Unmanned Formation Intelligent Air Combat Decision-making Method and System Based on Optimized Graph Neural Network

Abstract

The present invention discloses an intelligent air combat decision-making method and system for unmanned formations based on optimized graph neural networks, which relates to the field of future unmanned formation intelligent air combat technologies, and includes: obtaining simulated air combat confrontation data of unmanned formations participating in intelligent air combat confrontation; after performing dimensionality transformation on the simulated air combat confrontation data, inputting it into a temporal convolutional neural network to obtain a temporal feature vector; inputting the temporal feature vector into a trained graph neural network to obtain predicted confrontation data of the unmanned formations; and determining a decision set of the unmanned formations participating in intelligent air combat confrontation based on the predicted confrontation data. The present invention provides a solution idea for the problem in the prior art that the time and space information of distributed unmanned formation intelligent air combat cannot be represented quickly and simultaneously, and further provides technical support for solving the problem of unmanned formation intelligent air combat confrontation strategies.

Description

Technical Field

[0001] The present invention relates to the technical field of future unmanned aerial vehicle formation intelligent air combat, and specifically provides an unmanned formation intelligent air combat decision-making method and system based on an optimized graph neural network. Background Art

[0002] With the development of intelligent technology, the autonomy of unmanned intelligent agents within a formation is becoming stronger and stronger, and their autonomous decision-making ability brings machine decision-making advantages. Related domestic research on unmanned formation air combat decision-making usually simplifies multi-aircraft confrontation to a single game body problem. The default prerequisite for this method is that the states of all game intelligent agents are known, requiring real-time communication and a huge state space. Foreign research on unmanned formation tactical decision-making mainly focuses on collaborative confrontation levels such as territorial defense problems and pursuit and escape problems. Considering the incomplete interaction information between unmanned aerial vehicles in the actual complex combat environment, it cannot meet the requirements of unmanned formation intelligent air combat confrontation strategies. Summary of the Invention

[0003] The purpose of the present invention is to provide an unmanned formation intelligent air combat decision-making method and system based on an optimized graph neural network to solve at least one of the above technical problems.

[0004] In a first aspect, an embodiment of the present invention provides an unmanned formation intelligent air combat decision-making method based on an optimized graph neural network, including: obtaining simulated air combat confrontation data of an unmanned formation participating in intelligent air combat confrontation; after performing dimensionality change on the simulated air combat confrontation data, inputting it into a temporal convolutional neural network to obtain a temporal feature vector; inputting the temporal feature vector into a trained graph neural network to obtain the predicted confrontation data of the unmanned formation; and determining a decision set of the unmanned formation participating in intelligent air combat confrontation based on the predicted confrontation data.

[0005] Further, the calculation formula of the temporal convolutional neural network includes:

[0006]

[0007] where X TCN represents the temporal feature vector, represents the value after linear variable expansion of the input variable, f(·) represents an activation function, d represents a dilation factor, k represents the convolution kernel size, and s - d·i represents the past direction.

[0008] Further, the method further includes: constructing a preset graph neural network based on the initial confrontation data of the unmanned formation participating in intelligent air combat confrontation; and training the preset graph neural network based on a preset training set to obtain the trained graph neural network.

[0009] Further, the initial adversarial data includes: the longitude, latitude, yaw angle, pitch angle, roll angle, and speed of each unmanned aerial vehicle in the unmanned formation participating in the intelligent air combat confrontation.

[0010] Further, the mathematical calculation formula of the preset graph neural network includes:

[0011]

[0012] where, W (k) , B (k) represent the weight parameters to be trained for the k-th layer of the graph neural network, J represents the number of nodes, X TCN represents the temporal feature vector, represents the prediction result of the j-th node in the k-th layer of the graph neural network;

[0013] The training loss function of the preset graph neural network includes:

[0014]

[0015] where, N represents the total amount of data, P represents the prediction degree, K represents the number of layers of the preset graph neural network, represents the predicted value, y itjk represents the true value.

[0016] Further, based on the predicted adversarial data, determining the decision set of the unmanned formation participating in the intelligent air combat confrontation includes: based on the predicted adversarial data and the current adversarial data of the unmanned formation participating in the intelligent air combat confrontation, determining the change amount of the state quantity of each unmanned aerial vehicle in the unmanned formation participating in the intelligent air combat confrontation within a unit time, and taking the change amount as the decision set.

