A method for identifying deceptive jamming based on semi-supervised generative adversarial network

Abstract

The application discloses a kind of based on semi-supervised generation confrontation network's cheating interference identification method, comprising: obtaining the K group slow time complex envelope sequence of the target to be identified;K group slow time complex envelope sequence is sequentially horizontally linked to obtain two-dimensional data block;Interference identification network is constructed, including generator network and discriminator network;White noise data is input into generator network and is trained, obtains and the pseudo-label data similar to the echo signal to be identified;Pseudo-label data, the data block that radar obtains and is not marked echo signal data and the data block that is jointly constructed with marked echo signal data are input into discriminator network and are trained;Two-dimensional data block of the signal to be identified is input into the discriminator network after training, obtains classification result.The application uses a small amount of labeled data, a large amount of unlabeled data and a large amount of pseudo-labeled data generated by generator network to train network model, greatly improves the cheating interference identification probability of interference identification network under small sample condition.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of interference discrimination, and particularly relates to a spoofing interference discrimination method based on a semi-supervised generative adversarial network. BACKGROUND

[0002] Modern electromagnetic environment is more and more complex, and electronic technology represented by electronic jamming brings a serious threat to the detection performance of radar systems. Compared with suppressive jamming, the jammer generates false target information by delaying, modulating and forwarding, and acts on the radar signal receiving system, resulting in that the radar receives multiple spoofing target signals, and affecting the detection and tracking effect. In order to complete target detection, the radar system needs to adopt corresponding electronic countermeasure technology, and strong jamming countermeasure ability is the basic guarantee for the normal operation of radar in modern electronic warfare. In the existing research, for spoofing jamming, the single-station radar system has a single perspective in the target detection process, and it is difficult to obtain rich environmental information. Therefore, in the face of high-fidelity spoofing false targets, the anti-jamming effect of single-station radar is not ideal, and it is difficult to compete with existing complex electronic jamming.

[0003] In view of the limitations of single-station radar, the working mode of multi-radar cooperation has been recognized by more scholars. By connecting multiple radars with different spatial distribution, a networked detection system is constructed, multiple observation angles, multiple frequency bands and multiple working modes are occupied, and a high-density signal space is formed. Compared with single-station radar system, multi-platform and multi-sensor can effectively improve the target detection performance and parameter estimation accuracy. The information collected by the multi-station radar system is cooperatively processed in the fusion center, thereby enhancing the recognition and suppression of jamming, and greatly improving the system systematized anti-jamming ability. Therefore, the key problem in signal processing technology is how to efficiently process the large amount of target echo data set captured by the multi-station radar system, and better adapt to spoofing jamming countermeasures.

[0004] The countermeasures of traditional multi-station radar system against active deceptive jamming can be divided into two parts: data level and signal level. For the data level jamming countermeasure method, the isomorphic multi-station radar system can complete the target discrimination work more stably through the artificial extraction of features such as spatial position distribution aggregation, radial velocity, SNR (Signal to Noise Ratio) chi-square test, etc. Facing the heterogeneous multi-station radar system, through the fusion structure of centralized and distributed, the multi-station radar system can counteract the cooperative jamming. The signal level jamming countermeasure method is based on the difference of target space scattering characteristics, and the judgment standard of true and false targets is that the real target echoes are independent of each other, and the deceptive false target echoes are completely correlated. However, the traditional algorithm needs special discrimination features for different fusion levels and different system combination modes of radar. In the real environment, it is difficult to flexibly combine the available radar stations, and in the face of complex jamming environment, there is still room for improvement in the effectiveness and universality of the jamming countermeasure system. At the same time, although the artificial feature extraction process is intuitive and clear, it is time-consuming and cumbersome, not only requiring the designer to have certain prior knowledge, but also its strong pertinence will make the database update and upgrade slowly, which is difficult to meet the needs of modern electronic warfare for automation and informatization.

