Radar working mode recognition method based on combination of GNN and attention mechanism

Abstract

The invention relates to a radar working mode recognition method based on combination of GNN and an attention mechanism, and belongs to the technical field of radar signal classification, and the method comprises the following steps: inputting a radar working mode pulse data set, and mapping the radar working mode pulse data set into a plurality of radar pulse sequence data with features; the method comprises the following steps: constructing a GNN network containing an attention mechanism, initializing a target node, and inputting radar pulse sequence data with target characteristics into the target node; the GNN uses an attention mechanism to extract a context vector of the peripheral information of the target node, and supplements and updates the target node according to the context vector of the peripheral information; and inputting a final target node output by the GNN into a classification model formed by two full connection layers and an activation function, and predicting radar working mode classification of the target node. According to the method, the accuracy and robustness of radar working mode recognition are remarkably improved, and the limitation of a single GNN model in radar signal processing is overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal classification, and specifically relates to a radar working mode recognition method based on the combination of GNN and attention mechanism. Background Art

[0002] Radar operating mode recognition, that is, passive detection systems such as RWR (Radar Warning Receiver), ELINTR (Electronic Intelligence Receiver) and ESM (Electronic Support Measurement) identify the current operating mode of the radar by intercepting and analyzing radar signals.

[0003] The radar working mode is identified based on the specific waveform field semantics of the radar signal sequence. The difficulty of radar working mode identification is mainly reflected in two aspects: (1) The radar signal sequence contains a high information density, and its signals are closely connected before and after, which makes it easy to misjudge the semantics of a single waveform without considering the context; (2) The amount of information is large, the information is messy and not focused enough, and it is impossible to quickly and accurately identify relevant information. Summary of the invention

[0004] The purpose of the present invention is to provide a radar working mode recognition method based on the combination of GNN and attention mechanism, so as to solve the problem that the existing radar working mode recognition is prone to misjudgment of the semantics of a single waveform and cannot quickly and accurately identify related information.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows: The present invention relates to a radar working mode recognition method based on the combination of GNN and attention mechanism, which comprises the following steps: S1, inputting a radar working mode pulse data set and mapping it into a plurality of characteristic radar pulse sequence data; S2. Construct a GNN network with an attention mechanism, initialize a target node, the node is used to represent the feature classification of pulse data recognition in the radar working mode, and input the radar pulse sequence data with target characteristics into the target node; S3, GNN network uses the attention mechanism to extract the context vector of the target node's surrounding information, and supplements and updates the target node based on the context vector of the surrounding information; S4. Input the final target node output by the GNN network into a classification model consisting of two fully connected layers and an activation function to predict the radar working mode classification of the target node.

[0006] Preferably, the S1 slices the radar working mode pulse data using a time slicing method, and then maps it into a plurality of characteristic radar pulse sequence data, and the radar working mode pulse data is represented as X={x1,x2,…,x t}, where x t Time domain data representing a radar pulse train.

[0007] Preferably, the number of radar pulse sequence data formed by mapping in S1 is consistent with the number of nodes in the GNN network constructed in S2.

[0008] Preferably, S2 constructs a three-layer GNN network containing an attention mechanism.

[0009] Preferably, the expression of the context vector of the target node surrounding information extracted by S3 using the attention mechanism is: (1), in, v represents the target node, u represents the adjacent nodes, N ( v ) represents all adjacent nodes of the target node, C v The context vector representing the surrounding information of the target node extracted by the attention mechanism, M represents the attention mechanism function, h v represents the feature represented by the target node, h u represents the features represented by the adjacent nodes, k Indicates the update times of the target node; The update formula for updating the target node based on the context vector of the surrounding information is: (2), in, GNN Represents a graph neural network.