[0017] In a second aspect, an embodiment of the present invention further provides an intelligent air combat decision-making system for an unmanned formation based on an optimized graph neural network, including: an acquisition module, a temporal feature extraction module, a spatial feature extraction module, and a decision-making module; wherein, the acquisition module is used to acquire the simulated air combat confrontation data of the unmanned formation participating in the intelligent air combat confrontation; the temporal feature extraction module is used to input the simulated air combat confrontation data after dimension change into a temporal convolutional neural network to obtain a temporal feature vector; the spatial feature extraction module is used to input the temporal feature vector into a trained graph neural network to obtain the predicted adversarial data of the unmanned formation; the decision-making module is used to determine the decision set of the unmanned formation participating in the intelligent air combat confrontation based on the predicted adversarial data.

[0018] Further, it further includes a training module for: constructing a preset graph neural network based on the initial confrontation data of the unmanned formation participating in the intelligent air combat confrontation; training the preset graph neural network based on a preset training set to obtain the trained graph neural network.

[0019] In a third aspect, an embodiment of the present invention further provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the computer program, the method described in the first aspect above is implemented.

[0020] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, where the computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method described in the first aspect above is implemented.

[0021] The present invention provides a method and system for intelligent air combat decision-making of an unmanned formation based on an optimized graph neural network. By integrating a temporal convolutional network and a graph neural network into a learning model of the optimized graph neural network, it effectively represents the relative situation of multiple aircraft within the formation and the cumulative situation of a single aircraft within the formation. Through the training and learning of situation slices, the maneuvering actions of a single platform are obtained, and the maneuvering actions decided by the machine are connected into a situation chain to study the decision-making tactics of a single aircraft and even the decision-making tactics of the formation, providing a feasible idea and a beneficial attempt for the research of unmanned formation tactical decision-making, and alleviating the technical problem that the prior art cannot meet the requirements of the intelligent air combat confrontation strategy of the unmanned formation. Description of the Drawings

[0022] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the specific embodiments or the description of the prior art. Obviously, the following drawings are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a flowchart of a method for intelligent air combat decision-making of an unmanned formation based on an optimized graph neural network provided by an embodiment of the present invention;

[0024] Figure 2 It is a graph neural network representation graph and a computational graph provided by an embodiment of the present invention;

[0025] Figure 3 It is a schematic diagram of TCN convolution of two convolutional kernels provided by an embodiment of the present invention;

[0026] Figure 4 It is a schematic diagram of a two-layer decision-making model for unmanned formation tactics provided by an embodiment of the present invention;

[0027] Figure 5 Schematic diagram of an optimized graph neural network model training framework provided by an embodiment of the present invention;

[0028] Figure 6 Schematic diagram of a curve of change in training loss value provided by an embodiment of the present invention;

[0029] Figure 7 Schematic diagram of the situation when the red and blue sides first launch medium-range missiles;

[0030] Figure 8 Schematic diagram of the situation when the red and blue sides second launch medium-range missiles;

[0031] Figure 9 Schematic diagram of the final situation provided by an embodiment of the present invention;

[0032] Figure 10 Schematic diagram of an unmanned formation intelligent air combat decision-making system based on an optimized graph neural network provided by an embodiment of the present invention. Specific implementation manners

[0033] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0034] Embodiment 1

[0035] Figure 1 It is a flowchart of an unmanned formation intelligent air combat decision-making method based on an optimized graph neural network provided by an embodiment of the present invention. As Figure 1 shown, the method specifically includes the following steps:

[0036] Step S102, obtain the simulated air combat confrontation data of the unmanned formation participating in the intelligent air combat confrontation.

[0037] Step S104, after changing the dimensions of the simulated air combat confrontation data, input it into the temporal convolutional neural network to obtain the temporal feature vector.

[0038] Step S106, input the temporal feature vector into the trained graph neural network to obtain the predicted confrontation data of the unmanned formation.

[0039] Step S108, based on the predicted confrontation data, determine the decision set of the unmanned formation participating in the intelligent air combat confrontation.

[0040] Specifically, the calculation formula of the temporal convolutional neural network includes:

[0041]

[0042] Among them, X TCN represents the temporal feature vector, represents the value after linear variable expansion of the input variable, d represents the dilation factor, f(·) represents the activation function, k represents the convolutional kernel size, and s - d·i represents the past direction.

[0043] The method provided by the embodiments of the present invention further includes the training process of the graph neural network. Specifically, it includes:

[0044] Construct a preset graph neural network based on the initial confrontation data of the unmanned formation participating in the intelligent air combat confrontation;

[0045] Train the preset graph neural network based on the preset training set to obtain a trained graph neural network.