[0005] Compared with the traditional artificial feature extraction jamming countermeasure, the deep neural network model can extract multi-dimensional features from the target echo, and the multi-feature works together to obtain the deeper differences between the real target and the false target, promote the fusion information between the radar echoes from the receivers of different systems, and improve the effectiveness and universality of the jamming countermeasure, providing more technical methods for counteracting complex electronic jamming environment.

[0006] In the existing research, Liu proposed a multi-radar system jamming discrimination method based on convolutional neural network (CNN), which combined the advantages of multi-radar system cooperative detection technology and convolutional neural network, effectively applied in the field of anti-deceptive jamming, fully utilized the unknown information of echo data, and alleviated the influence of radar distribution on jamming discrimination under non-ideal conditions, and widened the boundary conditions of the application process. Luo proposed a semi-supervised deceptive jamming discrimination method based on convolutional deep belief network (CDBN), which used enough unlabeled data set to train the CRBM layer to obtain effective data features, and then used a small amount of labeled data set to train the classifier to complete the task of real and false echo classification, more fully utilized the labeled echo data, and realized the construction of more accurate deceptive jamming discrimination network under small sample condition.

[0007] The above methods all use various deep neural networks to solve technical bottlenecks in traditional signal processing processes, such as insufficient feature extraction, single discrimination method, information loss in conversion process and the like, so as to realize effective discrimination of false targets. The existing deep neural network method improves the defects of the traditional method to some extent, but still highly depends on a large amount of data and cannot cope with the real battlefield environment. The existing intelligent jamming countermeasure method needs to spend a long time to process data and learn features, and the discrimination probability is not high under the condition of a small sample, especially when the SNR or sampling rate is low, the existing method has poor ability to learn features and low running efficiency. SUMMARY

[0008] In order to solve the above problems in the prior art, the present application provides a deception jamming discrimination method based on a semi-supervised generative adversarial network.

[0009] The present application provides a deception jamming discrimination method based on a semi-supervised generative adversarial network, comprising:

[0010] S1: obtaining K sets of slow-time complex envelope sequences of a target to be identified according to K radars in a multi-station radar system, wherein K≥2;

[0011] S2: sequentially horizontally linking the K sets of slow-time complex envelope sequences to obtain a two-dimensional data block of the echo signal to be identified;

[0012] S3: constructing a semi-supervised generative adversarial jamming discrimination network, wherein the network comprises a generator network and a discriminator network;

[0013] S4: inputting white noise data into the generator network for training, and obtaining a large amount of pseudo-label data similar to the echo signal to be identified;

[0014] S5: inputting the pseudo-label data generated by the generator network, the labeled echo signal data captured by the radar and the unlabeled echo signal data jointly constructed data block into the discriminator network for training, and obtaining the trained discriminator network;

[0015] S6: inputting the two-dimensional data block of the echo signal of the target to be identified into the discriminator network to obtain a final classification output result.

[0016] In an embodiment of the present application, the S1 comprises:

[0017] S1.1: after each radar performs matched filtering, coherent accumulation and constant false alarm detection on the received signal, a plurality of preset targets are obtained, wherein the preset targets are real targets or jammers;

[0018] S1.2: all preset targets in the same distance unit in all preset targets are screened out, all pulse repetition times of the coherent processing period are obtained, and after the distance unit is matched filtered in all the pulse repetition times, a plurality of complex amplitudes are obtained, and all complex amplitudes of each radar constitute a slow-time complex envelope sequence of the radar.

[0019] In an embodiment of the present application, the generator network comprises a first upsampling layer, a data resetting module, a second upsampling layer, a first convolutional layer, a first function activation layer, a third upsampling layer, a second convolutional layer, a second function activation layer, a fourth upsampling layer, a third convolutional layer and a third function activation layer connected in sequence, wherein the first upsampling layer is used to upsample the input noise signal to obtain original data of the generative adversarial network; the data resetting module is used to reset the output data of the upsampling layer to convert it into three-dimensional data; the first function activation layer and the second function activation layer are LeakyRelu function activations, and the third function activation layer is a Tanh function activation.

