data distribution, and the interference signal is effectively suppressed, so the test is completed, otherwise k=k+1 and return to step two.

[0071] In the above formula, m is the number of samples in the test data set, i = 1, 2,..., m, and k = 1, 2,...

[0072] After obtaining the trained and tested generative adversarial network, it is applied to practice, specifically:

[0073] The radar echo real signal is sampled for 8192 points, and the data obtained by 8192 sampling points is formed into a 1x8192 matrix. The 1x8192 matrix is transformed into a 256x32 matrix. The 256x32 matrix is input into the trained generative adversarial network to obtain a result matrix after removing interference. Then the result matrix is inversely transformed into a 1x8192 matrix to obtain a signal similar to the target echo signal, and the elimination of active suppression jamming is completed.

[0074] In combination with Figure 5 and Figure 6 The performance of the GAN network in the embodiment is further simulated and verified:

[0075] Experimental scenario: The embodiment uses matlab language to build a data set, generates radar LFM target echo simulation signals, and adds coherent suppression jamming signals to the target echo simulation signals to produce radar echo simulation signals. The target echo simulation signals and the radar echo simulation signals are respectively converted into 256x32 matrices, wherein the matrix converted from the target echo simulation signals is S t , and the matrix converted from the radar echo simulation signals is S r Then change the time delay and Doppler shift of the target echo simulation signals to produce 3000 groups of S t and S r to build a data sample set. 80% of the samples in the data sample set are used as training data set, and 20% of the samples are used as test data set.

[0076] The network is built using python language, and the model is trained and tested based on tensorflow2.0. The performance verification is performed using pycharm software. The input batch size of the network is 10. When testing, the matrix S r is input into the trained generator. After training and testing for 12 times, the RMSE curve has converged, and the convergence curve is shown in Figure 5 . Then view the data distribution image of the test result, and the result is shown in Figure 6 . Wherein (a) is the data distribution image of S t , the texture is clear, (b) is the data distribution image of S r , because of the existence of the interference signal, the data distribution image is relatively fuzzy. (c) is the network output data G(S rthe data distribution of the test generated data is consistent with the data distribution of S t *. By comparing the RMSE convergence curve and the data distribution image, it can be seen that the data G(S r * generated by the GAN has a similar data distribution with S t *. After the matrix inverse transformation, the obtained data has a similar data distribution with the target echo simulation signal, and the interference is effectively suppressed.

[0077] In summary, the radar anti-active suppression jamming method based on the generative adversarial network in the embodiment, the network model is composed of a generator and a discriminator, and the generator adopts a U-net structure. In the embodiment, the matrix S r * converted from the echo simulation signal with the interference signal is input into the network, and the matrix S t * converted from the target echo simulation signal is input into the network as a data set for network training. Through the mutual game of the generator and the discriminator, the generator generates a signal more and more similar to the target echo simulation signal. In the test, the matrix S r * converted from the echo simulation signal with the interference signal is input into the trained network, and the root mean square error (RMSE) is added as a performance measurement index. When the RMSE curve converges, the data G(S r * generated by the generator has a similar data distribution with S t *. After the matrix inverse transformation, the obtained data has a similar data distribution with the target echo simulation signal, and the interference is effectively suppressed.

[0078] Although the present application is described herein with reference to particular embodiments, it is to be understood that these examples are merely hypothetical illustrations of the principles and applications of the present application. It should therefore be understood that numerous modifications can be made to the exemplary embodiments and that other arrangements can be devised without departing from the spirit and scope of the present application as defined by the appended claims. It should be understood that the features described in connection with separate embodiments can be used in combination with features described in connection with other embodiments. It should also be understood that features described in connection with individual embodiments can be used in other described embodiments.

Claims

1. A radar anti-active suppression jamming method based on generative adversarial network, characterized in that: Input the real radar echo signal into the trained generative adversarial network to obtain the target echo signal after removing interference; The generative adversarial network includes a generator and a discriminator. The generator is a U-net structure, which includes 5 encoding blocks, 4 decoding blocks and 1 fully connected layer, and the discriminator includes 3 encoding blocks and 1 fully connected layer. The encoding block includes a convolution layer, a LeakyReLU activation function layer and a maximum pooling layer, and the decoding block includes an upsampling layer, a convolution layer and a LeakyReLU activation function layer; Construct a data sample set. The specific method is: Adjust the time delay and Doppler frequency shift to generate multiple different target echo simulation signals, apply active suppression interference signal to each target echo simulation signal, and obtain the corresponding radar echo simulation signal. Convert each target echo simulation signal and radar echo simulation signal into a matrix of the same dimension, All matrices are taken as a data sample set, and the data in the data sample set is divided into a training data set and a test data set; The specific method of converting the target echo simulation signal and the radar echo simulation signal into matrices of the same dimension is: The target echo simulation signal is sampled at N points, and the data obtained from the N sampling points are formed into a matrix with 1 row and N columns. The matrix with 1 row and N columns is transformed into a matrix with a row and b columns, and the matrix with a row and b columns is used as the matrix S for the target echo simulation signal transformation. t ; The radar echo simulation signal is sampled at N points, and the data obtained from the N sampling points are formed into a matrix with 1 row and N columns. The matrix with 1 row and N columns is transformed into a matrix with a row and b columns, and the matrix with a row and b columns is used as the matrix S for transforming the radar echo simulation signal. r ; Wherein a*b=N, a and b are both integers greater than 1 and less than N, and N is the number of sampling points.

2. The radar anti-active suppression jamming method based on generative adversarial network according to claim 1 is characterized in that: The training method of the generative adversarial network is as follows: Step 1: Build the model structure of the generative adversarial network; Step 2: Convert the radar echo simulation signal in the training data set into the matrix S r As the input of the generator, the output data G(S r ), the matrix S converted from the target echo simulation signal in the training data set t and the output data G(S r ) as the input of the discriminator, and obtain the output data D(S t ) and D[G(S r )], The output data G(S r ) represents the result of learning the target echo simulation signal, and the output data D(S t ) and D[G(S r )] respectively indicate that the input data is S t and G(S r ) probability; Step 3: D(S t ) and D[G(S r )] Substitute the following loss function: Optimizing the weights of the generator and the discriminator according to the result of the loss function; Step 4: Determine whether all samples in the training data set have been traversed. If so, complete the training; otherwise, return to step 2. In the above formula, E represents expectation, P data (·) represents the distribution of data, and V(·) represents the loss function.

3. The radar anti-active suppression jamming method based on generative adversarial network according to claim 2 is characterized in that: After the training is completed in step 4, the following steps are further included: Step 5: Convert the radar echo simulation signal in the test data set into a matrix As input to the trained generator, obtain its output data Step 6: Output the data and the matrix converted from the target echo simulation signal in the test data set Substitute the following formula to obtain the root mean square error RMSE of the kth iteration k : In the above formula, ||·|| represents the matrix norm; Step 7: Determine whether the curve formed by the root mean square error values ​​obtained in the first k iterations converges. If so, complete the test; otherwise, set k = k + 1 and return to step 2; In the above formula, m is the number of samples in the test data set, i = 1, 2, ..., m, k = 1, 2, ....

Images