[0046] Preferably, the initial confrontation data includes: the longitude, latitude, yaw angle, pitch angle, roll angle, and speed of each unmanned aerial vehicle in the unmanned formation participating in the intelligent air combat confrontation.

[0047] Specifically, the mathematical calculation formula of the preset graph neural network includes:

[0048]

[0049] Among them, W (k) , B (k) represent the weight parameters to be trained for the k-th layer of the graph neural network, J represents the number of nodes, X TCN represents the temporal feature vector, represents the prediction result of the j-th node in the k-th layer of the graph neural network;

[0050] The training loss function of the preset graph neural network includes:

[0051]

[0052] Among them, N represents the total amount of data, P represents the prediction degree, K represents the number of layers of the preset graph neural network, represents the predicted value, and y itjk represents the true value.

[0053] Specifically, the basic Graph Neural Networks (GNNs) can capture the dependencies of a graph through message passing between graph nodes. The main idea is to iteratively aggregate the feature information from neighboring nodes and integrate the aggregated information with the current central node. In recent years, model architectures based on graph neural networks have shown breakthrough performance in tasks related to graph data. In this invention, the unmanned aerial vehicles are set as the nodes of the graph neural network model, and the node values are the feature vectors corresponding to the unmanned aerial vehicles. Other unmanned aerial vehicles within the detectable range or within a certain airspace range are used as related nodes to construct a graph structure. The node data at each moment is the feature vector corresponding to that moment, and the value of each edge represents a measure of the relative influence between the corresponding nodes. According to the actual situation, the Euclidean distance between two points can be taken, such as Figure 2 shown Figure 2 is a graph neural network representation graph and a computational graph provided according to an embodiment of the present invention.

[0054] The graph neural network can achieve spatial aggregation of the current situation. However, in intelligent air combat decision-making, attention should also be paid to grasping the historical situation. For example, a pilot may sacrifice the current situation to make dangerous moves to accumulate advantages for subsequent actions. Solving this problem is similar to most spatio-temporal sequence prediction problems. To solve spatio-temporal sequence prediction problems, spatial modules and time modules are often used. Among them, the commonly used spatial module model structures are Convolutional Neural Network (CNN), Graph Convolutional Network (GCN), and Attention mechanism network (Attention). The commonly used time module model structures are Long Short-Term Neural Network (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), and Attention mechanism network (Attention). The common aggregation methods of the two modules mainly include sequence structure, coupling structure, and spatio-temporal synchronous learning. A large number of experiments show that GNN+TCN is one of the frameworks with the highest accuracy at present, and there is no RNN and Attention in this model, and the calculation and training speed of the model are faster. Therefore, the present invention selects the temporal convolutional network to optimize the graph neural network.

[0055] Specifically, the Temporal Convolutional Network (TCN) can be used for processing temporal data and has been proven to achieve good results in extracting high-level features in structured data. It is suitable for the temporal representation of temporal data and can provide a field of view for temporal modeling. When training the TCN to predict l values under the input time series, assuming the input sequence is x0, x1, …, xL, it is desired to predict some corresponding outputs y0, y1, …, y L , whose value is equal to the input value shifted forward by l units. During the prediction process, only the observed inputs before time step t can be used to predict the output y at a certain time step t tTo meet the above principles, TCN uses a one-dimensional fully convolutional network. Each layer of convolution has the same length and has a zero-padding function, as shown in Figure 3 shown, where Figure 3 is a schematic diagram of the TCN convolution of two convolutional kernels provided according to an embodiment of the present invention.

[0056] Unmanned formation intelligent air combat decision-making is a formation-level distributed decision-making problem. This problem can be decoupled into formation tactical decision-making and single-aircraft tactical decision-making problems to reduce the difficulty of solving. Among them, single-aircraft tactical decision-making is the end implementation link. Figure 4 is a schematic diagram of a two-layer tactical decision-making model for unmanned formations provided according to an embodiment of the present invention. The basic construction idea of unmanned formation intelligent air combat decision-making is as shown in Figure 4 shown.

[0057] The essence of unmanned formation intelligent air combat confrontation decision-making is to obtain the to-be-predicted eigenvalue of the future time step through a mapping function by giving the eigenvalue of the current confrontation topology map of the two sides of the unmanned aircraft and the historical time step. The embodiment of the present invention proposes to input the processed unmanned formation air combat confrontation data into the TCN network for extracting temporal features and the GNN network for extracting spatial features in sequence. The training block diagram of its optimized graph neural network model is as shown in Figure 5 shown.

