Ship radiation source formation identification method and device based on graph neural network, and medium

Abstract

The invention discloses a ship radiation source formation identification method and device based on a graph neural network and a medium, and belongs to the field of electronic reconnaissance, and the method comprises the steps: forming a radiation source adjacent matrix through the position features and electromagnetic signal features between group radiation sources; constructing a hierarchical graph coding and scoring network based on a graph neural network based on the radiation source adjacent matrix; and identifying fleet information by using the hierarchical graph coding and scoring network. The problems that when a new radiation source appears, existing knowledge cannot distinguish military and civilian attributes of the new radiation source and parameters of a military radiation source and a civilian radiation source are close and difficult to distinguish are solved.

Description

Technical Field

[0001] The present invention relates to the field of electronic reconnaissance, and more specifically, to a method, device and medium for identifying a ship radiation source formation based on a graph neural network. Background Art

[0002] With the development of electronic reconnaissance technology, the sensitivity of sensors has been continuously improved, and more and more signals can be intercepted. At the same time, with the development of radar technology, the types, quantities and working conditions of ship radars have become more and more numerous, and the data types are complex. It is difficult to effectively judge and detect new radiation sources based on the detected radar pulse description words (PDW). Most of the existing electromagnetic data fusion associations only sort and identify from the radar PDW, but ignore the positional relationship between ships and the cooperative working relationship between shipborne radars, resulting in the inability to effectively distinguish the radiation source characteristics of military and civilian ships.

[0003] For example, the patent application scheme of "A method for analyzing and predicting electronic target sequences based on graph neural networks" with publication number CN116340822A uses graph convolutional neural networks to analyze electronic target sequences in real-time scenarios. The sequences include all platform targets, electronic targets, and various parameters of targets in the scene, including various types of electronic equipment, such as radars, radio stations, etc., and also include various dimensional relationships between all nodes, such as command relationships, association relationships, communication relationships, and various information such as dynamic changes in sequences, and outputs which known electronic target sequence the electronic target sequence belongs to. The input conditions of this patent application are communication, command, and association relationships between nodes, and the possible sequence structure is output. The problem with this method is that it is difficult to obtain the command and association relationships between nodes in actual non-cooperative electronic reconnaissance. In addition, although this method also uses graph neural networks to process electronic target data, it does not consider the position information between targets, and only judges the electronic target sequence based on electromagnetic characteristics.

[0004] The patent application scheme of "Aircraft Formation Identification Method Based on Multi-Observation Point Shape Context" with announcement number CN111914910A constructs multiple observation points based on Archimedean spiral; establishes shape context operator describing global information; establishes probability density function describing local information; establishes similarity metric function to distinguish formation type. The problem with this method is that it can only be suitable for a few special scenarios by using observation points on Archimedean spiral to observe aircraft formation. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method, device and medium for identifying ship radiation source formations based on graph neural networks. It specifically proposes to use the radiation source characteristics and position information of ships to find radiation source group targets that meet the fleet characteristics among massive electromagnetic targets, and to achieve deep-level law mining of radiation source group behavior, laying the foundation for intelligence work such as continuous monitoring and tracking of fleet targets, extraction of target activity laws and anomaly detection.

[0006] The object of the present invention is achieved through the following solutions:

[0007] A method for identifying ship radiation source formation based on graph neural network includes the following steps:

[0008] The location characteristics and electromagnetic signal characteristics of the group radiation sources are used to form a radiation source adjacency matrix; based on the radiation source adjacency matrix, a hierarchical graph encoding and scoring network based on graph neural network is constructed;

[0009] Fleet information is identified using the hierarchical graph encoding and scoring network.

[0010] Furthermore, the electromagnetic signal characteristics include carrier frequency, repetition frequency and pulse width.

