CN116796284A - Unmanned aerial vehicle cluster-oriented intention recognition method applied to countermeasure environment - Google Patents

Abstract

The invention discloses an intention recognition method for unmanned aerial vehicle clusters in an antagonistic environment. Firstly, threat coefficient similarity and distance similarity masks between unmanned aerial vehicles are calculated by using physical characteristics of unmanned aerial vehicle clusters, and battlefield states are mapped into a graph structure. And then, the graph neural network is utilized to introduce an attention mechanism to process graph structure information, the mutual influence weight between the neighbor unmanned aerial vehicles is calculated, and the unmanned aerial vehicle characteristic information is deeply extracted. And fusing the extracted feature expression by using a pooling method to obtain the comprehensive feature description of the unmanned aerial vehicle cluster. And then high-precision intention classification recognition is realized through supervised learning by utilizing full connection. The invention can identify the intention of the unmanned aerial vehicle cluster under the condition of complex battlefield environment and various deception information. The method has the advantages that the problem of intention recognition of the unmanned aerial vehicle cluster is transferred to the field of graph data processing, and a graph neural network method is introduced, so that the difficulty that expert experience is excessively relied on in the field of intention recognition and recognition accuracy is low is solved.

Description

An intent recognition method for UAV swarms in a confrontational environment

Technical field

The invention relates to an intention recognition method for UAV clusters in a confrontational environment, and belongs to the technical field of UAV intention identification.

Background technique

Unmanned equipment is becoming more and more popular among people, and the importance of drones is also reflected in modern warfare. Under the premise that UAV swarm operations bring huge scale advantages, research on countermeasures for UAV swarm operations is also an important topic. However, in the field of air defense, there are problems such as "difficulty in early warning detection and command and control". UAV swarms are mostly micro-UAVs with weak infrared characteristics, small radar reflection surfaces, and low flight altitudes. In addition, the swarm system Most of them have adaptive capabilities such as autonomous formation and intelligent decision-making, making countermeasures against UAV swarms more difficult. Faced with various challenges, the technology to identify the intentions of hostile drones and their swarms has become particularly critical. Accurately and quickly predicting the enemy's combat intentions will help the air combat offensive and defensive decision-making system make more appropriate decision-making behaviors, and can help our own side seize air supremacy and even win the war.

In the problem of intention recognition, the intention of the cluster is displayed through a series of time-series actions of a plurality of individual drones. Due to the hidden and complicated interactions within the cluster, it is an urgent matter to realize the intention recognition of the cluster in the complex and changeable battlefield environment. The problem.

Contents of the invention

The technical problem solved by the present invention is to overcome the shortcomings of the existing technology and propose an intention recognition method for UAV clusters in a confrontational environment. This method can conceal complex interactions within the cluster and the battlefield environment is complex and changeable. Under such circumstances, accurate identification of cluster intentions is achieved. In terms of data preprocessing, threat coefficient similarity and distance similarity masking methods are used to construct clustered graph structure data. In terms of intent recognition, the graph attention mechanism method and the graph pooling method are used to extract the deep features of the clustered drones, and then combined with the fully connected network to identify the intent of the drone cluster.

In order to achieve the above objects, the technical solutions provided by the present invention are:

An intent recognition method for UAV swarms applied in a confrontational environment, including the following steps:

The first step is to calculate the threat coefficient and spatial distance mask of the UAV cluster based on the physical characteristics of the UAV cluster at the same time, and construct the UAV based on the calculated threat coefficient and spatial distance mask of the UAV cluster. Graph structured data form of clusters;

In the second step, use the graph attention network to process the graph structure data obtained in the first step to obtain the deep feature information of each drone;

The third step is to obtain the deep feature information of the entire UAV cluster based on the deep feature information of each UAV obtained in the second step, and input the obtained deep feature information of the entire UAV cluster into the fully connected network. After supervised training and learning, the classification of UAV swarm intentions is achieved. The classification of UAV swarm intentions includes attack, feint attack, retreat, electronic jamming, and low-altitude raid.

