Gcn-based radar emitter operating mode recognition method and device thereof

Abstract

The application discloses a radar radiation source working mode recognition method and device based on GCN, relates to the technical field of radar signal classification, and comprises the following steps: simulating each working mode of a radar radiation source; extracting feature parameters of each simulated working mode and forming a feature matrix; wherein the feature parameters comprise pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency visibility, echo, the number of pulses on each CPI, and bandwidth; performing normalization processing on the feature matrix; training a preset GCN recognition network by using the processed feature matrix, so as to obtain a radar radiation source recognition model; identifying the processed radar radiation source signals to be identified by using the radar radiation source recognition model, so as to obtain the working mode of the radar radiation source signals; and performing dimension reduction and visualization on the working mode of the radar radiation source signals. The application can solve the problem that the connection between feature parameters of different working modes of the existing radar radiation source is weak and sparse.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal classification, and particularly relates to a radar radiation source working mode recognition method based on GCN and a device thereof. BACKGROUND

[0002] With the development of radar technology, especially the use of new systems and new purposes of radar, the modulation form becomes complex, and the modulation parameter conversion is faster and faster. Just as a phased array radar can enter different working modes according to different functional requirements, different modes represent different threat levels, which directly affects the measures taken to respond. Obviously, the phased array radar has multiple working modes, which brings great challenges to electronic search.

[0003] Traditional working mode recognition is often to extract parameters of radar signals, and then to compare templates manually or semi-automatically according to the extracted parameters. Obviously, a large amount of time consumption is extremely unsuitable for the rapidly changing environment. In recent years, the field of artificial intelligence has developed rapidly, and CNN and RNN are very classic algorithms, which can achieve relatively advanced results in many fields. In fact, it is found through research that the classic algorithm is only suitable for processing some data that can be established as a Euclidean space model (such as images, texts, and speeches), and this type of data has translational invariance. However, with the development of society, more and more non-Euclidean space structure problems are in front of us, such as social networks and recommendation systems. Similarly, due to the small difference between the working modes of the phased array radar and the difficulty of extraction, each node in the constructed graph has different connections, some nodes have multiple connections, and some nodes have only one connection, which is an irregular structure. These graph structure data makes the model suitable for processing Euclidean space no longer applicable, so the model for processing graph structure data is challenged.

[0004] Therefore, it is urgent to propose a method suitable for recognizing the working mode of a radar radiation source. SUMMARY

[0005] In order to solve the above problems existing in the prior art, the application provides a radar radiation source working mode recognition method based on GCN and a device thereof. The technical problem to be solved by the application is realized through the following technical scheme:

[0006] In a first aspect, the application provides a radar radiation source working mode recognition method based on GCN, comprising:

[0007] simulating each working mode of the radar radiation source;

[0008] Extract feature parameters of each simulation working mode, and form a feature matrix; wherein, the feature parameters include pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency view number, echo, number of pulses on each CPI and bandwidth;

[0009] The feature matrix is normalized;

[0010] The processed feature matrix is used to train the preset GCN recognition network, and a radar emitter identification model is obtained;

[0011] The radar emitter identification model is used to identify the processed radar emitter signal to be identified, and the working mode of the radar emitter signal is obtained;

[0012] The working mode of the radar emitter signal is reduced and visualized.

[0013] Optionally, the process of normalizing the feature matrix comprises:

[0014] The Row-Normalize normalization algorithm is obtained, and the feature matrix is processed, and the expression is:

[0015] ;

[0016] wherein, is the normalized result, is the th feature of the th node, is the feature matrix, is the number of feature parameters.

[0017] Optionally, it further comprises:

[0018] Obtain the connection relationship of each feature parameter under the same simulation working mode, and construct an adjacency matrix ;

[0019] The adjacency matrix is normalized.

[0020] Optionally, the preset GCN layer recognition network comprises multiple GCN layers;

[0021] wherein, the transmission relationship between adjacent GCN layers is:

[0022] ;

[0023] wherein, is the feature of the node in the th layer, is a nonlinear transformation, is an adjacency matrix with self-loop, is a corresponding degree matrix, a weight of the i-th layer, a weight of the i-th layer.