[0011] Furthermore, the method of utilizing the position characteristics and electromagnetic signal characteristics between the group radiation sources to form a radiation source adjacency matrix specifically includes the following sub-steps:

[0012] The hierarchical graph encoding and scoring network is defined as HGSESN(A,X); where A is the adjacency matrix and X is the attribute matrix;

[0013] Assume that there are n radiation sources in a reconnaissance, then the size of the A matrix is ​​n×n; for the undirected graph A, assume that the maximum spacing between the radiation sources is k, then:

[0014]

[0015] A(i,j) represents the adjacency matrix, d i,j represents the distance of the ship in row i and column j;

[0016] The position information between radiation sources is converted into an adjacency matrix that can be input into the neural network; the size of the X matrix is ​​n×m, where m is the number of radiation source attribute information;

[0017] Through conversion, the radiation source location information and attribute information obtained by electronic reconnaissance are converted into two matrices A and X.

[0018] Furthermore, the construction of a hierarchical graph encoding and scoring network based on a graph neural network based on the radiation source adjacency matrix specifically includes the following sub-steps:

[0019] The hierarchical graph encoding and scoring network HGSESN (A, X) includes a graph isomorphism network GIN, which uses the graph isomorphism network GIN to globally pool the input graph structure data and finally obtain the score of each child node.

[0020] Furthermore, the hierarchical graph coding and scoring network HGSESN (A, X) also includes a graph coding network, and the graph coding network is composed of several layers of GCN.

[0021] Furthermore, the graph isomorphism network GIN is used to globally pool the input graph structure data, and finally the score of each child node is obtained, which specifically includes the following sub-steps:

[0022] c (k+1) (v) = GINConv(c (k) (v),{c (k) (u)} u∈A(u) )

[0023] Among them, c (k+1) (v) is the new feature after v nodes are aggregated, c (k) (u) is the feature of the adjacent nodes of node v, GINConv is a graph isomorphic network, c (k) (v) is the feature of node v before aggregation, u is the adjacent node of node v, and A(u) is the adjacency matrix;

[0024] After the graph isomorphism network operation, the characteristics of each node include not only its own characteristics but also the characteristics of its adjacent nodes. Finally, the attributes of each node are pooled into the score of the node using the multilayer perceptron (MLP). Through the GIN operation, the score of each node is obtained as L s .

[0025] Furthermore, the method further comprises the steps of:

[0026] Output the subgraph in the graph data to GCN to obtain the subgraph score L sub , input the full image into the image encoder to get L F , then the final score of each node is L Node for:

[0027] L Node =L s -(L sub -L F )

[0028] By operating the node scores through the activation function, the fleet information can be identified in the interference radiation source through the hierarchical graph encoding and scoring network HGSESN (A, X). It should be noted that after the activation function, the radiation sources with outstanding scores can be screened out and the interference radiation sources with low scores can be filtered out, thereby identifying the fleet information.

[0029] Furthermore, the activation function includes a softmax function.

[0030] A ship radiation source formation identification device based on graph neural network includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, any method described above is executed.

[0031] A computer-readable storage medium stores a computer program, wherein the computer program executes any of the above methods when loaded by a processor.

[0032] The beneficial effects of the present invention include:

[0033] The method for identifying ship radiation source formations based on graph neural networks proposed in this invention can solve the problem that when a new radiation source appears, existing knowledge cannot distinguish its military and civilian attributes, and the problem that the parameters of military and civilian radiation sources are close and difficult to distinguish. Its military and civilian attributes can be determined by the formation position of the formation and the combination relationship between the radiation sources.

[0034] The present invention uses a graph neural network structure to analyze the location and attribute information of the radiation source in the current scenario. A graph autoencoder is used to aggregate the features and edge parameters of the radiation source nodes, and each node is scored to generate new graph features. In order to prevent the data of abnormal nodes from being overwhelmed by normal nodes during the convolution process, a graph encoder network is designed to increase the description of nodes from the perspective of the full graph and subgraph, and finally the fleet radiation source can be identified in the presence of a large amount of background noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0036] Figure 1 The distribution of target radiation sources under a lot of background noise;

[0037] Figure 2 Encoding and scoring networks for hierarchical graphs;

[0038] Figure 3 Added test data of ship formations and civilian radiation sources to the experimental scenario;

[0039] Figure 4 The test data is encoded according to the encoding method of the present invention and then input into the layered graph encoding and scoring network processing proposed by the present invention to obtain the test result. DETAILED DESCRIPTION

[0040] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0041] The specific implementation process of the present invention is as follows:

[0042] like Figure 1 As shown, with the development of electronic reconnaissance technology, the sensitivity of sensors continues to increase, and more and more signals can be intercepted. Electronic reconnaissance will discover a large number of electromagnetic radiation sources. Only measuring individual radiation sources often cannot distinguish between military and civilian radiation sources.