[0010] Preferably, the specific steps of said S3 supplementing and updating the target node according to the context vector of the surrounding information are: S3.1. Traverse the neighboring nodes of the target node, find other neighboring nodes except the target node, and transform and aggregate the node information of other neighboring nodes and the target node into the corresponding neighboring nodes; S3.2. Aggregate the information of adjacent nodes into the target node through the GNN network to complete an update of the target node; S3.3. Repeat S3.1-S3.2 until the features represented by all nodes are aggregated into the target node.

[0011] Preferably, the GNN network in S3 is described as an edge from node B to node A, and the edge is the weight of the feature connecting the nodes, expressed as: (3), in, x A Representation Node A Input radar working mode pulse data, x B Representation Node A Neighboring nodes B Input radar working mode pulse data, x A ∈ R F , e B,A Represents a slave node B The attributes of the edge to node A, e B,A ∈ R D , R F and R D Respectively F dimensional space vector and D dimensional space vector, k express k Layer, which is the number of node updates. F and c represents a multi-layer perceptron, f represents a differentiable, permutation-invariant function, N ( A ) represents all adjacent nodes of node A; The S4 inputs the last layer of target nodes in the GNN network into the classification model.

[0012] Preferably, the classification model is optimized by training in S4, and the training process is as follows: S4.1. Establish a hypothetical function. The expression of the hypothetical function is: (4), in, h θ ( x （i） ) represents the hypothesis function, x （i） and y （i） Respectively represent the input radar pulse sequence data and its classification result label, i represents the parameters of the classifier,e represents the natural logarithm, t Represents the time parameter, i Indicates the input data sample label of the current classification, j Indicates the classification number, n Indicates the number of categories of classification results, T Indicates network moment; The probability of each classification result is represented by the hypothesis function, which is expressed as: (5) S4.2. According to the maximum likelihood estimation classifier loss function, the loss function is expressed as: (6) in, J represents the loss function, m represents the number of training samples, o Represents the output of each category of classification; S4.3. Use gradient descent method to train and update parameters.

[0013] Preferably, the activation function in the classification model adopts the ReLU activation function.

[0014] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects: The present invention relates to a radar working mode recognition method based on a combination of GNN and an attention mechanism. Aiming at the unique properties of radar signal data, GNN is introduced for the recognition task of radar working mode. On this basis, considering that the traditional GNN model may ignore the difference in the importance of different features in the radar signal to the recognition result, the present invention further integrates the attention mechanism, extracts the context vector of the surrounding information of the target node through the attention mechanism, supplements and updates the target node according to the context vector of the surrounding information based on the previous representation of the target node, so that the model can automatically adjust the weights of different features, thereby significantly improving the accuracy and robustness of radar working mode recognition, effectively overcoming the limitations of the single GNN model in radar signal processing, and providing a new technical means for the intelligentization of radar systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of a radar working mode recognition model based on the combination of GNN and attention mechanism in the present invention; Figure 2 This is a schematic diagram of the GNN radar working mode recognition network structure of the present invention; Figure 3 Schematic diagram of the process of updating the target node based on the context vector of surrounding information; Figure 4Schematic diagram of the classifier structure containing the attention mechanism of the present invention. DETAILED DESCRIPTION

[0016] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with examples. The following examples are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0017] See attached Figure 1 As shown, the present invention relates to a radar working mode recognition method based on the combination of GNN and attention mechanism, which comprises the following steps: S1. Input a radar working mode pulse data set, and map it into a plurality of characteristic radar pulse sequence data. In this embodiment, the radar working mode pulse data is sliced ​​by a time slicing method, and then mapped into a plurality of characteristic radar pulse sequence data. The radar working mode pulse data is represented by X={x1,x2,…,x t}, where x t Time domain data representing a radar pulse train.

[0018] S2. Construct a three-layer GNN network with an attention mechanism. Each layer of the GNN network is equipped with nodes equal to the number of radar pulse sequence data. Each node represents a type of radar working mode and carries the corresponding radar signal characteristics. Initialize the target node, input the radar pulse sequence data with target characteristics into the target node, and input the remaining radar pulse sequence data into other corresponding nodes; the nodes here represent the feature classification of radar working mode pulse data recognition.

