Sensor-weapon-target distribution method and device based on graph structure
By modeling the combat scenarios into graph structures and using graph learning models for feature learning, the problems of low computational efficiency, poor adaptability and insufficient accuracy in complex combat environments are solved, and efficient and accurate target allocation is achieved.
Patent Information
- Application Number
- CN202510103444.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-22
- Publication Date
- 2025-06-24
AI Technical Summary
Traditional target allocation methods have low computational efficiency, poor adaptability and insufficient accuracy in complex combat environments, making it difficult to quickly adapt to dynamically changing combat environments.
Using a graph isomorphic network method, the combat scenario is modeled as a graph structure, and the graph learning model is used to learn and reconstruct the features of nodes and edges, thereby realizing the target allocation of sensors and weapons.
It improves the efficiency and accuracy of target allocation, enhances adaptability and generalization capabilities in different combat scenarios, and can quickly respond and optimize target allocation decisions.
Smart Images
Figure CN120197465A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of artificial intelligence, deep learning, and military automation, and particularly to a method and device for allocating sensors, weapons, and targets based on a graph isomorphism network. Background Art
[0002] With the continuous improvement of the demand for intelligence and automation in modern warfare, target allocation, as an important link in combat decision-making, has become a research hotspot. Traditional target allocation methods mostly rely on rule-driven or heuristic algorithms, and often have the following limitations:
[0003] Low computational efficiency: Traditional methods require manual setting of rules and constraint conditions, with high computational complexity and inability to quickly adapt to the dynamically changing combat environment.
[0004] Poor adaptability: Traditional methods are often optimized for specific scenarios and are difficult to operate efficiently in different combat scenarios.
[0005] Insufficient accuracy: Manually set rules and models may not fully consider the complex relationships in the combat environment, resulting in insufficient accuracy of target allocation.
[0006] In recent years, graph neural network methods based on deep learning have gradually shown great potential in the analysis of graph data, especially having obvious advantages when dealing with complex dependency relationships between nodes. However, there is still little research on the application of existing graph neural networks in target allocation, and there is a lack of in-depth optimization for the specific needs of the combat environment. Summary of the Invention
[0007] In view of this, the present invention provides a method for allocating sensors, weapons, and targets based on a graph isomorphism network, which can, in a complex combat environment, model the combat scenario as a graph structure, and combine a graph learning model to learn and reconstruct the features of nodes and edges, so as to quickly and accurately achieve the target allocation of sensors and weapons, and solve the problems of poor adaptability and low computational efficiency of traditional methods in complex combat environments.
[0008] To solve the above technical problems, the present invention is implemented as follows.
[0009] A sensor-weapon-target allocation method based on a graph structure, comprising:
[0010] Step 1: Model the combat scenario as a graph structure G1; model the sensors, weapons, and targets in the combat scenario as nodes of the graph structure; model the interaction relationships between the sensors, weapons, and targets as edges between the nodes; and represent the edge features as the importance of the interaction relationships.
[0011] Step 2: Use the graph learning model to reconstruct the features of nodes and edges in the graph structure G1 to capture the complex relationships between nodes and obtain the reconstructed graph G2;
[0012] Step 3: Fuse the graph structure G1 and the reconstructed graph G2 by weighting the importance of the same edges to obtain the fused graph G3;
[0013] Step 4: For each target node in the fused graph G3, with the idea that the better the allocation plan is when the sum of the importance on both sides of the node is higher, perform target allocation and assign weapons and sensors to the targets.
[0014] Preferably, the modeling of the interaction relationships between sensors, weapons, and targets as edges between nodes is as follows:
[0015] The edge features between sensors and targets include the detection probability of sensors for targets;
[0016] The edge features between weapons and targets include the damage probability of weapons for targets;
[0017] The edge features between weapons and sensors include the matching relationship between weapons and sensors.
[0018] Preferably, the graph learning model includes a graph neural network and a multi-layer perceptron; the graph neural network is used to reconstruct the node features; the multi-layer perceptron is used to reconstruct the edge features of each edge in the graph structure after feature reconstruction and output the importance of each edge.
[0019] Preferably, the graph neural network adopts a graph isomorphism network.