[0043] The present invention includes intelligent association and fusion of electromagnetic data from heterogeneous sources on multiple platforms and mining of electromagnetic perception target activity patterns. The intelligent association and fusion of electromagnetic data from heterogeneous sources on multiple platforms conduct feature similarity analysis and fusion from two aspects: electromagnetic data attribute characteristics and track characteristics; the mining of electromagnetic perception target activity patterns intends to analyze and mine from two aspects: the multi-dimensional attribute change law of a single radiation source and the correlation relationship between multiple radiation sources.

[0044] Specifically, Figure 2 The structure of the hierarchical graph encoding and scoring network constructed by the present invention. In order to solve the problem that the radiation source characteristics of military and civilian ships cannot be effectively distinguished in a dense and complex electromagnetic signal environment, the present invention conceives to use the position characteristics between group radiation sources and electromagnetic signal characteristics such as carrier frequency, repetition frequency, pulse width, etc., and proposes a radiation source adjacency matrix construction method and a hierarchical graph encoding and scoring network (HGESN) based on graph neural network.

[0045] The definition of the Hierarchical Graph Encoding and Scoring Network is HGSESN(A,X). A is the adjacency matrix and X is the attribute matrix. Assuming there are n radiation sources in a reconnaissance, the size of the A matrix is ​​n×n.

[0046] For an undirected graph A, assuming that the maximum spacing between the radiation sources is k, then:

[0047]

[0048] In this way, the position information between the radiation sources is converted into an adjacency matrix that can be input into the neural network. The size of the X matrix is ​​n×m, where m is the number of radiation source attribute information. Through the above conversion, the radiation source position information and attribute information obtained by electronic reconnaissance are converted into two matrices A and X.

[0049] The Hierarchical Graph Encoding and Scoring Network HGSESN (A, X) consists of a graph isomorphism network and a graph encoding network. The graph isomorphism network (GIN) is used to globally pool the input graph structure data, and finally obtain the score of each child node. The specific operations are as follows:

[0050] c (k+1) (v) = GINConv(c (k) (v),{c (k) (u)} u∈A(u) )

[0051] where c (k) (u) is the feature of the node v’s neighboring nodes, c (k+1) (v) is the new feature after the aggregation of v nodes. After the graph isomorphism network operation, the features of each node include not only its own features but also the features of its adjacent nodes. Finally, MLP is used to pool the attributes of each node into the score of the node. Finally, through the GIN operation, the score of each node is L s .

[0052] However, if all the nodes near an abnormal node (military radiation source) are normal nodes, the score of the abnormal node will decrease. In this case, the present invention designs a graph encoder network after the GIN network. The graph encoder network consists of several layers of GCN. Output the subgraph in the graph data to GCN to obtain the subgraph score L sub , input the full image into the image encoder to get L F , then the final score of each node is:

[0053] L Node =L s -(L sub -L F )

[0054] According to the node scores and the softmax operation, the fleet information can be identified among many interference radiation sources through the hierarchical graph encoding and scoring network HGSESN (A, X).

[0055] like Figure 3 As shown, the experimental scenario is to randomly set up 4 ship formations, the number of radiation sources in each ship formation is between 4 and 8, and 1,000 civilian radiation sources are randomly added.

[0056] After the test data is encoded according to the encoding method of the present invention and input into the hierarchical graph encoding and scoring network proposed by the present invention, the obtained result is as follows: Figure 4 As shown, it can be seen that the radiation sources of the four formations are all accurately identified.

[0057] In summary, the present invention, through the study of the ship radiation source formation identification technology based on graph neural network, proposes to use the radiation source characteristics and position information of the ship to find the radiation source group targets that meet the fleet characteristics among massive electromagnetic targets, and realizes the deep mining of the radiation source group behavior rules, laying the foundation for intelligence work such as continuous monitoring and tracking of fleet targets, extraction of target activity rules and anomaly detection.