[0058] As shown in Figure 5 shown, a method for unmanned formation intelligent air combat decision-making based on an optimized graph neural network provided by an embodiment of the present invention. The main steps for solving the problem of unmanned formation intelligent air combat confrontation strategy are as follows:

[0059] First, extract the key influencing factors in the air combat confrontation data in combination with the actual problem. It mainly includes the aircraft types, aircraft positions, attitudes, types and quantities of mounted missiles, key actions (mainly considering intercepting the target 00, turning on interference 01, launching missiles 02), time identifiers (identifying the situation time tags), etc. The key information is described in a structured manner. The data structure of a single aircraft as a node of the graph neural network is shown in Table 1. Through comparative analysis, the present invention selects 8, 9, 11, 12, 13, and 14 as the eigenvalues of the input vector of the graph neural network model. That is, the node eigenvalues of the graph neural network in the embodiment of the present invention include the longitude, latitude, yaw angle, pitch angle, roll angle, and speed of each unmanned aircraft in the unmanned formation participating in the intelligent air combat confrontation.

[0060] Table 1 Single Aircraft Situation Description Table

[0061]

[0062]

[0063] Secondly, the historical cumulative situation of each aircraft in the air combat confrontation between both sides is the primary key issue in the decision-making problem. After changing the dimensions of the simulated air combat confrontation data, the TCN model is input to parallelize the processing of the historical cumulative situation and extract the time information in the data. The calculation formula is as shown in (1):

[0064]

[0065] Among them, X TCN represents the time series feature vector, represents the value after the input variable is linearly expanded, d represents the dilation factor, which takes the value of 1 in the embodiments of the present invention; k represents the convolutional kernel size, which takes the value of 3 in the embodiments of the present invention; s - d·i represents the past direction.

[0066] In addition, the TCN in the embodiments of the present invention uses a residual network. The calculation formula of the l-th residual module is as shown in (2):

[0067]

[0068] Among them, is the output result of the l-th residual module, is the output result of the previous residual module, φ(·) is the mapping function of the residual module; after passing through multiple residual modules, the output result X of the TCN model is obtained TCN ∈R N×O×8×E .

[0069] Then, dynamically capturing the spatial correlation between the complex confrontation sides and the friendly side is another key issue in the unmanned formation decision-making. The output result of the TCN model is input into the GNN to extract the spatial information of the confrontation data. The calculation formulas are as shown in (3) and (4):

[0070]

[0071] Among them, W (k) , B (k) represent the weight parameters that need to be trained for the k-th layer of the graph neural network, and J represents the number of nodes, which takes the value of 8 in the embodiments of the present invention.

[0072] During the model training process, it is assumed that the model is trained on data without abnormal information. For the data in each batch, after optimizing the graph neural network model, the predicted value (representing the predicted value of the k-th feature of the fighter j at the t-th moment that needs to be predicted in the i-th multi-aircraft confrontation) and the true value y itjk are numerically different, and the loss in this batch is recorded. Until all batches are traversed, the total loss function can be obtained as shown in formula (5):

[0073]

[0074] Among them, N represents the total amount of data, P represents the prediction degree, K represents the number of layers of the preset graph neural network, and the loss function obtains the model parameters through backpropagation.

[0075] Specifically, step S108 further includes: determining the change amount of the state quantity of each unmanned aerial vehicle in the unmanned formation participating in the intelligent air combat confrontation per unit time based on the predicted adversarial data and the current adversarial data of the unmanned formation participating in the intelligent air combat confrontation, and using the change amount as the decision set.

[0076] Preferably, the decision set includes: longitude change amount, latitude change amount, speed change amount, yaw angle change amount, pitch angle change amount, and roll angle change amount.

[0077] That is, finally, after the training of the optimized graph neural network model is completed and verified to be effective, when making predictions, the change amount of the variable per unit time can be used as the reference output of the decision set, as shown in Table 2:

[0078] Table 2 Model Output Variable List

[0079] Number Basic Operations of Aircraft Platform Number Basic Operations of Aircraft Platform 1 <![CDATA[Longitude change amount △L O > 4 Yaw Angle Variation △h 2 <![CDATA[Latitude change amount △L a > 5 Pitch Angle Variation △p 3 Speed Variation △v 6 Roll Angle Variation △r

[0080] Embodiment 2

[0081] The embodiment of the present invention provides a simulation experiment and prediction error evaluation of an intelligent air combat decision-making method for an unmanned formation based on an optimized graph neural network. Specifically,

[0082] (1) Scenario setting and data acquisition.