[0019] S3, GNN network uses the attention mechanism to extract the context vector of the target node's surrounding information, and supplements and updates the target node based on the context vector of the surrounding information; See attached Figure 2 As shown in the figure, the expression of the attention mechanism to extract the context vector of the surrounding information of the target node is: (1), in, v represents the target node, u represents the adjacent nodes, N ( v ) represents all adjacent nodes of the target node, C v The context vector representing the surrounding information of the target node extracted by the attention mechanism, M represents the attention mechanism function, h v represents the feature represented by the target node, h u represents the features represented by the adjacent nodes, kIndicates the update times of the target node; Based on the previous representation of the target node, the target node is updated according to the context vector of the surrounding information. The update formula is: , in, GNN Represents a graph neural network.

[0020] See attached Figure 3 As shown, the specific steps of updating the target node are: S3.1. Traverse the neighboring nodes of the target node, find other neighboring nodes except the target node, and transform and aggregate the node information of other neighboring nodes and the target node into the corresponding neighboring nodes; S3.2. Aggregate the information of adjacent nodes into the target node through the GNN network, and complete a node representation update of the target node; S3.3. Repeat S3.1-S3.2 until the features represented by all nodes are aggregated into the target node.

[0021] In this step, the GNN network is described as an edge from node B to node A, expressed as: (3), in, x A Representation Node A Input radar working mode pulse data, x A ∈ R F , when the node A When it is the target node, it can be considered x A Equivalent to the feature represented by the target node described in formula (1) h v , x B Representation Node A Neighboring nodes B Input radar working mode pulse data, e B,A represents the attributes of the edge from node B to node A, e B,A ∈ R D , R F and R D Respectively F dimensional space vector and D dimensional space vector, k express k layer,k That is, the number of node updates. The number of updates corresponds to the number of layers of nodes in the model structure. F and c represents a multi-layer perceptron, f Represents a differentiable, permutation-invariant function; The last layer of target nodes in the GNN network is input into the classification model.

[0022] S4. Refer to the attached Figure 4 As shown in the figure, the final target node output by the GNN network is input into a classification model consisting of two fully connected layers and an activation function to predict the radar working mode classification of the target node.

[0023] Prior to this, the classification model needs to be trained and optimized. The training process is as follows: S4.1. Establish a hypothesis function, assuming that the training set is and ,in is the number of samples, assuming the function expression is: (4), in, h θ ( x （i） ) represents the hypothesis function, x （i） and y （i） Respectively represent the input radar pulse sequence data and its classification result label, i represents the parameters of the classifier, e represents the natural logarithm, t Represents the time parameter, i Indicates the input data sample label of the current classification, j Indicates the current classification category number. n Indicates the number of categories of classification results, T Indicates network moment; The hypothesis function represents the probability of each classification result, which is expressed as: (5) The probability of the classification result indicates that when the input data only corresponds to a certain working mode, One of the outputs is close to 1, and the others are approximately 0; when the input data includes two or more working modes, Some of the outputs are not 0, and the rest are approximately 0; S4.2. According to the maximum likelihood estimation classifier loss function, the loss function is expressed as: (6) in, J represents the loss function, m represents the number of training samples, o Represents the output of each classification, , ; S4.3. Use gradient descent method to train and update parameters; The activation function in the classification model adopts the ReLU activation function.

[0024] The features represented by the final target node output by the GNN network are input into the above classification model, and finally a two-layer fully connected layer and an activation function are used. The constructed classification model predicts the target node The classification is: (7), in, A represents the target node, y A Indicates the classification of the target node.