[0020] Preferably, the method further includes training the graph learning model, including:
[0021] Construct a graph structure G for the combat scenario as a sample 10 , and at the same time determine the optimal allocation plan and construct the graph structure G0 of the optimal allocation plan;
[0022] Use the graph learning model to reconstruct the features of nodes and edges in the graph structure G 10 to obtain the reconstructed graph G 20 ;
[0023] Calculate the loss function between the reconstructed graph G 20 and the graph structure G0 of the optimal allocation plan, and backpropagate to optimize the model parameters of the graph learning model.
[0024] Preferably, the loss function is set as:
[0025]
[0026] where, e i,jis the importance of the edge between node i and node j in the optimized allocation scheme graph structure G0, e' i,j is the reconstructed graph G 20 the importance of the edge between node i and node j in it;
[0027] The model parameters of the backpropagation optimization graph learning model are obtained, and the optimal hyperparameters of the graph learning model are obtained through cross-validation.
[0028] Preferably, the step 4 includes:
[0029] Sort the target nodes according to the threat value of the target from large to small. Starting from the target node with a large threat value, perform scheme allocation for each target node;
[0030] When performing scheme allocation for a target node, according to the importance of the edges in the fusion graph G3, select the sensors and weapons that have an interaction relationship with the current target node to generate all possible feasible chains composed of sensor-target-weapon; calculate the evaluation value for each feasible chain, and the evaluation value is obtained by adding the importance of the two edges on the feasible chain; select the feasible chain with the highest evaluation value and add it to the allocation scheme, and at the same time mark the allocated nodes to avoid repeated allocation.
[0031] The present invention also provides a sensor-weapon-target allocation device based on a graph structure. The device includes a graph structure representation module, a graph learning model, a fusion module, an allocation module, and a training module;
[0032] The graph structure representation module is used to model the combat scenario as a graph structure G1 and provide it to the graph learning model and the fusion module; the sensors, weapons, and targets in the combat scenario modeling are modeled as nodes of the graph structure; the interaction relationships between the sensors, weapons, and targets are modeled as edges between the nodes; the edge features represent the importance of the interaction relationships; among them, the edge features between the sensor and the target include the detection probability of the sensor for the target; the edge features between the weapon and the target include the damage probability of the weapon for the target; the edge features between the weapon and the sensor include the matching relationship between the weapon and the sensor;
[0033] The graph learning model is used to reconstruct the features of the nodes and edges of the graph structure G1 to capture the complex relationships between the nodes, obtain the reconstructed graph G2, and output it to the fusion module;
[0034] The fusion module is used to fuse the graph structure G1 and the reconstructed graph G2 by weighting the importance of the same edges to obtain the fusion graph G3;
[0035] The allocation module is used to perform target allocation for each target node in the fusion graph G3, with the idea that the higher the sum of the importance of the two sides of the node, the better the allocation scheme, and allocate the weapons and sensors to the targets;
[0036] The training module is used to train the learning module: input the combat scenario as a sample into the graph structure representation module to obtain the combat scenario construction graph structure G 10 , and use the graph learning model to 10 perform feature reconstruction on the nodes and edges of the graph structure G to obtain the reconstructed graph G 20 ; meanwhile, for the combat scenario as the sample, determine the optimal allocation plan and construct the graph structure G0 of the optimal allocation plan; use the reconstructed graph G 20 and the loss function between the graph structure G0 of the optimal allocation plan to perform backpropagation to optimize the model parameters of the graph learning model.
[0037] Preferably, the graph learning model includes a graph isomorphism network and a multi-layer perceptron; the graph isomorphism network model is used to reconstruct node features; the multi-layer perceptron is used to reconstruct the edge features of each edge in the graph structure after feature reconstruction and output the importance of each edge.
[0038] Preferably, the allocation module includes a sorting unit, a feasible chain generation unit, an evaluation unit, and a plan allocation unit;
[0039] The sorting unit is used to sort the target nodes from largest to smallest according to the threat value of the target, and starting from the target node with the largest threat value, perform plan allocation for each target node using the feasible chain generation unit, the evaluation unit, and the plan allocation unit;
[0040] The feasible chain generation unit is used to, when performing plan allocation for a target node, select the sensors and weapons having an interaction relationship with the current target node according to the importance of the edges in the fusion graph G3, generate all possible feasible chains composed of sensor-target-weapon, and form a feasible chain set;
[0041] The evaluation unit is used to calculate an evaluation value for each feasible chain, and the evaluation value is obtained by adding the importance of two edges on the feasible chain;
[0042] The plan allocation unit is used to select the feasible chain with the highest evaluation value according to the evaluation value calculated by the evaluation unit, add it to the allocation plan, and at the same time mark the allocated nodes to avoid repeated allocation.