[0020] In an embodiment of the present application, the discriminator network comprises a feedforward network module, and a C_Model classification module and a D_Model discrimination module connected to the feedforward network module respectively, wherein

[0021] The feedforward network module comprises a first deconvolutional layer, a fourth function activation layer, a first pooling layer, a second deconvolutional layer, a fifth function activation layer, a second pooling layer, a third deconvolutional layer, a sixth function activation layer, a third pooling layer, a first full connection layer and a second full connection layer connected in sequence;

[0022] The C_Model classification module is a Softmax activation function, the D_Model discrimination module is a Sigmoid activation function, and the output data of the feedforward network module obtains a prediction result through two types of activation functions of the D_Model classification module and the C_Model discrimination module respectively.

[0023] In an embodiment of the present application, the S5 comprises:

[0024] S5.1: a training data set is constructed, the training data set comprises real echo data received by the multi-station radar system, and the real echo data is divided into labeled data and unlabeled data;

[0025] S5.2: a data block composed of the labeled data and the unlabeled data and pseudo-label data generated by the generator network are jointly input into the discriminator network for training;

[0026] S5.3: the training data set and the pseudo-label data are used for iterative training, and the parameter structure of the interference discrimination network is optimized according to a loss function.

[0027] S5.4: According to the loss function, the whole semi-supervised generative adversarial interference discrimination network is trained by back propagation to reduce the loss function.

[0028] In an embodiment of the present application, the loss function is:

[0029]

[0030] wherein G(z) represents a generator network function, D(x) represents a discriminator network function, represents the loss of the discriminator network, E x～Pz(z) [log(1-D(G(z)))] represents the loss of the sample generated by the generator network after passing through the discriminator network. Compared with the prior art, the present application has the beneficial effects that:

[0031] 1. The interference discrimination method of the present application improves the utilization rate of the sampling data in the information processing process of the multi-station radar system and improves the discrimination probability of the deceptive jamming. Compared with the prior art of manually extracting single features and only analyzing the correlation of the data to discriminate the interference, the present application applies the artificial intelligence generative adversarial network to the multi-station radar system, learns the unknown data in depth, excavates more comprehensive, more diverse and more essential data features other than the correlation, improves the utilization rate of the sampling data, jointly processes multiple features, trains an efficient discrimination network, makes up for the problems of single feature extraction by hand and unsatisfactory discrimination effect, and improves the discrimination probability of the multi-station radar system to the deceptive jamming.

[0032] 2. The deceptive jamming discrimination method based on the semi-supervised generative adversarial network of the present application constructs an end-to-end generative adversarial network, introduces a semi-supervised neural network training model from the perspective of learning the feature representation of the real distribution of the echo data, uses a small amount of labeled data, a large amount of unlabeled data and a large amount of pseudo-labeled data generated by the generator network as training samples, trains the network model by borrowing the idea of game theory, and greatly improves the deceptive jamming discrimination probability of the interference discrimination network under the condition of small samples.

[0033] 3. For the influence of noise, the present application effectively reduces the influence of noise and pulse number on the interference discrimination probability and improves the overall anti-deceptive jamming capability of the multi-station radar system. The discrimination effect of the prior art is not ideal in the noise environment, while the present application can effectively reduce the influence of noise on the interference discrimination probability due to the reference analysis of more information by the generative adversarial network, and improve the overall anti-deceptive jamming capability of the multi-station radar system.

[0034] The present application will be further described in detail below in combination with the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a flow chart of a deception jamming discrimination method based on a semi-supervised generative adversarial network provided by an embodiment of the present application;

[0036] Figure 2 is a multi-radar system target space scattering schematic diagram provided by an embodiment of the present application;

[0037] Figure 3 is an echo data linking schematic diagram provided by an embodiment of the present application;

[0038] Figure 4 is a structural schematic diagram of a semi-supervised generative adversarial jamming discrimination network provided by an embodiment of the present application;

[0039] Figure 5 is a relationship between the number of iterations and jamming discrimination probability when having different labeling rates;

[0040] Figure 6 is jamming discrimination effect of the deception jamming discrimination method based on a semi-supervised generative adversarial network provided by an embodiment of the present application under different PRIs;

[0041] Figure 7a is a comparison of jamming discrimination effects between a prior art CNN method and an SGAN method of an embodiment of the present application;

[0042] Figure 7b is a comparison of jamming discrimination effects between a prior art CDBN method and an SGAN method of an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined purposes, a deception jamming discrimination method based on a semi-supervised generative adversarial network according to the present application is described in detail below in combination with the accompanying drawings and specific embodiments.