In the first step, the physical characteristics include the spatial position, heading angle, pitch angle, speed, radar on status, and drone type of each drone;

In the first step, assume that the blue machine cluster B has a total of M drones B _i (=1,2...M), and the red machine cluster R has a total of N drones R _j (j=1,2...N) , the threat coefficient of UAV R _j to UAV B _i includes distance threat coefficient Angle threat coefficient/> Speed threat coefficient/> and air combat effectiveness threat coefficient thR _ji R _ij , the formula is as follows:

In the formula, thR _ji D _ji ∈[0,1], D _ji is the distance between UAV R _j and UAV B _i ; λ ₁ , λ ₂ , ε are constant parameters; R _aR is the effective range of the red aircraft missile; T _rR is the maximum effect of the red aircraft missile distance;/> is the speed of drone B _i ;/> is the speed of UAV R _j ; C _i is the air combat effectiveness index of UAV B _i , and C _j is the air combat effectiveness index of UAV R _j ;

Assume that the distance mask between the red drone R _j and the red drone R _j′ (j′=1,2…N) is DI _jj′ , and its formula is as follows:

DI _jj′ =INT(d(R _j′ ,R _j )≤μ)

Among them, μ is a constant, μ∈[0.6,0.75]; INT represents the true value conversion function; d(R _j′ ,R _j ) represents the relative distance between the drone R _j and the red drone R _j′ ;

According to the distance threat coefficient, angle threat coefficient, speed threat coefficient, air combat effectiveness threat coefficient and distance mask, the graph structure data of the constructed UAV cluster is:

Among them, j″=1,2…N, th _jj′ is the UAV R _j and UAV calculated through the gray correlation matrix method using the distance threat coefficient, angle threat coefficient, speed threat coefficient and air combat effectiveness threat coefficient. The comprehensive threat coefficient of machine R _j′ to blue team;

G _jj′∈ [0,1]

Select the maximum value of G _jj′ obtained, and set G _jz as the maximum value, then connect the drone R _j to the drone R _z to complete the construction of the graph data structure of the drone cluster;

In the second step, the method of obtaining deep feature information is: calculating the attention coefficient between the drone R _j and the drone R _j′ , and obtaining the deep feature information based on the calculated attention coefficient;

Among them, the calculation method of the attention coefficient α _jj′ is: use a learnable shared parameter W to enhance the features of the connected UAV R _j and UAV R _j′ , and then splice the enhanced features to The characteristics of UAV R _j and UAV R _j′ are combined into a vector. Finally, mapping α is used to map the synthesized vector to a real number. After normalization by SoftMax, the connected UAV R j and UAV R _j ′ are obtained. The attention coefficient of human-machine R _j′ , the formula is:

in, is the set of other drones in the red machine cluster connected to drone R _j , /> e _jj′ represents the mapping of spliced features;

e _jj′ =α([Wh _j ||Wh _j′ ])

h _j is the set of physical characteristics of UAV R _j ;

Deep feature information of UAV R _j for:

In the third step, the deep feature information of the entire UAV cluster is obtained through the graph pooling operation based on the deep feature information of each UAV. Specifically, let A ^(l) be the lth layer of the pooling layer. For the corresponding ^{adjacency matrix} , inferred:

A ^(l+1) ,X ^(l+1) =DiffPool(A ^(l) ,Z ^(l) )

Among them, DiffPool means to solve the weight matrix S ^(l) through a graph neural network. S ^(l) numerically represents the weight of each node in the upper layer to be allocated to the nodes in the next layer. Find The formula looks like this:

S ^(l) =softmax(GNN _l,pool (A ^(l) ,X ^(l) ))

Among them, GNN _l,pool represents the graph neural network used in graph pooling. The iteration formula between layers in the graph neural network is:

The optimization constraint function of the graph neural network includes the connection regular term constraint and the entropy function regular term constraint. The expression formula of the connection regular term constraint is:

L _LP =||A ^(l) ,S ^(l) S ^(l)T || _F

The expression formula of the regular term constraint of the entropy function is:

Among them, n represents the number of nodes in the current pooling layer, p=1,2...n;

The UAV cluster includes multi-rotor decoy UAVs, integrated investigation and attack UAVs, multi-rotor relay UAVs and small reconnaissance UAVs.

beneficial effects

(1) The method of the present invention uses the drone mapping method combined with the physical information of the drone cluster to build a graph structure model for the drone cluster to describe the internal functional relationships of the drone cluster.