[0043] Beneficial effects:
[0044] (1) Target allocation efficiency: The present invention models the combat scenario through a graph structure and learns the graph learning network based on the known scenario, can effectively capture the complex dependence relationship between sensors, weapons, and targets, thereby greatly improving the efficiency of target allocation.
[0045] (2) Adaptability and generalization ability: The present invention combines two stages of offline training and online inference. Through model training and cross-validation in the offline stage, the adaptability and generalization ability of the model in different combat scenarios are ensured. Whether it is a static scenario or a dynamically changing combat environment, the model can quickly adapt and perform target allocation through the online inference stage, demonstrating strong scenario adaptability.
[0046] (3) Target allocation accuracy: In a preferred implementation, through the precise reconstruction and learning of the features of nodes (sensors, weapons, targets) by the graph neural network, combined with the extraction and quantization of edge features by the multi-layer perceptron, the present invention can accurately evaluate the threat level and task importance of each target, optimize the target allocation scheme, and thus improve the accuracy of target allocation. In a multi-target complex environment, it can ensure that each target obtains the most suitable sensor and weapon combination, avoiding the uneven or inefficient allocation in traditional methods.
[0047] (4) Quick response ability: The present invention uses the greedy algorithm for target allocation decision-making. On the premise of ensuring the decision-making quality, it can quickly complete target allocation. The greedy algorithm gives priority to allocating the most urgent and critical targets according to the threat level and comprehensive importance value of the targets, thus ensuring the efficient execution of combat tasks. This method can make real-time responses in a dynamically changing combat environment, optimizing the speed and effect of target allocation decisions. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 is a flowchart of the sensor-weapon-target allocation method based on the graph structure according to the embodiment of the present invention;
[0049] Figure 2 is an explanatory diagram of the graph data structure representation of the combat scenario;
[0050] Figure 3 is a flowchart of the offline training stage;
[0051] Figure 4 is a curve graph of the change of the loss function during the offline training process;
[0052] Figure 5 is a flowchart of the online inference stage;
[0053] Figure 6 is a flowchart of the greedy algorithm target allocation;
[0054] Figure 7 is a graph of the comparison results of algorithm performance;
[0055] Figure 8 is a block diagram of the composition of the sensor-weapon-target allocation device based on the graph structure according to the embodiment of the present invention;
[0056] Figure 9 For Figure 8 the block diagram of the composition of the allocation module in Specific implementation manner
[0057] The present invention will be described in detail below in conjunction with the accompanying drawings and by way of examples.
[0058] The present invention provides a sensor - weapon - target allocation scheme based on a graph structure. Its basic idea is that for a complex scenario such as a combat environment that includes multiple sensors, weapons, and targets, it is modeled as a graph structure, and a graph learning model is used to optimize the importance of the edges of the graph structure so that it can capture the complex relationships between nodes; then, based on the optimized graph structure and the original graph structure are fused, and then based on the importance of the target edges, sensors and weapons are allocated to the targets, so that in a multi - target complex environment, it can ensure that each target obtains the most suitable combination of sensors and weapons, avoiding the uneven or inefficient allocation in traditional methods, and achieving efficient and accurate combat decision - making.
[0059] Figure 1 The sensor - weapon - target allocation method based on the graph structure of the present invention is shown, which includes the following steps:
[0060] Step 1: Initialize the combat scenario.
[0061] The information for initialization mainly includes the number information of sensors, weapons, and targets; the threat value information of each target; the interaction relationships between sensors, weapons, and targets. Among them, the interaction relationships at least include the damage probability information of each weapon against each target, the detection probability information of each sensor against each target, and the matching information of each weapon against each sensor. Here, the matching information refers to whether the weapon and the sensor can cooperate with each other. In practice, other information can also be added to enrich and improve the interaction relationships. For example, for the interaction relationship between the sensor and the target, the capabilities of the sensor itself, such as sensitivity, communication ability, etc., can also be added.
[0062] Step 2: Model the combat scenario as a graph structure G1.