[0044] The foregoing and other technical contents, features and effects of the present application can be clearly presented in the detailed description of the specific embodiments below in combination with the accompanying drawings. Through the description of the specific embodiments, the technical means and effects taken by the present application to achieve the predetermined purposes can be understood more deeply and specifically. However, the accompanying drawings are provided for reference and illustration only, and are not used to limit the technical solutions of the present application.

[0045] It should be noted that, in this document, the terms "first", "second", and the like, merely denote different instances of an entity or action, and do not necessarily require or imply any actual relationship or order between or among such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process or method. Without further limitation, an element defined by an occurrence of the words "comprising a" does not exclude the presence of additional identical elements in the process or method comprising the element.

[0046] Referring to Figure 1 , Figure 1 is a flowchart of a method for identifying spoofing jamming based on a semi-supervised generative adversarial network according to an embodiment of the present application. The method for identifying spoofing jamming comprises the following steps.

[0047] S1: obtaining K sets of slow-time complex envelope sequences from K radars in a multi-station radar system, where K≥2.

[0048] Specifically, referring to Figure 2 , Figure 2 is a schematic diagram of target spatial scattering of a multi-radar system according to an embodiment of the present application. It is assumed that there are K node radars in the multi-station radar system of the embodiment, which constitute a networked detection system. Each node radar receives a set of slow-time complex envelope sequences after receiving a signal.

[0049] In the embodiment, step S1 comprises:

[0050] S1.1: obtaining a plurality of preset targets by performing matched filtering, coherent accumulation and constant false alarm detection on the received signal by each radar.

[0051] Specifically, it is assumed that each radar receives a signal. After receiving the signal, the signal is sequentially subjected to matched filtering, coherent accumulation and constant false alarm detection processing. After the above processing, a plurality of preset targets can be obtained, wherein the preset target is a target or jamming.

[0052] S1.2: screening preset targets in all preset targets in the same distance unit. In the distance unit, all pulse repetition times of the coherent processing period are obtained. After matched filtering is performed on the distance unit in all pulse repetition times, a plurality of complex amplitudes are obtained. Then, all complex amplitudes of each radar constitute a slow-time complex envelope sequence of the radar.

[0053] Specifically, in the distance unit, the preset targets in the distance unit are determined among all preset targets, and then all pulse repetition times in the coherent processing period of the distance unit are obtained for the distance unit, so that the corresponding all complex amplitudes of the distance unit can be obtained by performing matched filtering on the distance unit with all the obtained pulse repetition times, and all complex amplitudes of each radar constitute a slow-time complex envelope sequence of the radar, and thus K groups of slow-time complex envelope sequences are obtained by K radars. The distance unit is a resolution unit where the target is located.

[0054] S2: sequentially horizontally link the K groups of slow-time complex envelope sequences to obtain a two-dimensional data block of the to-be-identified echo signal.

[0055] Specifically, K groups of slow-time complex envelope sequences can be obtained through step S1, and the slow-time complex envelope sequences A k are sequentially horizontally linked, and the slow-time complex envelope sequence of each radar is used as a row of information, so as to construct a two-dimensional data block, which is denoted as:

[0056] Data 2D =[A 1 ;A 2 ;...;A K ]

[0057] wherein, Q represents that there are a total of Q PRIs (Pulse Repetition Interval), that is, one complex amplitude is obtained each time sampling, and a total of Q times of sampling are performed. 1 denotes a group of data obtained by radar 1 (the first radar among the K radars), wherein, denotes a complex amplitude obtained by radar 1 each time sampling.

[0058] S3: constructing a semi-supervised generative adversarial interference discrimination network, the network comprising a generator network and a discriminator network.