[0058] It should be noted that within the scope of protection defined in the claims of the present invention, the following embodiments can be combined and / or expanded or replaced in any logical way from the above specific implementation methods, such as disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0059] Example 1

[0060] A method for identifying ship radiation source formation based on graph neural network includes the following steps:

[0061] The location characteristics and electromagnetic signal characteristics of the group radiation sources are used to form a radiation source adjacency matrix; based on the radiation source adjacency matrix, a hierarchical graph encoding and scoring network based on graph neural network is constructed;

[0062] Fleet information is identified using the hierarchical graph encoding and scoring network.

[0063] Example 2

[0064] Based on Example 1, the electromagnetic signal characteristics include carrier frequency, repetition frequency and pulse width.

[0065] Example 3

[0066] On the basis of Example 1, the method of using the position characteristics and electromagnetic signal characteristics between the group radiation sources to form a radiation source adjacency matrix specifically includes the following sub-steps:

[0067] The hierarchical graph encoding and scoring network is defined as HGSESN(A,X); where A is the adjacency matrix and X is the attribute matrix;

[0068] Assume that there are n radiation sources in a reconnaissance, then the size of the A matrix is ​​n×n; for the undirected graph A, assume that the maximum spacing between the radiation sources is k, then:

[0069]

[0070] Among them, A(i,j) represents the adjacency matrix, di,j represents the distance of the ship in row i and column j;

[0071] The position information between radiation sources is converted into an adjacency matrix that can be input into the neural network; the size of the X matrix is ​​n×m, where m is the number of radiation source attribute information;

[0072] Through conversion, the radiation source location information and attribute information obtained by electronic reconnaissance are converted into two matrices A and X.

[0073] Example 4

[0074] On the basis of Example 3, the hierarchical graph coding and scoring network based on the graph neural network is constructed based on the radiation source adjacency matrix, which specifically includes the following sub-steps:

[0075] The hierarchical graph encoding and scoring network HGSESN (A, X) includes a graph isomorphism network GIN, which uses the graph isomorphism network GIN to globally pool the input graph structure data and finally obtain the score of each child node.

[0076] Example 5

[0077] On the basis of Example 4, the hierarchical graph coding and scoring network HGSESN (A, X) further includes a graph coding network, which is composed of several layers of GCN.

[0078] Example 6

[0079] On the basis of Example 5, the graph isomorphism network GIN is used to perform global pooling on the input graph structure data, and finally obtain the score of each child node, which specifically includes the following sub-steps:

[0080] c (k+1) (v) = GINConv(c (k) (v),{c (k) (u)} u∈A(u) )

[0081] Among them, c (k+1) (v) is the new feature after v nodes are aggregated, c (k) (u) is the feature of the adjacent nodes of node v, GINConv is a graph isomorphic network, c (k) (v) is the feature of node v before aggregation, u is the adjacent node of node v, and A(u) is the adjacency matrix;

[0082] After the graph isomorphism network operation, the characteristics of each node include not only its own characteristics but also the characteristics of its adjacent nodes. Finally, the attributes of each node are pooled into the score of the node using the multilayer perceptron (MLP). Through the GIN operation, the score of each node is obtained as Ls .

[0083] Example 7

[0084] On the basis of Example 6, the method further comprises the following steps:

[0085] Output the subgraph in the graph data to GCN to obtain the subgraph score L sub , input the full image into the image encoder to get L F , then the final score of each node is L Node for:

[0086] L Node =L s -(L sub -L F )

[0087] By operating the node scores through the activation function, the fleet information can be identified in the interference radiation source through the hierarchical graph encoding and scoring network HGSESN (A, X). It should be noted that after the activation function, the radiation sources with outstanding scores can be screened out and the interference radiation sources with low scores can be filtered out, thereby identifying the fleet information.

[0088] Example 8

[0089] Based on Example 7, the activation function includes a softmax function.

[0090] Example 9

[0091] A ship radiation source formation identification device based on graph neural network includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, the method described in any one of Examples 1 to 8 is executed.

[0092] Example 10

[0093] A computer-readable storage medium stores a computer program, and when the computer program is loaded by a processor, the method described in any one of Embodiments 1 to 8 is executed.