[0063] According to the information after initialization, the combat environment is modeled as a graph structure. Among them, each sensor, weapon, and target in the combat scenario is regarded as a node in the graph, and the edges between the nodes represent their relationships, such as the detection probability, damage probability, and matching relationship mentioned above. The edge features represent the importance of the interaction relationships.
[0064] In this embodiment, the interaction relationships among sensors, weapons, and targets are modeled as edges between nodes as follows: The edge features between a sensor and a target include the detection probability of the sensor for the target; the edge features between a weapon and a target include the damage probability of the weapon for the target; and the edge features between a weapon and a sensor include the matching relationship between the weapon and the sensor.
[0065] Based on the information of each node and edge, the initial features can be constructed as shown in the following table:
[0066]
[0067] Among them, h t represents the threat value of target t, p st represents the probability that sensor s detects target t, q wt represents the damage probability of weapon w for target t, and m sw represents whether sensor s and weapon w can be matched, which is 1 if they can, and 0 if they cannot. The illustration of the graph data structure of the combat scenario is as Figure 2 shown. The features of all nodes and edges are initialized and ready for graph neural network processing.
[0068] Step 3: Use the graph learning model to perform model inference on the constructed scenario graph structure.
[0069] The model inference includes two processes: node feature reconstruction and edge feature importance reconstruction (calculation).
[0070] The graph learning model of this embodiment includes a graph neural network and a multi-layer perceptron. The graph neural network preferably adopts a graph isomorphism network. The graph isomorphism network is used to realize the reconstruction of node features; the multi-layer perceptron is used to realize the edge feature reconstruction of each edge in the graph structure after feature reconstruction, that is, importance calculation.
[0071] The graph learning model for inference needs to be pre-trained offline to ensure that efficient target allocation decisions can be made in real time during the online inference phase, and then the offline-trained model is used for online inference.
[0072] The offline training process is as Figure 3 shown, specifically:
[0073] Step A1: Obtain a large number of combat scenario data as samples. For each combat scenario sample, construct a graph structure, denoted as G 10 . Use an exact algorithm to determine the optimal allocation plan for the combat scenario sample, and construct an optimized allocation plan graph structure G0. Among them, the exact algorithm can adopt the branch and bound method, or the exhaustive method combined with evaluation value calculation to find the optimal allocation plan.
[0074] Step A2: Use the graph learning model to process the graph structure G 10Reconstruct the features of nodes and edges to obtain the reconstructed graph G 20 。
[0075] Step A3: Calculate the loss function between the reconstructed graph G 20 and the graph structure G0 of the optimized allocation scheme, and backpropagate to optimize the model parameters of the graph learning model.
[0076] Among them, the loss function during training can be set as:
[0077]
[0078] Among them, e i,j is the importance of the edge between node i and node j in the graph structure G0 of the optimized allocation scheme, and e' i,j is the importance of the edge between node i and node j in the reconstructed graph G 20 。
[0079] Among them, when backpropagating to optimize the model parameters of the graph learning model, the optimal hyperparameters of the model can be obtained by adopting a cross-validation scheme according to the L calculated in multiple rounds.
[0080] The change curve of the loss function during training is as Figure 4 shown. The curve is smooth and stable, and no obvious overfitting phenomenon occurs, indicating that the model has good generalization ability and stability and can achieve good results in different scenarios.
[0081] The online inference process of the graph learning model is as Figure 5 shown. In this embodiment, first, the graph isomorphism network is used to reconstruct the node features. The features of each node are fused with the information of its neighbor nodes through the neighborhood aggregation operation. The node features updated by the graph isomorphism network are:
[0082]
[0083] Among them, represents the feature vector of node v at the kth layer; N(v) represents the set of neighbor nodes of node v; MLP (k) is the multi-layer perceptron at the kth layer in the graph neural network; ε (k) is a learnable parameter.
[0084] The graph isomorphism network outputs the graph structure after reconstructing the node features. The multi-layer perceptron obtains the information of the two nodes connected to each edge for each edge, and obtains node-edge-node (target node - damage probability edge - weapon node, target node - detection probability edge - sensor node), concatenates their feature vectors, and then uses the multi-layer perceptron to learn and quantify the features of the edge. The features of each edge are updated by the following formula:
[0085]
[0086] Among them, represents the feature of the edge between node i and node j at the t-th layer. The output layer of the last layer of the multi-layer perceptron contains only one node output, and the sigmoid activation function is used to normalize the output to between [0, 1], which is used to represent the importance of the current edge in the entire allocation.