[0059] Please refer to Figure 4 , Figure 4is a structural diagram of a semi-supervised generative adversarial interference discrimination network provided by an embodiment of the present application. The interference discrimination network comprises a generator network and a discriminator network, the generator network is used to construct a pseudo-label data set, and the pseudo-label data set contains a large amount of pseudo-label data generated by the generator network. Specifically, the generator network of the present embodiment comprises a first up-sampling layer, a data resetting module, a second up-sampling layer, a first convolutional layer, a first function activation layer, a third up-sampling layer, a second convolutional layer, a second function activation layer, a fourth up-sampling layer, a third convolutional layer and a third function activation layer connected in sequence, wherein the first up-sampling layer is used to up-sample an input noise signal such as Gaussian white noise into more abundant data to obtain original data of the generative adversarial network; the data resetting module is used to reset the output data of the up-sampling layer into three-dimensional data; the parameters of the first convolutional layer, the second convolutional layer and the third convolutional layer are shown in Table 1. The first function activation layer and the second function activation layer are LeakyRelu function activations, and the third function activation layer is a Tanh function activation.

[0060] Table 1. Parameters of the semi-supervised generative adversarial interference discrimination network of the present embodiment

[0061]

[0062] Further, the discriminator network of the present embodiment comprises a feedforward network module, a C_Model classification module with a Softmax activation function and a D_Model discrimination module with a Sigmoid activation function, and the C_Model classification module and the D_Model discrimination module are connected to the feedforward network module respectively, wherein the D_Model discrimination module and the C_Model classification module share most of the weights of the feedforward network module. The feedforward network module comprises a first inverse convolutional layer, a fourth function activation layer, a first pooling layer, a second inverse convolutional layer, a fifth function activation layer, a second pooling layer, a third inverse convolutional layer, a sixth function activation layer, a third pooling layer, a first full connection layer and a second full connection layer connected in sequence, wherein the step lengths of the first pooling layer, the second pooling layer and the third pooling layer are all 2. The fourth function activation layer, the fifth function activation layer and the sixth function activation layer are all LeakyRelu function activations. The specific parameters of each layer of the discriminator network can be referred to Table 1.

[0063] The C_Model classification module is a Softmax activation function, and the D_Model discrimination module is a Sigmoid activation function, and the output data of the feedforward network module obtains a prediction result through the two types of activation functions of the D_Model classification module and the C_Model discrimination module respectively.

[0064] Specifically, in the D_Model discrimination module, it is judged whether the data output by the feedforward network module is real echo data of the received radar or pseudo-label data generated by the generator network. A Sigmoid activation function is used in the output layer of the D_Model discrimination module, and a binary cross-entropy loss function is used to optimize the parameters of the D_Model discrimination module. By comparing a large amount of pseudo-label data generated by the generator network with real echo data, the purpose of pre-training the shared network of the discriminator network is achieved.

[0065] In the C_Model classification module, it is judged whether the input data is a target or interference, that is, the model category of the input data is judged. A softmax activation function is used in the output layer of the C_Model classification module, and a classification cross-entropy loss function is used to optimize the parameters of the discriminator network. By using a small amount of real target echo and interference echo data, the discriminator network is further trained, and the network weight of the shared part of the discriminator network is optimized.

[0066] S4: input white noise data into the generator network for training, and obtain a large amount of pseudo-label data similar to the to-be-identified echo signal.

[0067] It should be noted that the white noise data is randomly generated according to a certain distribution rule. In this embodiment, the white noise data is Gaussian white noise. Specifically, the Gaussian white noise is transmitted into the generator network for training, and the generator network is used to generate pseudo-label data close to real radar echo data as much as possible from the input Gaussian white noise.

[0068] S5: input the data block composed of the pseudo-label data generated by the generator network, the labeled echo signal data captured by the radar, and the unlabeled echo signal data into the discriminator network for training, and obtain the trained discriminator network.

[0069] In this embodiment, step S5 includes:

[0070] S5.1: construct a training data set, wherein the training data set includes real echo data received by the multi-station radar system, and the real echo data is divided into labeled data (i.e., labeled as target data or interference data) and unlabeled data.