(2) The method of the present invention uses the graph neural network method to solve for the mutual influence weight between two drones in the cluster, and then combines the interaction relationship between drones in the graph structure to calculate and extract the drone cluster. The deep hidden characteristics of each drone in .

(3) The method of the present invention adopts the graph pooling method to integrate individual characteristics in the UAV cluster, and uses the final node characteristics after fusion as the overall characteristics of the UAV cluster. Then, according to the fully connected classification network method, the fused overall features are processed to classify the UAV cluster.

(4) In the method of the present invention, the threat coefficient describes the threat level of each UAV to the enemy, and the similarity degree expresses the degree of coordination between each UAV. Spatial distance mask blocks the relationship between drones by comparing the distance between drones and the masking threshold. The graph structure model for UAV clusters established by this method is the basis for subsequent feature extraction, information concentration and intent recognition.

(5) In the method of the present invention, by minimizing the constraint function, the UAV cluster is condensed to one node in the iterative process. There are two optimization constraint ideas in the iterative process. The first is to make it easier for two UAVs with closer interactions to be mapped to the same enrichment node in the next layer. The second is to try to avoid the occurrence of UAV pairs. The situation where the mapping weights of all condensed nodes in a layer are equal. After obtaining the final cluster condensed node characteristics, the cluster can be identified through the fully connected network.

(6) Compared with classic methods such as template matching and D-S evidence reasoning that rely on expert experience, the present invention eliminates dependence on expert experience by introducing a supervised neural network method, and will not be affected by the lack of reliable, complete and matching Expert experience leads to a significant drop in accuracy. When constructing the graph data structure of the cluster from the physical characteristics of the UAV cluster, the present invention proposes a spatial distance mask to reduce complexity. Group modeling is performed based on the spatial distance threshold between cluster members, and the entire group is broken into parts, which not only reduces the complexity of interactive relationship reasoning, but also improves the sophistication of interactive relationship modeling. Compared with simply and directly weighting the sum of UAVs and their neighbors, this invention pays more attention to the interaction relationship of clusters and uses the graph attention mechanism method proposed in the second step to consider the inequality between UAVs. It realizes the sharing of information features between drones and can extract more representative deep features of drones. Compared with the classic method that directly uses the characteristic average of all UAVs in the cluster to represent the characteristic information of the UAV cluster, this invention can fully consider the correlation within the UAV cluster and use the proposed graph pooling method to Continuously search for similar associated nodes and provide fusion solutions to ensure the rationality of concentration. Through the pooling method of the present invention, the structural information of the graph can be well extracted at a coarse-grained level, and feature vectors that can better express cluster information can be retained.

(7) The present invention discloses an intention recognition method for UAV clusters in a confrontational environment. This method builds a battlefield intention recognition model based on the depth graph neural network, which mainly includes the following steps: first, the physical characteristics such as direction angle, position, speed and so on of the UAV cluster are used to calculate the threat coefficient similarity and distance similarity mask between the UAVs. code to map the battlefield state into a graph structure. Afterwards, the graph neural network is used to introduce the attention mechanism to process the graph structure information, calculate the mutual influence weights between neighboring drones, and perform deep extraction of drone feature information. Afterwards, the graph pooling method is used to fuse the extracted feature expressions to obtain a comprehensive feature description of the UAV cluster. Finally, modules such as full connection are used to build a learning network, and high-precision intent classification and recognition is achieved through supervised learning. The invention can identify the intention of the UAV cluster under the condition that the battlefield environment is complex and there is a lot of deception information. The problem of intention recognition of drone clusters is transferred to the field of graph data processing, and the graph neural network method is introduced, which solves the difficulty of over-reliance on expert experience and low recognition accuracy in the field of intention recognition.