[0024] Optionally, the preset GCN recognition network sequentially comprises a first GCN layer, a ReLu activation layer, a Dropout layer and a second GCN layer.

[0025] Optionally, the process of training the preset GCN recognition network using the processed feature matrix to obtain the radar emitter recognition model comprises:

[0026] inputting the processed feature matrix into the preset GCN recognition network;

[0027] initializing the weight of the preset GCN recognition network using a Glorot algorithm;

[0028] updating the weight of the preset GCN recognition network using an Adam algorithm;

[0029] obtaining the radar emitter recognition model.

[0030] Optionally, the expression of the Adam algorithm is:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] wherein, is a gradient of a loss function at a current iteration, is an assignment operation, is a gradient operator, is a loss function at a current iteration when iteratively training the preset GCN recognition network, is a weight at a current iteration when iteratively training the preset GCN recognition network, is a first moment estimation of the gradient with an initial value of 0, is an exponential decay rate of the first moment estimation with a value of 0.9, is a second moment estimation of the gradient with an initial value of 0, is an exponential decay rate of the second moment estimation with a value of 0.999, is a transpose operation, is a parameter with a preset value of 0.001, is a smoothing constant with a value of .

[0037] Optionally, the iteration learning rate is 0.01, and the weight decay is 5e-4.

[0038] Optionally, the process of dimensionality reduction visualization of the recognition result of the to-be-identified working mode comprises:

[0039] representing the recognition result as high-dimensional data;

[0040] using a t-SNE algorithm to convert the high-dimensional data into low-dimensional data.

[0041] In a second aspect, the present application further provides a GCN-based radar emitter working mode recognition device, comprising:

[0042] a simulation module configured to simulate each working mode of a radar emitter;

[0043] a feature extraction module configured to extract feature parameters of each simulated working mode and form a feature matrix; wherein the feature parameters comprise pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency visibility, number of pulses on each CPI, and bandwidth;

[0044] a processing module configured to normalize the feature matrix;

[0045] a training module configured to train a preset GCN recognition network using the processed feature matrix to obtain a radar emitter recognition model;

[0046] a recognition module configured to use the radar emitter recognition model to recognize a to-be-identified radar emitter signal after processing to obtain a working mode of the radar emitter signal;

[0047] a dimensionality reduction visualization module configured to perform dimensionality reduction visualization on the working mode of the radar emitter signal.

[0048] The present application has the following advantages:

[0049] The GCN-based radar emitter working mode recognition method and device provided by the present application are based on a GCN mode, process graph structure data, and perform working mode recognition, can mine certain associations between data nodes and data nodes, only use randomly initialized parameters to perform radar working mode recognition, and have obvious effects, greatly reduce the calculation amount of deep network model recognition, and can further solve the problems of weak connection and sparsity of feature parameters between different working modes of the existing radar emitter.

[0050] The application will be described in further detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a flow chart of the GCN-based radar emitter operating mode recognition method provided by the embodiments of the application. DETAILED DESCRIPTION

[0052] The application will be described in further detail below with reference to the drawings and embodiments.

[0053] In the prior art, a radar emitter signal recognition method based on a deep learning network is proposed. Pulse description words and spectral amplitude are taken as primary features and are fused, and the obtained feature vector is input into a sparse deep belief network based on a smooth L0 norm for classification and recognition. Since a sparse penalty term is added on the basis of the deep belief network to establish a sparse deep neural network, the operation efficiency is increased. However, this method only uses a small amount of features of samples for recognition, and does not consider the relevance between features of different operating modes, and the information mining of signals is not sufficient.

[0054] A radar operating mode recognition method based on a random forest algorithm is also proposed. The method first obtains a radar operating mode data set, and pre-processes radar emitter signals, then trains a decision tree basic classifier, at each node, randomly selects a subset of all features to calculate the best segmentation method, then performs prediction classification on the emitter signals, and outputs the results. According to the number of samples and the number of features, the method uses the feature with the minimum Gini index as the tree set point to establish the tree set, but for feature data with different values of radar signals, the weight value produced by the random forest on this data is not reliable. Moreover, the calculation cost of this method will significantly increase with the increase of the number of samples and the number of features.