[0071] S5.2: input the data block composed of the labeled data and the unlabeled data and the pseudo-label data generated by the generator network into the discriminator network for training.

[0072] Specifically, as described above, the discriminator network of the embodiment includes a C_Model classification module with a Softmax activation function, and a D_Model discrimination module with a Sigmoid activation function, wherein the D_Model discrimination module and the C_Model classification module share most of the weights of the feedforward network module. In the structure of the feedforward network module shared by the discriminator network, the feature extraction layer is three deconvolution layers, and the output of each deconvolution layer contains a pooling operation with a step of 2. After the data passes through the three deconvolution and pooling operations, the output results are transmitted to two layers of fully connected neural networks. Subsequently, the output data of the fully connected neural network are obtained through two types of activation functions of the D_Model discrimination module and the C_Model classification module to obtain the prediction results.

[0073] S5.3: Optimizing the parameter structure of the interference discrimination network according to the loss function.

[0074] In the embodiment, the loss function of the interference discrimination network based on semi-supervised generative adversarial is as follows:

[0075]

[0076] The loss function of the interference discrimination network is consistent with the generative adversarial network. Wherein, G(z) represents the generator network function, D(x) represents the discriminator network function, represents the loss of the discriminator network, E x～Pz(z) [log(1-D(G(z)))] represents the loss of the sample generated by the generator network after passing through the discriminator network.

[0077] S5.4: According to the loss function, the entire interference discrimination network based on semi-supervised generative adversarial is trained by back propagation to reduce the loss function.

[0078] Specifically, according to the Nash equilibrium principle, when the probability distribution p G of the radar echo data data , the model reaches the equilibrium point, and at this time the interference discrimination network training is completed.

[0079] S6: Inputting the two-dimensional data block into the discriminator network to obtain the final classification output result.

[0080] Specifically, the two-dimensional data block to be discriminated is obtained by steps S1 and S2, and the two-dimensional data block is input into the trained discriminator network to obtain the final classification output result.

[0081] The semi-supervised generative adversarial network-based deception jamming discrimination method of the present application uses a small amount of labeled data, a large amount of unlabeled data and pseudo-labeled data generated by a generator network to train a discriminator network. The pseudo-labeled data reduces the dependence of the jamming discrimination network on the amount of data, thereby improving the robustness and universality of the discrimination performance under small sample conditions. The deception jamming discrimination method of the embodiment of the present application can achieve the same performance as the existing convolutional neural network (CNN)-based jamming recognition method under 10% label rate (the label rate is the proportion of labeled data in the total training data) of the training process, which is better than the performance of the convolutional deep belief network (CDBN) under the same label proportion. At the same time, the method of the embodiment of the present application is significantly better than the existing convolutional neural network jamming discrimination method under 50% labeled data. The simulation results show that the semi-supervised generative adversarial network-based deception jamming discrimination method of the embodiment of the present application reduces the data requirement, enhances the practicability of the network, and is more suitable for real battlefield environment.

[0082] The comparative experimental data of the deception jamming discrimination method of the embodiment of the present application and the existing method under different label rates and signal-to-noise ratios are as follows:

[0083] (1) When the PRI (pulse repetition interval) is set to 12 and the TNR (target noise ratio, the signal-to-noise ratio SNR for the real target and the jam-to-noise ratio JNR for the jamming) is set to 6dB, the simulation results are shown in Table 2. Under the condition of 10% data label rate, the semi-supervised generative adversarial network is obviously better than the discrimination method using only artificial extraction of single feature (traditional method), indicating that the semi-supervised generative adversarial network (SGAN) can extract multi-dimensional essential features of signal data and is suitable for discrimination of real and false targets. Compared with the fully connected neural network (DNN) and the convolutional neural network (CNN) trained using 20,000 labeled data, the jamming discrimination accuracy of the semi-supervised generative adversarial network of the embodiment remains basically unchanged, but the required number of labels is reduced by 90%. At the same time, compared with the semi-supervised convolutional deep belief network (CDBN) under the same other conditions, the accuracy is improved from 91.3% to 99.1%.