Description of the drawings

Figure 1 is a schematic diagram of the air combat situation;

Figure 2 is a schematic diagram of the attention mechanism;

Figure 3 is a weighted schematic diagram of feature extraction;

Figure 4 is a schematic diagram of graph pooling.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific implementation examples.

An intent recognition method for UAV swarms applied in a confrontational environment, including the following steps:

Step 1: Establish the graph data structure of the cluster based on the physical characteristics of the UAV cluster

The present invention uses threat coefficient similarity and spatial distance mask to construct the graph data structure of the cluster. Next, the blue UAV group B and the red UAV group R are used as examples to illustrate the establishment process of the graph data structure. Assume that cluster B has a total of M drones, and cluster R has a total of N drones. The attack postures of B _i (=1,2...M) and R _j (j=1,2...N) are shown in Figure 1. set up and thR _ji R _ij are the distance threat coefficient, angle threat coefficient, speed threat coefficient and air combat effectiveness threat coefficient of UAV R _j to UAV B _i respectively. Their formulas are as follows:

In the formula: thR _ji D _ji ∈[0,1], D _ji is the distance between UAVs; λ ₁ and λ ₂ are constant parameters; R _aR is the effective range of the red aircraft missile; T _rR is the maximum action range of the red aircraft missile;/> is the speed of drone B _i ;/> is the speed of UAV R _j ; C _i is the air combat effectiveness index of UAV B _i , and C _j is the air combat effectiveness index of UAV R _j ;

Since the status information from neighbor drones is more important than the status information of drones from other locations. Therefore, when the distance between two drones exceeds a certain threshold, they can be considered to have no relationship. Therefore, the present invention proposes a distance mask method to solve the position similarity. Assume that the distance mask between the j-th red-side UAV and the j′-th red-side UAV is DI _jj′ , and its formula is as follows:

DI _jj′ =INT(d(R _j′ ,R _j )≤μ)#(5)

Among them, μ is a constant, μ∈[0.6,0.75].

Based on the distance threat coefficient, angle threat coefficient, speed threat coefficient and air combat effectiveness threat coefficient combined with the distance mask, the composition formula can be given:

The establishment of the formula imitates the form of the sigmoid function, and the sizes of all G _jj′ are between 0 and 1. After calculating the relationship metric between the j-th red drone and all other red drones, select the maximum value G _zj and connect the j-th red drone to the z-th red drone it represents, thereby completing the unmanned aerial vehicle control. Construction of graph data structure for machine clusters.

Step 2: Use graph attention network to extract deep features of UAV

The following problems will be encountered when extracting UAV features in a cluster: For each UAV in the cluster, the influence weight of its neighbor UAVs is different. If this problem is ignored, then UAV The drone cannot effectively utilize the characteristic information of neighboring drones. For example, if you directly perform a weighted summation of the characteristics of neighbor drones, you will definitely ignore the information implicit in the interaction relationship of the drone cluster. Therefore, in step two, it is necessary to first calculate and solve the attention coefficients between drones based on the graph data structure of drone interaction, and then combine the attention coefficients to extract the deep features of the drones.

The extraction method of attention coefficient α _jj′ is shown in Figure 2. The specific process is: first use a learnable shared parameter W to enhance the features of two connected UAVs R _j and UAV R _j′ . The enhanced features are spliced to combine the features of UAV R _j and UAV R _j′ into a vector. Finally, mapping α is used to map the synthesized vector to a real number, which can be obtained after normalization by SoftMax. The attention coefficient of the connected UAV R _j and UAV R _j′ . The formula is:

in, is the set of other drones in the red machine cluster connected to drone R _j , /> e _jj′ represents the mapping of spliced features, and the formula is as follows:

e _jj′ =a([Wh _j ||Wh _j′ ])#(8)

As shown in Figure 3, after obtaining the attention coefficients of all neighbor UAVs, the deep implicit features of the UAV after fusing neighborhood information can be obtained through the weighted sum of the attention coefficients and the original physical features. The specific formula is as follows:

Step 3: Perform graph pooling operations on the graph data structure of the cluster and classify it

The graph pooling operation is similar to the pooling layer in a convolutional neural network, which can continuously condense features and generate higher-level data. The flow chart of graph pooling is shown in Figure 4. Each time graph pooling is passed, nodes marked with the same color in the graph will undergo feature fusion, and the number of nodes in the graph structure will be reduced. After the last pooling operation, only the last node will be retained. Later, the characteristic information of this node can be used to predict the intention of the cluster.

The specific pooling operation can be expressed by the following formula. Let A ^(l) be the adjacency matrix corresponding to the l-th layer, X ^(l) be the feature matrix of all nodes in the l-th layer, Z ^(l) be the feature of each node extracted on the l-th layer graph, Then the graph structure of the next layer can be obtained by the following formula:

A ^(l+1) ,X ^(l+1) =DiffPool(A ^(l) ,Z ^(l) )#(10)

The operation represented by DiffPool is to solve the weight matrix through a graph neural network, and numerically represent the weight of each node in the upper layer to be allocated to the nodes in the next layer. The formula is as follows:

S ^(l) =softmax(GNN _l,pool (A ^(l) ,X ^(l) ))#(11)

Next, perform weighted aggregation according to the learned weight matrix to obtain the next layer of graph structure. The formula is as follows:

At the same time, in order to reduce the learning difficulty of the network, regular terms need to be added to constrain the network. First, the connection regular term constraint needs to be carried out, and the formula is shown in Equation (14), where, Represents the layer node and the connection strength between nodes. The stronger the connection between two nodes, /> The larger the value. /> Represents the row sum/> of S ^(l) The corresponding elements of the column are multiplied and then added. That is to say, the greater the probability that two nodes are mapped to the same condensed node, the corresponding The larger the value. Therefore, by minimizing this constraint, two nodes with greater connection strength can be more easily mapped to the same node in the next layer. The second step is to add the entropy function regular term constraint, the formula is shown in Equation (15). Entropy represents the degree of chaos of the system distribution, and increases the uncertainty of the mapping distribution by reducing entropy. Avoid the situation where the mapping weight of a node to all nodes in the next layer is equal.

L _LP =||A ^(l) ,S ^(l) S ^(l)T |||| _F #(14)

After completing the graph pooling operation, the final condensed node features are fed into a learnable fully-connected neural network to classify the intent of the cluster.

Example

In order to further understand the implementation steps of the present invention, examples are as follows:

There is a red aircraft cluster with N = 3 drones to attack a blue aircraft cluster with M = 5 drones. The physical characteristics of the red aircraft cluster relative to the blue aircraft B ₁ are as follows

By incorporating the above threat coefficient calculation formula, we can get the threat coefficient of each red machine to the blue machine B ₁ :

In the same way, the threat coefficient of each red machine to other blue machines can be obtained.

Since the red machine cluster is a small cluster, the distance between two drones in the cluster is relatively close, so the spatial distance mask is all 1.

The comprehensive threat coefficient G _jj ′ of UAV R _j and UAV R _j′ to the blue side can be calculated through the gray correlation matrix method, and the maximum value of G _jj′ obtained is selected to construct the graph data of the UAV cluster. Structure, the corresponding graph representation matrix in this example is: This matrix indicates that UAV R ₁ and UAV R ₂ are connected, and UAV R ₁ and UAV R ₃ are connected.

Input the obtained graph representation matrix and the original physical feature information of the cluster into the graph neural network, use the graph attention network to calculate the attention coefficient between pairs within the red machine cluster, and obtain the deep feature information based on the calculated attention coefficient. This step is calculated by the neural network model, and the manual calculation process is no longer shown. The final deep feature information of the UAVs R ₁ , R ₂ , and R ₃ are as follows:

Input the deep feature information of the UAVs R ₁ , R ₂ , and R ₃ into the graph pooling network, and continuously minimize the two constraints through parameter updates to obtain the deep feature information of the red machine cluster.