[0087] Use the trained graph isomorphism network to update and reconstruct the features of sensors, weapons, and target nodes, capture the complex dependencies between them, so as to accurately reflect the interactions between sensors, weapons, and targets. Use the trained multi-layer perceptron to extract edge features and evaluate the importance of each edge for subsequent target allocation decisions.
[0088] Step 4: By the way of weighting the importance of the same edges, fuse the graph structure G1 and the reconstructed graph G2 to obtain the fused graph G3.
[0089] As Figure 6 shown, after the previous steps, the graph structure G1 and the reconstructed graph structure G2 are obtained. In this step, graph fusion is entered. The importance of the same edge in G1 and the importance in G1 are weighted and summed to comprehensively reflect the importance of the edge and the initial probability information, and then put back into the graph structure to obtain the fused graph G3.
[0090] Step 5: For each target node in the fused graph G3, with the idea that the higher the sum of the importance on both sides of the node, the better the allocation scheme, perform target allocation and allocate weapons and sensors to the targets.
[0091] In this embodiment, in order to further improve the efficiency of target allocation, this step can use the greedy algorithm for quick decision-making and give priority to allocating the most threatening targets. Specifically:
[0092] Sort the target nodes according to the threat values of the targets from large to small, and give priority to processing the targets with higher threat values. When allocating a scheme for each target node, according to the importance of the edges in the fused graph G3, as long as the importance is greater than 0, it is considered that there is an interaction relationship between the nodes at both ends of the edge. Select the sensors and weapons that have an interaction relationship with the current target node to generate all possible feasible chains composed of sensor-target-weapon; calculate the evaluation value for each feasible chain, and the evaluation value is obtained by adding the importance of the two edges on the feasible chain; select the feasible chain with the highest evaluation value and add it to the allocation scheme, and at the same time mark the allocated nodes, or directly delete the allocated nodes to avoid repeated allocation. Repeat this process until all targets are allocated to ensure the overall optimality and efficiency of the scheme.
[0093] Step 6: Finally, output the allocation results of weapons and sensors for each target to ensure the efficient execution of combat missions.
[0094] Under the example cases of different combat scenarios, the method of the present invention and other algorithms (PSO and GA) were run multiple times, and the average performance of each algorithm was compared. The comparison results are as Figure 7 shown. The results show that the method of the present invention is slightly lower than GA in the evaluation value of small-scale scenarios, but is superior to PSO and GA in medium-scale and large-scale scenarios. Moreover, the method of the present invention shows significant advantages both in terms of evaluation value and efficiency, especially showing better stability and robustness in complex scenarios.
[0095] Based on the above allocation method, the present invention also provides a sensor-weapon-target allocation device based on a graph structure, as Figure 8 shown. The device includes a graph structure representation module, a graph learning model, a fusion module, an allocation module, and a training module;
[0096] The graph structure representation module is used to model the combat scenario as a graph structure G1 and provide it to the graph learning model and the fusion module; in the modeling of the combat scenario, sensors, weapons, and targets are modeled as nodes of the graph structure; the interaction relationships between sensors, weapons, and targets are modeled as edges between nodes; the edge features represent the importance of the interaction relationships; among them, the edge features between sensors and targets include the detection probability of the sensor for the target; the edge features between weapons and targets include the damage probability of the weapon for the target; the edge features between weapons and sensors include the matching relationship between the weapon and the sensor;
[0097] The graph learning model is used to reconstruct the features of nodes and edges of the graph structure G1 to capture the complex relationships between nodes, obtain the reconstructed graph G2, and output it to the fusion module;
[0098] The fusion module is used to fuse the graph structure G1 and the reconstructed graph G2 by weighting the importance of the same edges to obtain the fused graph G3;
[0099] The allocation module is used to perform target allocation for each target node in the fused graph G3. Taking the idea that the higher the sum of the importance on both sides of the node, the better the allocation scheme, the weapons and sensors are allocated to the targets;
[0100] The training module is used to train the learning module: input the combat scenario as a sample into the graph structure representation module to obtain the graph structure G of the combat scenario construction 10 , and use the graph learning model to reconstruct the features of nodes and edges of the graph structure G 10 to obtain the reconstructed graph G 20 ; at the same time, for the combat scenario as the sample, determine the optimal allocation scheme and construct the graph structure G0 of the optimal allocation scheme; use the reconstructed graph G20 Backpropagate the loss function between the optimized allocation scheme graph structure G0 to optimize the model parameters of the graph learning model.