[0084] Table 2. Discrimination performance of different algorithms

[0085]

[0086] The parameter settings remain the same as before, Figure 5The relationship between the discrimination probability and the number of iterations is demonstrated. The semi-supervised generative adversarial network method of the embodiment of the application improves the discrimination probability of interference under the condition of a small sample. With the increase of the number of iterations, the final accuracy is greater than 98%. The experiment proves the effectiveness of the semi-supervised generative adversarial network interference confrontation of the embodiment of the application. The increase of the label rate improves the discrimination probability of the deceptive interference, and finally obtains more stable output results. However, it should be noted that due to the small number of available labels, after multiple training, the 5% label rate is unstable when the number of iterations is small. The 10% label rate can ensure a stable discrimination probability of more than 99%, and therefore it is recommended to use more than 10% labels for training of the semi-supervised generative adversarial network interference discrimination network, so as to reduce the data dependency of the intelligent algorithm and ensure the discrimination performance.

[0087] (2) Other parameters are consistent with (1), and the influence of the number of pulses on the false target (i.e. interference) recognition performance is verified by changing the SNR and PRI. The range of TNR (the signal-to-noise ratio of the real target and the dry ratio of the deceptive false target) is set to be between -3dB and 18dB, and the target discrimination probability is obtained under different pulse numbers by the semi-supervised generative adversarial network, wherein the number of PRI is from 4 to 20 with a step of 4.

[0088] As shown in Figure 6 , with the increase of the number of pulses PRI, the discrimination probability is significantly improved. Because the more samples used for interference discrimination in each group, the more information the deep neural network can refer to, and the discrimination performance is effectively improved. When the amount of information meets the discrimination requirement, the discrimination probability tends to be stable and no longer changes with PRI.

[0089] At the same time, with the increase of TNR, the discrimination probability is significantly increased. Because the increase of TNR reduces the influence of noise on the echo signal, the network can more easily obtain the basic characteristics or active deceptive interference of the real target, thereby improving the discrimination performance. The above experiment shows that the semi-supervised generative adversarial network interference discrimination method of the embodiment of the application can ensure the interference discrimination performance under different PRI conditions.

[0090] (3) Compared with the convolutional neural network interference discrimination method under the same condition. The simulation result of FIG. 7(a) shows that when the labeled data is 10%, the semi-supervised generative adversarial network method (SGAN) of the embodiment of the application is slightly worse than the convolutional neural network (CNN) at low TNR. When TNR is greater than 0dB, the discrimination probability of the SGAN and CNN methods converges and finally is greater than 99%. In addition, when the proportion of the labeled data increases to 50%, the performance of the semi-supervised generative adversarial network discrimination network is completely superior to that of the fully supervised convolutional neural network discrimination network. The above comparison experiment is sufficient to prove that the semi-supervised generative adversarial network of the embodiment of the application is superior to the existing convolutional neural network interference discrimination algorithm.

[0091] As shown in Figure 7(b), compared with the convolutional deep belief network interference discrimination method (CDBN), the SGAN method of the embodiment of the application can ensure that the discrimination performance is significantly improved under the same label rate. The experiment shows that the semi-supervised generative adversarial network is a robust interference discrimination method superior to the convolutional deep belief network semi-supervised interference discrimination network in different environments.

[0092] The fraud interference discrimination method based on the semi-supervised generative adversarial network of the embodiment of the application constructs an end-to-end generative adversarial network, introduces a semi-supervised neural network training model from the perspective of learning the feature representation of the real distribution of echo data, uses a small amount of labeled data, a large amount of unlabeled data and a large amount of pseudo-labeled data generated by the generator network as training samples, trains the network model by borrowing the idea of game theory, and greatly improves the fraud interference discrimination probability of the interference discrimination network under the condition of small samples.

[0093] The above is a further detailed description of the application in combination with specific preferred embodiments, and the specific implementation of the application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the application belongs, some simple deductions or substitutions can be made without departing from the concept of the application, and all of them should be regarded as falling within the protection scope of the application.