Bundle Input into the trained fully connected network, classify the cluster's intention, and the final output result is offensive, which is consistent with the real label.

The above are only preferred embodiments of the present invention. The present invention is not limited to the above-mentioned embodiments. All modifications, substitutions, combinations, cuts, etc. made within the spirit and principles of the present invention shall be included in within the protection scope of the present invention.

Claims (10)

1. The intent recognition method for the unmanned aerial vehicle cluster in the countermeasure environment is characterized by comprising the following steps of:

firstly, calculating threat coefficients and a spatial distance mask of an unmanned aerial vehicle cluster according to physical characteristics of the unmanned aerial vehicle cluster at the same time, and constructing a graph structure data form of the unmanned aerial vehicle cluster according to the calculated threat coefficients and the spatial distance mask of the unmanned aerial vehicle cluster;

secondly, using the graph annotation force network to process the graph structure data obtained in the first step to obtain deep characteristic information of each unmanned aerial vehicle;

thirdly, deep characteristic information of the whole unmanned aerial vehicle cluster is obtained according to the deep characteristic information of each unmanned aerial vehicle obtained in the second step, the obtained deep characteristic information of the whole unmanned aerial vehicle cluster is input into a fully-connected network, and classification of unmanned aerial vehicle cluster intentions is achieved after supervised training and learning.

2. The method for identifying intention of unmanned aerial vehicle clusters in countermeasure environment according to claim 1, wherein:

in the first step, the physical characteristics include a spatial position, a course angle, a pitch angle, a speed, a radar opening state and a type of unmanned aerial vehicle of each unmanned aerial vehicle.

3. An intention recognition method for unmanned aerial vehicle-oriented clusters in a countering environment according to claim 1 or 2, characterized in that:

in the first step, a blue machine cluster B is arranged to share M unmanned aerial vehicles B _i I=1, 2 … M, red cluster R has N unmanned aerial vehicles R in total _j J=1, 2 … N unmanned plane R _j To unmanned plane B _i Threat coefficients of (a) include distance threat coefficientsAngle threat coefficientSpeed threat coefficient->And air combat efficiency threat coefficient thR _ji R _ij The formula is as follows:

in the formula ,thR_ji D _ji ∈[0,1],D _ji Is unmanned plane R _j With unmanned plane B _i A distance therebetween; lambda (lambda) ₁ 、λ ₂ Is a constant parameter; />Is unmanned plane R _j To unmanned plane B _i Is a direction angle; k (K) _aR The effective acting distance of the missile is the effective acting distance of the missile; t (T) _rR The maximum acting distance of the missile is the maximum acting distance of the missile; />Is unmanned plane B _i Is a speed of (2);is unmanned plane R _j Is a speed of (2); c (C) _i Is unmanned plane B _i Air combat effectiveness index, C _j Is unmanned plane R _j Is an air combat effectiveness index.

4. A method for identifying intent of unmanned aerial vehicle-oriented clusters in a challenge environment according to claim 3, wherein:

unmanned aerial vehicle R with red square _j Unmanned with red squareMachine R _j′ The distance mask between (j' =1, 2 … N) is DI _jj′ The formula is as follows:

DI _jj′ ＝INT(d(R _j′ ,R _j )≤μ)

wherein mu is a constant and mu is [0.6,0.75]]The method comprises the steps of carrying out a first treatment on the surface of the INT represents the truth transformation function; d (R) _j′ ,R _j ) Representing unmanned plane R _j Red square unmanned plane R _j′ Is a relative distance of (2);

5. the method for identifying the intention of the unmanned aerial vehicle cluster in the countermeasure environment according to claim 4, wherein:

according to the distance threat coefficient, the angle threat coefficient, the speed threat coefficient, the air combat effectiveness threat coefficient and the distance mask, the graph structure data of the constructed unmanned aerial vehicle cluster are as follows:

wherein j "=1, 2 … N, th _jj′ In order to utilize the distance threat coefficient, the angle threat coefficient, the speed threat coefficient and the air combat effectiveness threat coefficient, the unmanned plane R is calculated by a gray correlation matrix method _j And unmanned plane R _j′ Comprehensive threat coefficients to the blue party;

G _jj′ ∈[0,1]

selecting the G _jj′ Is set to G _jz Maximum value, unmanned plane R _j With unmanned aerial vehicle R _z And connecting the unmanned aerial vehicle clusters to complete the construction of the graph data structure of the unmanned aerial vehicle clusters.