[0101] Preferably, the graph learning model includes a graph isomorphism network and a multi-layer perceptron; the graph isomorphism network model is used to reconstruct node features; the multi-layer perceptron is used to reconstruct the edge features of each edge in the graph structure after feature reconstruction, and output the importance of each edge.
[0102] Figure 9 The composition block diagram of the allocation module is shown. As shown in the figure, the allocation module includes a sorting unit, a feasible chain generation unit, an evaluation unit, and a scheme allocation unit;
[0103] The sorting unit is used to sort the target nodes in descending order according to the threat value of the target, and starting from the target node with the largest threat value, perform scheme allocation for each target node using the feasible chain generation unit, the evaluation unit, and the scheme allocation unit;
[0104] The feasible chain generation unit is used to, when performing scheme allocation for a target node, select sensors and weapons that have an interaction relationship with the current target node according to the importance of the edges in the fusion graph G3, generate all possible feasible chains composed of sensor-target-weapon, and form a feasible chain set;
[0105] The evaluation unit is used to calculate an evaluation value for each feasible chain, and the evaluation value is obtained by adding the importance of two edges on the feasible chain;
[0106] The scheme allocation unit is used to select the feasible chain with the highest evaluation value according to the evaluation value calculated by the evaluation unit, add it to the allocation scheme, and at the same time mark the allocated nodes to avoid repeated allocation.
[0107] In summary, the present invention proposes a sensor-weapon-target allocation scheme based on a graph structure, which solves the problems of poor adaptability and low computational efficiency of traditional methods in complex combat environments. This method first models the combat scenario as a graph structure, learns and reconstructs the features of nodes and edges through a graph learning model, captures the complex dependency relationships between nodes, so as to accurately reflect the interactions between sensors, weapons, and targets. Subsequently, combining the two stages of offline training and online inference, the optimal allocation result is learned offline to ensure that efficient target allocation decisions can be made in real time during the online inference stage.
[0108] In order to further improve the efficiency of target allocation, a greedy algorithm is used for rapid decision-making, and the most threatening targets are preferentially allocated. This method comprehensively considers the deep learning ability of the graph neural network and the efficiency of the greedy algorithm, realizes rapid response and accurate allocation in a dynamic and complex environment, and has significant innovation and broad application prospects.
[0109] In summary, the above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. A sensor-weapon-target allocation method based on a graph structure, characterized in that: include: Step 1: Model the combat scenario as a graph structure G1; The sensors, weapons and targets in combat scenario modeling are modeled as nodes of a graph structure; the interaction relationships among sensors, weapons and targets are modeled as edges between nodes; Edge features represent the importance of the relationship; Step 2: Use the graph learning model to reconstruct the features of nodes and edges of the graph structure G1 to capture the complex relationship between nodes and obtain the reconstructed graph G2; Step 3: By weighting the importance of the same edges, the graph structure G1 and the reconstructed graph G2 are fused to obtain the fused graph G3; Step 4: For each target node in the fusion graph G3, the allocation strategy is to allocate weapons and sensors to the target based on the idea that the higher the sum of the importance of both sides of the node, the better the allocation scheme.
2. The method according to claim 1, characterized in that The interaction relationship between sensors, weapons and targets is modeled as the edges between nodes: The edge features between the sensor and the target include the detection probability of the sensor to the target; The edge features between the weapon and the target include the probability of the weapon damaging the target; The edge features between weapons and sensors include the matching relationship between weapons and sensors.
3. The method according to claim 1, characterized in that The graph learning model includes a graph neural network and a multi-layer perceptron; the graph neural network is used to reconstruct node features; the multi-layer perceptron is used to reconstruct the edge features of each edge in the graph structure after feature reconstruction, and output the importance of each edge.