[0025] The present invention is described in detail above in combination with the embodiments, but the contents described are only preferred embodiments of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A radar working mode recognition method based on the combination of GNN and attention mechanism, characterized by: It includes the following steps: S1, inputting a radar working mode pulse data set and mapping it into a plurality of characteristic radar pulse sequence data; S2. Construct a GNN network with an attention mechanism, initialize a target node, the node is used to represent the feature classification of pulse data recognition in the radar working mode, and input the radar pulse sequence data with target characteristics into the target node; S3, GNN network uses the attention mechanism to extract the context vector of the target node's surrounding information, and supplements and updates the target node based on the context vector of the surrounding information; S4. Input the final target node output by the GNN network into a classification model consisting of two fully connected layers and an activation function to predict the radar working mode classification of the target node.

2. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 1 is characterized in that: The S1 uses the time slicing method to slice the radar working mode pulse data, and then maps it into a plurality of characteristic radar pulse sequence data. The radar working mode pulse data is represented as X ={ x 1, x 2,…, x t }, in, x t Time domain data representing a radar pulse train.

3. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 2 is characterized in that: The number of radar pulse sequence data mapped in S1 is consistent with the number of nodes in the GNN network constructed in S2.

4. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 1 is characterized in that: The S2 constructs a three-layer GNN network with an attention mechanism.

5. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 1 is characterized in that: The expression of the context vector of the target node surrounding information extracted by S3 using the attention mechanism is: （1）， in, v represents the target node, u represents the adjacent nodes, N ( v ) represents all adjacent nodes of the target node, C v The context vector representing the surrounding information of the target node extracted by the attention mechanism, M represents the attention mechanism function, h v represents the feature represented by the target node, h u represents the features represented by the adjacent nodes, k Indicates the update times of the target node; The update formula for updating the target node based on the context vector of the surrounding information is: （2）， in, GNN Represents a graph neural network.

6. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 5 is characterized in that: The specific steps of S3 supplementing and updating the target node according to the context vector of the surrounding information are: S3.

1. Traverse the neighboring nodes of the target node, find other neighboring nodes except the target node, and transform and aggregate the node information of other neighboring nodes and the target node into the corresponding neighboring nodes; S3.

2. Aggregate the information of adjacent nodes into the target node through the GNN network to complete an update of the target node; S3.

3. Repeat S3.1-S3.2 until the features represented by all nodes are aggregated into the target node.

7. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 4 is characterized in that: The GNN network in S3 is described as an edge from node B to node A, and the edge is the weight of the feature connecting the nodes, expressed as: （3）， in, x A Representation Node A Input radar working mode pulse data, x B Represents the adjacent nodes of node A B Input radar working mode pulse data, x A ∈ R F , e B,A Represents a slave node B To Node A The properties of the edges, e B,A ∈ R D , R F and R D Respectively F dimensional space vector and D dimensional space vector, k express k Layer, which is the number of node updates. Ф and γ represents a multi-layer perceptron, f represents a differentiable, permutation-invariant function, N ( A ) represents a node A All adjacent nodes of The S4 inputs the last layer of target nodes in the GNN network into the classification model.

8. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 1 is characterized in that: In S4, the classification model is optimized through training, and the training process is as follows: S4.

1. Establish a hypothetical function. The expression of the hypothetical function is: （4）， in, h θ ( x （i） ) represents the hypothesis function, x （i） and y （i） Respectively represent the input radar pulse sequence data and its classification result label, θ represents the parameters of the classifier, e represents the natural logarithm, τ Represents the time parameter, i Indicates the input data sample label of the current classification, j Indicates the current classification category number. n Indicates the number of categories of classification results, T Indicates network moment; The probability of each classification result is represented by the hypothesis function, which is expressed as: (5)， S4.

2. According to the maximum likelihood estimation classifier loss function, the loss function is expressed as: (6)， in, J represents the loss function, m represents the number of training samples, o Represents the output of each category of classification; S4.

3. Use gradient descent method to train and update parameters.

9. The radar working mode recognition method based on the combination of GNN and attention mechanism according to claim 1 is characterized in that: The activation function in the classification model adopts the ReLU activation function.

Images