[0055] In summary, through the analysis of the prior art solutions, it is found that the existing deep learning network is mostly optimized for the network itself, and there are problems of weak connection between feature parameters of different operating modes of radar emitters and sparsity.

[0056] Therefore, the application provides a GCN-based radar emitter operating mode recognition method, which overcomes the problem that the traditional deep learning algorithm model cannot recognize various operating modes through parameters with small distinguishing degrees on a phased array radar and fails to mine the underlying features of the internal parameters of the data.

[0057] Please refer to Figure 1 , Figure 1 is a flowchart of the GCN-based radar emitter operating mode recognition method provided by the application, and the GCN-based radar emitter operating mode recognition method provided by the application comprises the following steps.

[0058] S101, simulating various operating modes of a radar emitter.

[0059] S102, extracting feature parameters of various simulated operating modes and forming a feature matrix; wherein the feature parameters comprise pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency visibility, echo, the number of pulses on each CPI and bandwidth.

[0060] Specifically, the differences of various simulated operating modes in feature parameters are analyzed and extracted, and the feature parameters comprise pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency visibility, echo, the number of pulses on each CPI, bandwidth and the like.

[0061] S103, performing normalization processing on the feature matrix.

[0062] Specifically, the connection relationship between various parameter samples under the same simulated operating mode is identified as "1", and an adjacency matrix A is constructed. In order to avoid increasing the range of output values in the subsequent multiplication operation, the feature matrix is subjected to Row-Normalize normalization processing; in order to avoid changing the original distribution of the features by the inner product multiplication of the adjacency matrix and the feature matrix, the adjacency matrix is subjected to normalization processing.

[0063] S104. Use the processed feature matrix to train the preset GCN recognition network to obtain the radar radiation source recognition model.

[0064] Specifically, the graph convolutional neural network adopts a (convolution-pooling) stacked + fully connected structure. Except for changing the number of neurons in the convolutional layer and the number of stacked layers, other parameters remain unchanged. The feature matrix after processing the simulated radar radiation source working mode is input into the preset GCN recognition network. The weights of the network are initialized using the Glorot algorithm, and then the weights of the network are iteratively updated using the Adam algorithm. The learning rate is set to 0.01 and the weight decay is set to 5e-4 to obtain the radar radiation source recognition model.

[0065] S105. Use the radar radiation source identification model to identify the processed radar radiation source signal to obtain the working mode of the radar radiation source signal.

[0066] S106. Visualize the working mode of the radar radiation source signal in a reduced dimension.

[0067] Specifically, the radar radiation source operating mode identification method based on GCN provided in this embodiment adopts the GCN algorithm, which reduces the processing complexity and makes the real-time performance of radar radiation source operating mode identification higher; and solves the problem of weak correlation and sparsity of feature parameters between different operating modes of existing radar radiation sources.

[0068] In an optional embodiment of the present invention, the process of normalizing the feature matrix includes:

[0069] Obtain the Row-Normalize normalization algorithm and process the feature matrix. Specifically, sum each row of the input sparse matrix and obtain the sum value, which is the normalized value of the non-zero elements in each row. The expression is as follows:

[0070] ;

[0071] in, For the normalized result, For the first The node of the first One characteristic, For the characteristic matrix, This represents the number of feature parameters.

[0072] In an optional embodiment of the present invention, it further includes:

[0073] Obtain the connection relationships of various feature parameters under the same simulation working mode, and construct the adjacency matrix. ;

[0074] For adjacency matrix Normalization processing is performed.

[0075] Specifically, in the same simulation mode of the same kind in the embodiment, the characteristic parameters are connected with each other, and then an adjacency matrix is constructed ; and the sparse matrix form of the adjacency matrix with self-loop is normalized; wherein , is a unit matrix.

[0076] The normalization of the adjacency matrix is realized by , in order to ensure that the adjacency matrix is symmetrical (undirected graph) when calculating, the above normalization is converted to .