4. The method according to claim 3, characterized in that The graph neural network adopts a graph isomorphism network.
5. The method according to claim 1, characterized in that The method further includes training the graph learning model, including: Construct a graph structure G for the combat scenario as a sample 10 , and at the same time determine the optimal allocation plan and construct the optimal allocation plan graph structure G0; Using graph learning model to learn graph structure G 10 Reconstruct the features of nodes and edges to obtain the reconstructed graph G 20 ; Compute the reconstructed graph G 20 The loss function between the graph structure G0 and the optimized allocation scheme is then back-propagated to optimize the model parameters of the graph learning model.
6. The method according to claim 5, characterized in that The loss function is set as: Among them, e i,j is the importance of the edge between node i and node j in the optimized allocation scheme graph structure G0, e' i,j is the reconstructed graph G 20 The importance of the edge between node i and node j; The back propagation optimizes the model parameters of the graph learning model, and obtains the optimal hyperparameters of the graph learning model through cross-validation.
7. The method according to claim 1, characterized in that The step 4 comprises: Sort the target nodes according to their threat values from large to small, starting from the target nodes with large threat values, and assign solutions to each target node; When allocating a plan for a target node, sensors and weapons that have an action relationship with the current target node are selected according to the importance of the edges in the fusion graph G3 to generate all possible feasible chains consisting of sensors, targets, and weapons; an evaluation value is calculated for each feasible chain, and the evaluation value is obtained by adding the importance of the two edges on the feasible chain; the feasible chain with the highest evaluation value is selected and added to the allocation plan, and the allocated nodes are marked to avoid duplicate allocation.
8. A sensor-weapon-target allocation device based on a graph structure, characterized in that: The device includes a graph structure representation module, a graph learning model, a fusion module, an allocation module and a training module; The graph structure representation module is used to model the combat scenario as a graph structure G1 and provide it to the graph learning model and fusion module; The sensors, weapons and targets in combat scenario modeling are modeled as nodes of a graph structure; the interaction relationships among sensors, weapons and targets are modeled as edges between nodes; The edge feature indicates the importance of the action relationship; the edge feature between the sensor and the target includes the detection probability of the sensor to the target; the edge feature between the weapon and the target includes the damage probability of the weapon to the target; the edge feature between the weapon and the sensor includes the matching relationship between the weapon and the sensor; The graph learning model is used to reconstruct the features of nodes and edges of the graph structure G1 to capture the complex relationship between nodes, obtain the reconstructed graph G2, and output it to the fusion module; The fusion module is used to fuse the graph structure G1 and the reconstructed graph G2 by weighting the importance of the same edges to obtain a fused graph G3; The allocation module is used to allocate the target to each target node in the fusion graph G3, taking the allocation scheme with the sum of importance on both sides of the node as the best allocation idea, and allocate the weapons and sensors to the target; The training module is used to train the learning module: input the combat scene as a sample into the graph structure representation module to obtain the combat scene construction graph structure G 10 , using the graph learning model to 10 Reconstruct the features of nodes and edges to obtain the reconstructed graph G 20 At the same time, for the combat scenario as a sample, determine the optimal allocation plan and construct the optimal allocation plan graph structure G0; use the reconstruction graph G 20 The loss function between the optimized allocation scheme graph structure G0 is back-propagated to optimize the model parameters of the graph learning model.
9. The device according to claim 8, characterized in that The graph learning model includes a graph isomorphism network and a multi-layer perceptron; the graph isomorphism network model is used to reconstruct node features; the multi-layer perceptron is used to reconstruct the edge features of each edge in the graph structure after feature reconstruction, and output the importance of each edge.
10. The device according to claim 8 or 9, characterized in that The allocation module includes a sorting unit, a feasible chain generation unit, an evaluation unit and a solution allocation unit; A sorting unit is used to sort the target nodes according to the threat value of the target from large to small, starting from the target node with the largest threat value, and for each target node, a feasible chain generation unit, an evaluation unit, and a solution allocation unit are used to allocate a solution; The feasible chain generation unit is used to select sensors and weapons that have an action relationship with the current target node according to the importance of the edges in the fusion graph G3 when allocating a plan for a target node, and generate all possible feasible chains consisting of sensors, targets and weapons to form a feasible chain set; An evaluation unit, used to calculate an evaluation value for each feasible chain, where the evaluation value is obtained by adding the importance of two edges on the feasible chain; The scheme allocation unit is used to select the feasible chain with the highest evaluation value according to the evaluation value calculated by the evaluation unit and add it to the allocation scheme, and mark the allocated nodes to avoid repeated allocation.