Claims

1. A deceptive interference identification method based on semi-supervised generative adversarial network, characterized in that: include: S1: Obtain K groups of slow-time complex envelope sequences of the target to be identified based on K radars in the multi-station radar system, where K ≥ 2; S2: horizontally linking the K groups of slow-time complex envelope sequences in sequence to obtain a two-dimensional data block of the echo signal to be identified; S3: construct a semi-supervised generative adversarial interference discriminator network, the network comprising a generator network and a discriminator network; S4: inputting white noise data into the generator network for training, and obtaining a large amount of pseudo-label data similar to the echo signal to be identified; S5: inputting the data block constructed by the pseudo-label data generated by the generator network, the labeled echo signal data captured by the radar, and the unlabeled echo signal data into the discriminator network for training to obtain a trained discriminator network; S6: Inputting the two-dimensional data block of the echo signal to be identified into the discriminator network to obtain a final classification output result.

2. The deceptive interference identification method based on semi-supervised generative adversarial network according to claim 1 is characterized in that Said S1 comprises: S1.1: Each radar performs matched filtering, coherent integration, and constant false alarm detection on the received signal to obtain a number of preset targets, where the preset targets are real targets or interference; S1.2: Filter and obtain preset targets in the same range unit from all preset targets. In the range unit, obtain all pulse repetition times of the coherent processing cycle. Perform matched filtering on the range unit within all the pulse repetition times to obtain a number of complex amplitudes. All the complex amplitudes of each radar constitute the slow-time complex envelope sequence of the radar.

3. The deceptive interference identification method based on semi-supervised generative adversarial network according to claim 1 is characterized in that The generator network includes a first upsampling layer, a data reset module, a second upsampling layer, a first convolution layer, a first function activation layer, a third upsampling layer, a second convolution layer, a second function activation layer, a fourth upsampling layer, a third convolution layer and a third function activation layer connected in sequence, wherein the first upsampling layer is used to upsample the input noise signal to obtain the original data of the generative adversarial network; the data reset module is used to reset the output data of the upsampling layer and convert it into three-dimensional data; the first function activation layer and the second function activation layer are LeakyRelu function activations, and the third function activation layer is Tanh function activation.

4. The deceptive interference identification method based on a semi-supervised generative adversarial network according to claim 1 is characterized in that The discriminator network includes a feedforward network module and a C_Model classification module and a D_Model identification module respectively connected to the feedforward network module, wherein The feedforward network module includes a first deconvolution layer, a fourth function activation layer, a first pooling layer, a second deconvolution layer, a fifth function activation layer, a second pooling layer, a third deconvolution layer, a sixth function activation layer, a third pooling layer, a first fully connected layer and a second fully connected layer, which are connected in sequence; The C_Model classification module is a Softmax activation function, the D_Model identification module is a Sigmoid activation function, and the output data of the feedforward network module obtains prediction results through the two types of activation functions of the D_Mode identification module and the C_Model classification module respectively.

5. The deceptive interference identification method based on semi-supervised generative adversarial network according to claim 1 is characterized in that The S5 includes: S5.1: Construct a training data set, where the training data set includes real echo data received by the multi-station radar system, where the real echo data is divided into labeled data and unlabeled data; S5.2: The blocks of labeled and unlabeled data, along with the pseudo-labeled data generated by the generator network, are fed into the discriminator network for training. S5.3: Performing iterative training using the training data set and the pseudo-labeled data while optimizing the parameter structure of the interference identification network according to a loss function; S5.4: According to the loss function, the entire semi-supervised generative adversarial interference discrimination network is trained through backpropagation to reduce the loss function.

6. The deceptive interference identification method based on semi-supervised generative adversarial network according to claim 5 is characterized in that: The loss function is: Among them, G(z) represents the generator network function, D(x) represents the discriminator network function, represents the loss of the discriminator network, E x～Pz(z) [log(1-D(G(z)))] represents the loss of the samples generated by the generator network after passing through the discriminator network.

Images