[0094] The units involved in the embodiments of the present invention may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not, in some cases, limit the units themselves.

[0095] According to one aspect of an embodiment of the present invention, a computer program product or a computer program is provided, the computer program product or the computer program includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. A processor of a computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in the above various optional implementations.

[0096] As another aspect, an embodiment of the present invention further provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiment.

Claims

1. A method for identifying ship radiation source formations based on graph neural networks, characterized in that: The steps include: The location characteristics and electromagnetic signal characteristics of the group radiation sources are used to form a radiation source adjacency matrix; based on the radiation source adjacency matrix, a hierarchical graph encoding and scoring network based on graph neural network is constructed; Fleet information is identified using the hierarchical graph encoding and scoring network.

2. The method for identifying ship radiation source formations based on graph neural network according to claim 1 is characterized in that: The electromagnetic signal characteristics include carrier frequency, repetition frequency and pulse width.

3. The method for identifying ship radiation source formations based on graph neural network according to claim 1 is characterized in that: The method of using the position characteristics and electromagnetic signal characteristics between the group radiation sources to form a radiation source adjacency matrix specifically includes the following sub-steps: The hierarchical graph encoding and scoring network is defined as HGSESN(A,X); where A is the adjacency matrix and X is the attribute matrix; Assume that there are n radiation sources in a reconnaissance, then the size of the A matrix is ​​n×n; for the undirected graph A, assume that the maximum spacing between the radiation sources is k, then: A(i,j) represents the adjacency matrix, d i,j represents the distance of the ship in row i and column j; The position information between radiation sources is converted into an adjacency matrix that can be input into the neural network; the size of the X matrix is ​​n×m, where m is the number of radiation source attribute information; Through conversion, the radiation source location information and attribute information obtained by electronic reconnaissance are converted into two matrices A and X.

4. The method for identifying ship radiation source formations based on graph neural network according to claim 3 is characterized in that: The method of constructing a hierarchical graph encoding and scoring network based on a graph neural network based on the radiation source adjacency matrix specifically includes the following sub-steps: The hierarchical graph encoding and scoring network HGSESN (A, X) includes a graph isomorphism network GIN, which uses the graph isomorphism network GIN to globally pool the input graph structure data and finally obtain the score of each child node.

5. The method for identifying ship radiation source formations based on graph neural network according to claim 4 is characterized in that: The hierarchical graph coding and scoring network HGSESN (A, X) also includes a graph coding network, and the graph coding network is composed of several layers of GCN.

6. The method for identifying ship radiation source formations based on graph neural network according to claim 5 is characterized in that: The method of using the graph isomorphism network GIN to globally pool the input graph structure data and finally obtain the score of each child node specifically includes the following sub-steps: c (k+1) (v)=GINConv(c (k) (v),{c (k) (u)} u∈A ( u) ) Among them, c (k+1) (v) is the new feature after v nodes are aggregated, c (k) (u) is the feature of the adjacent nodes of node v, GINConv is a graph isomorphic network, c (k) (v) is the feature of node v before aggregation, u is the adjacent node of node v, and A(u) is the adjacency matrix; After the graph isomorphism network operation, the characteristics of each node include not only its own characteristics but also the characteristics of its adjacent nodes. Finally, the multi-layer perceptron MLP is used to pool the attributes of each node into the score of the node. Through the GIN operation, the score of each node is obtained as L s .

7. The method for identifying ship radiation source formations based on graph neural network according to claim 6 is characterized in that: Also includes the steps: Output the subgraph in the graph data to GCN to obtain the subgraph score L sub , input the full image into the image encoder to get L F , then the final score of each node is L Node for: L Node =L s -(L sub -L F ) The node scores are operated through activation functions, and the fleet information can be identified in the interference radiation source through the hierarchical graph encoding and scoring network HGSESN (A, X).

8. The method for identifying ship radiation source formations based on graph neural network according to claim 7 is characterized in that: The activation function includes a softmax function.

9. A ship radiation source formation identification device based on graph neural network, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, the method according to any one of claims 1 to 8 is executed.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a readable storage medium, and when the computer program is loaded by a processor, the method according to any one of claims 1 to 8 is executed.

Images