[0077] In an optional embodiment of the present application, the preset GCN layer recognition network comprises multiple GCN layers;

[0078] Wherein, the transmission relationship between adjacent GCN layers is:

[0079] ;

[0080] Wherein, is the feature of the node in the th layer, is a nonlinear transformation, is the adjacency matrix with self-loop, is the corresponding degree matrix, is the weight of the th layer, is the intercept of the th layer.

[0081] In an optional embodiment of the present application, the preset GCN recognition network comprises a first GCN layer, a ReLu activation layer, a Dropout layer and a second GCN layer in sequence.

[0082] Specifically, in the embodiment, the output dimensions of the first GCN layer and the second GCN layer are 16 and 7 respectively, the activation function is ReLU function, and the Dropout is set to 0.2.

[0083] In an optional embodiment of the present application, the process of training the preset GCN recognition network using the processed feature matrix to obtain a radar radiation source recognition model comprises:

[0084] Inputting the processed feature matrix into the preset GCN recognition network;

[0085] Initializing the weight of the preset GCN recognition network using Glorot algorithm;

[0086] updating the preset GCN recognition network using the Adam algorithm iteratively;

[0087] obtaining a radar radiation source recognition model.

[0088] In an optional embodiment of the present application, the expression of the Adam algorithm is:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] wherein, is the gradient of the loss function at the current iteration, is an assignment operation, is a gradient operator, is the loss function at the current iteration when iteratively training the preset GCN recognition network, is the weight at the current iteration when iteratively training the preset GCN recognition network, is the first moment estimation of the gradient with an initial value of 0, is the exponential decay rate of the first moment estimation with a value of 0.9, is the second moment estimation of the gradient with an initial value of 0, is the exponential decay rate of the second moment estimation with a value of 0.999, is a transpose operation, is a preset parameter with a value of 0.001, is a smoothing constant with a value of In an optional embodiment of the present application, the iterative learning rate is 0.01, and the weight decay is 5e-4. In an optional embodiment of the present application, the process of dimensionality reduction visualization of the recognition result of the working mode to be recognized includes:

[0095] representing the recognition result as high-dimensional data;

[0096] using a t-SNE algorithm to convert the high-dimensional data into low-dimensional data.

[0097] representing the recognition result as high-dimensional data;

[0098] using a t-SNE algorithm to convert the high-dimensional data into low-dimensional data.

[0099] ​​​​​​​Specifically, in the embodiment, the recognition result is represented as high-dimensional data, in order to visualize the data and intuitively understand the distribution of the data; the t-SNE algorithm is used to maximize the similarity of the mapping of the high-dimensional data to the low-dimensional data; wherein, the t-SNE algorithm includes:

[0100] Suppose , Map to low-dimensional space , ;

[0101] The joint probability distribution of two points in the high-dimensional space:

[0102] ;

[0103] ;

[0104] The joint probability distribution of two points in the low-dimensional space:

[0105] ;

[0106] Loss function:

[0107] ;

[0108] Optimized gradient:

[0109] ;

[0110] Wherein, the t-SNE algorithm converts the similarity between data points into conditional probability, the similarity of data points in the original space is represented by Gaussian joint distribution, and the similarity of data points in the embedded space is represented by t distribution. The KL divergence of the joint probability distribution of the original space and the embedded space. The function related to the KL divergence is used as the loss function, and the loss function is minimized by the gradient descent algorithm.

[0111] Based on the same inventive concept, the application also provides a GCN-based radar emitter operating mode recognition device, which is applied to the recognition method provided in the above-mentioned embodiments of the application, and specific embodiments can be referred to the above, which will not be repeated here; the method comprises:

[0112] A simulation module is configured to simulate each operating mode of the radar emitter;

[0113] A feature extraction module is configured to extract feature parameters of each simulated operating mode and form a feature matrix; wherein, the feature parameters include pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency view number, echo, number of pulses on each CPI, and bandwidth;

[0114] A processing module is configured to normalize the feature matrix;

[0115] a training module, configured to train the preset GCN identification network using the processed feature matrix to obtain a radar emitter identification model;

[0116] an identification module, configured to identify the working mode of the radar emitter signal using the radar emitter identification model.