6. The method for identifying the intention of the unmanned aerial vehicle cluster in the countermeasure environment according to claim 5, wherein the method comprises the following steps:

in the second step, the method for obtaining the deep characteristic information comprises the following steps: computing unmanned aerial vehicle R _j And unmanned plane R _j′ Attention coefficients between the deep feature information and the deep feature information are obtained according to the attention coefficients obtained through calculation;

wherein the attention coefficient alpha _jj′ The calculation method of (1) is as follows: connected unmanned aerial vehicle R using a leachable shared parameter W pair _j And unmanned plane R _j′ Performing feature enhancement, then splicing the enhanced features, and performing unmanned plane R _j And unmanned plane R _j′ And finally mapping the synthesized vector to a real number by using mapping alpha, and normalizing by softMax to obtain the connected unmanned aerial vehicle R _j And unmanned plane R _j′ Is given by:

wherein ,is unmanned plane R _j A set of other unmanned aerial vehicles in a connected red cluster,/->e _jj′ Mapping representing splice features;

e _jj′ ＝α([Wh _j ||Wh _j′ ])

h _j is unmanned plane R _j Is a collection of physical features of (a);

unmanned plane R _j Deep characteristic information of (a)The method comprises the following steps:

7. the method for identifying the intention of the unmanned aerial vehicle cluster in the countermeasure environment according to claim 6, wherein:

in the third step, deep characteristic information of the whole unmanned aerial vehicle cluster is obtained through image pooling operation according to the deep characteristic information of each unmanned aerial vehicle, specifically: let A ^(l) For the adjacency matrix corresponding to the first layer of the pooling layer, X ^(l) For the feature matrix of all nodes in the first layer, Z ^(l) For the features on each node extracted on the layer i graph, then the graph structure of the next layer is derived from the following equation:

A ^(l+1) ,X ^(l+1) ＝DiffPool(A ^(l) ,Z ^(l) )

wherein DiffPool represents solving the weight matrix S through a graph neural network ^(l) ，S ^(l) The numerical expression is used for indicating how much weight is distributed to the nodes of the next layer by the nodes of the previous layer, and the solving formula is as follows:

S ^(l) ＝softmax(GNN _l,pool (A ^(l) ,X ^(l) ))

wherein ,GNN_l,pool The method is characterized in that a graph neural network used for graph pooling is shown, and an iterative formula among layers in the graph neural network is as follows:

8. the method for identifying intent of unmanned aerial vehicle clusters in a challenge environment according to claim 7, wherein:

the optimization constraint function of the graph neural network comprises a connection regular term constraint and an entropy function regular term constraint, wherein the expression formula of the connection regular term constraint is as follows:

L _LP ＝||A ^(l) ,S ^(l) S ^(l)T || _F

the expression formula of the constraint of the regular term of the entropy function is as follows:

where n represents the number of nodes of the current pooling layer, p=1, 2 … n.

9. The method for identifying intention of unmanned aerial vehicle clusters in countermeasure environment according to claim 1, wherein:

the unmanned aerial vehicle cluster comprises a multi-rotor-wing decoy unmanned aerial vehicle, a investigation integrated unmanned aerial vehicle, a multi-rotor-wing relay unmanned aerial vehicle and a small investigation unmanned aerial vehicle.

10. The method for identifying intention of unmanned aerial vehicle clusters in countermeasure environment according to claim 1, wherein:

in the third step, classification of unmanned aerial vehicle cluster intent includes attack, impersonation, withdrawal, electronic interference and low-altitude attack.