[0117] a dimension reduction visualization module, configured to perform dimension reduction visualization on the working mode of the radar emitter signal.

[0118] In an optional embodiment of the present application, the effect of the present application is further illustrated below in combination with a simulation experiment:

[0119] 1. Simulation conditions:

[0120] The hardware platform of the simulation experiment of the present application is: the processor is Intel(R) Core i7-10750H, the main frequency is 2.60 GHz, and the memory is 16 GB.

[0121] The software platform of the simulation experiment of the present application is: WINDOWS 10 operating system, MATLAB-R2016a, and Tensor-flow.

[0122] 2. Simulation content and result analysis:

[0123] The radar emitter samples to be identified used in the simulation experiment of the present application are generated by MATLAB simulation, a total of 7000 samples of seven types of working mode signals, and the features of each sample are extracted, including eight parameters: pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency view number, echo, number of pulses on each CPI, and bandwidth.

[0124] The radar emitter identification simulation experiment of the present application is to identify the working mode of each signal sample of the seven different working mode radar emitter signals generated by simulation using the method of the present application, and to train and test the network model under different signal-to-noise ratios to verify the robustness of the model. The total number of correctly identified samples of the seven working mode signals under each signal-to-noise ratio is counted, and then the total number of correctly identified samples of the seven working mode signals under each signal-to-noise ratio is divided by the total number of samples of the seven working mode signals under each signal-to-noise ratio, i.e. 7000, to obtain the correct rate of radar working mode identification under each signal-to-noise ratio. All the calculation results are plotted in Table 1.

[0125] The network model is trained and tested under different signal-to-noise ratios to verify the robustness of the model. Four data sets with different signal-to-noise ratios are selected: 5 dB, 10 dB, 15 dB, and 20 dB. The experimental results are as follows:

[0126] Table 1 recognition accuracy under different signal-to-noise ratios

[0127]

[0128] Table 2 number of iterations for stable accuracy under different signal-to-noise ratios

[0129]

[0130] From Table 1, first, it can be seen that under the above signal-to-noise ratio conditions, as the number of training increases, the recognition accuracy can eventually stabilize near 100%, which shows the effectiveness of the GCN network. Second, from Table 2, it can be found that when the signal-to-noise ratio is 5dB, about 200 iterations are needed for stable recognition accuracy, while when the signal-to-noise ratio is 10dB, more than 100 iterations are needed, and when the signal-to-noise ratio is increased to 15dB, only more than 80 iterations are needed for stable accuracy. It can be seen that as the signal-to-noise ratio increases, the number of iterations required to reach stable recognition rate also decreases, which shows that as the signal condition improves, the network model can learn the rules between different working mode signal features in a shorter training time, and the weight matrix of the GCN model is almost no longer adjusted, and the recognition effect is more and more stable.

[0131] It should be noted that in this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between 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. Elements defined by an expression "comprising a" do not exclude the presence of additional identical elements in the process or method comprising the defined element. The words "connected" or "coupled" or similar terms are not limited to direct physical or mechanical connections or couplings, but can also include electrical connections or couplings, whether direct or indirect. The terms "upper", "lower", "left", "right", and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only used for the convenience of description and simplification of the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0132] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the particular feature or characteristic being described is included in at least one embodiment or example of the present application. The illustrative description of these terms in this specification is not necessarily referring to the same embodiment or example. Moreover, the particular features or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can make several modifications or adaptations to the description of the specification without departing from the concept of the present application, and these should be considered as falling within the scope of the present application.

[0133] The above is a further detailed description of the present application in combination with specific preferred embodiments, and cannot be considered as limiting the specific implementation of the present application to these descriptions. For those skilled in the art, several simple deductions or substitutions can be made without departing from the concept of the present application, and all should be considered as falling within the scope of protection of the present application.

Claims

1. A GCN-based radar emitter operating mode recognition method, characterized in that, The method comprises the following steps: simulate each working mode of a radar radiation source; extract feature parameters of each simulated working mode and form a feature matrix; wherein the feature parameters include pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency view number, return, number of pulses on each CPI, and bandwidth; normalize the feature matrix; train a preset graph convolutional neural network (GCN) recognition network using the processed feature matrix to obtain a radar radiation source recognition model; wherein the preset GCN layer recognition network comprises multiple GCN layers, and the preset GCN recognition network comprises a first GCN layer, a ReLu activation layer, a Dropout layer, and a second GCN layer in sequence; wherein the transmission relationship between adjacent GCN layers is: ； wherein, is the node at the layer, is a non-linear transformation, is the adjacency matrix with self-loops added, is the corresponding degree matrix, is the weight of the layer, is the intercept of the layer; use the radar radiation source recognition model to identify the processed radar radiation source signal to be identified to obtain the working mode of the radar radiation source signal; dimensionality reduction visualization of the working mode of the radar radiation source signal.

2. The GCN-based radar emitter operating mode recognition method according to claim 1, characterized in that, The normalization process of the feature matrix comprises: obtain a Row-Normalize normalization algorithm and process the feature matrix, and the expression is: ； wherein, is the normalized result, is the first node, feature of the first feature matrix, is the number of feature parameters.

3. The GCN-based radar emitter operating mode recognition method according to claim 1, wherein, Further comprising: Obtain the connection relationship of each characteristic parameter in the same simulation working mode, and construct an adjacency matrix ; The adjacency matrix is normalized. is normalized.

4. The GCN-based radar emitter operating mode recognition method according to claim 1, characterized in that, The process of training the preset GCN recognition network using the processed feature matrix to obtain the radar radiation source recognition model comprises: input the processed feature matrix into the preset GCN recognition network; initialize the weight of the preset GCN recognition network using the Glorot algorithm; update the weight of the preset GCN recognition network using the Adam algorithm; obtain the radar radiation source recognition model.

5. The GCN-based radar emitter operating mode recognition method according to claim 4, characterized in that, The expression of the Adam algorithm is: ； ； ； ； ； in, The loss function at the current iteration gradient, For assignment operation, For gradient operators, This refers to the loss function for the current iteration during iterative training of the pre-defined GCN recognition network. The weights for the current iteration during iterative training of the pre-defined GCN recognition network. gradient with initial value of 0 The first moment estimate, The exponential decay rate is estimated by the first moment with a value of 0.

9. gradient with initial value of 0 The second moment estimate, The exponential decay rate is estimated using the second moment with a value of 0.

999. For transpose operation, The parameter has a default value of 0.

001. For the value to be The smoothing constant.

6. The GCN-based radar emitter operating mode recognition method according to claim 4, characterized in that, The iteration learning rate is 0.01, and the weight decay is 5e-4.

7. The GCN-based radar emitter operating mode recognition method according to claim 1, wherein, The process of dimensionality reduction visualization of the working mode of the radar radiation source signal comprises: express the identification result as high-dimensional data; convert the high-dimensional data into low-dimensional data using the t-SNE algorithm. 8.A GCN-based radar emitter operating mode recognition device, characterized in that, The method comprises the following steps: a simulation module for simulating each working mode of a radar radiation source; a feature extraction module for extracting feature parameters of each simulated working mode and forming a feature matrix; wherein the feature parameters include pulse repetition frequency, pulse width, duty cycle, intra-pulse modulation mode, frequency view number, return, number of pulses on each CPI, and bandwidth; a processing module for normalizing the feature matrix; a training module for training a preset graph convolutional neural network (GCN) recognition network using the processed feature matrix to obtain a radar radiation source recognition model; wherein the preset GCN layer recognition network comprises multiple GCN layers, and the preset GCN recognition network comprises a first GCN layer, a ReLu activation layer, a Dropout layer, and a second GCN layer in sequence; wherein the transmission relationship between adjacent GCN layers is: ； wherein, is a feature of the node at the layer, is a non-linear transformation, is an adjacency matrix with self-loops added, is a corresponding degree matrix, is a weight of the layer, is an intercept of the layer; an identification module for using the radar radiation source recognition model to identify the processed radar radiation source signal to be identified to obtain the working mode of the radar radiation source signal; a dimensionality reduction visualization module for dimensionality reduction visualization of the working mode of the radar radiation source signal.

Images