[0083] To verify the effectiveness of the method provided by the embodiment of the present invention, the present invention conducts a four-on-four scale air combat simulation experiment based on a certain air combat formation tactical confrontation platform, collects the data information of both sides of the confrontation 1000 times, the simulation time of each confrontation is 3 hours, and 10,800 pieces of data are collected for each confrontation. The data is stored using a certain data structure. The basic situation of the data set is shown in Table 3, and the data set is divided into a training set, a validation set, and a prediction set according to a ratio of 6:2:2.

[0084] Table 3 Basic Attribute Information of the Data Set

[0085]

[0086] The red and blue opposing aircraft are used as nodes, and the relative feature differences between every two aircraft are used as the edge attributes between nodes. The Euclidean distance between each pair of nodes is used as the weight to construct an undirected graph, thereby characterizing the topological information of both sides in the air combat confrontation. On the above dataset, 10 observation points are used, and the data with an observation step of 300 (i.e., the historical time series O = 10), that is, the air combat situation in the next 5 minutes is predicted with a 5-minute historical time window.

[0087] (2) Model training based on the optimized graph neural network.

[0088] In the embodiment of the present invention, Python 3.10 is used as the programming environment. The GNN convolution kernel size and the TCN convolution kernel size are both set to 3, the batch_size parameter is 4, the Adam optimizer is used for optimization, the iteration period is 50, and the learning rate LEARNING_RATE parameter is 0.0001. Some codes are shown in the following pseudo-code form:

[0089] Input: X ∈ R N×O×J×K , linearly transformed into E represents the expanded dimension Output:

[0090]

[0091]

[0092] Figure 6 is a schematic diagram of the change curve of the training loss value provided according to the embodiment of the present invention. As Figure 6 shown, in order to make the change curve of the loss function more intuitive, the present invention selects the abscissa as the training step step and the ordinate as the loss value. It can be seen from Figure 6 that the loss value gradually converges during the model training process and finally tends to be stable, indicating that the model plays a good fitting role for the tested dataset.

[0093] (3) Simulation result analysis and visualization.

[0094] After training, the generated data is externally connected to the intelligent air combat formation tactical confrontation platform. In order to control variables, the maneuver decision of the red side uses the intelligent agent after learning and training based on the optimized graph neural network, and the blue side uses the original intelligent agent based on the expert system in the platform for confrontation. The present invention assumes that medium-range air-to-air missiles can be simulated and launched under the condition of unmanned formation intelligent air combat, and the missile launch command is an external input command. Now the confrontation data is read and visualized through Tacview. The present invention selects a typical piece of data for analysis. The red and blue sides are initially 100 km apart. After at least 1 aircraft of the red side obtains the position information of all opposing aircraft through radar detection, a graph connection is established, and the red side can obtain the next maneuver action based on the training model.Figure 7 It is a schematic diagram of the situation when the red and blue sides first launch medium-range missiles according to an embodiment of the present invention. As Figure 7 shown, when the confrontation reaches about 5 minutes, the two sides shoot medium-range missiles at each other, and the red and blue sides successively perform tail-setting maneuvers to avoid, which basically conforms to the tactical action selection in this situation.

[0095] Figure 8 It is a schematic diagram of the situation when the red and blue sides second launch medium-range missiles according to an embodiment of the present invention. As Figure 8 shown, when the confrontation reaches about 8 minutes, at this time, 1 fighter plane of the red side is attacked by 3 fighter planes of the blue side, and 1 fighter plane of the blue side is attacked by 2 fighter planes of the red side. Both sides are committed to forming an asymmetric situation.

[0096] Figure 9 It is a schematic diagram of the final situation according to an embodiment of the present invention. As Figure 9 shown, when the confrontation reaches 20 minutes, the red side loses 2 aircraft, the blue side loses 1 aircraft, and the red side chooses to escape.

[0097] Through the analysis of the confrontation situation, basically some practical tactics can emerge on the red side, which proves that the optimized graph neural network has great potential in solving the problem of intelligent air combat decision-making for unmanned formations, and basically can achieve the transfer of knowledge and experience through learning, and the transfer effect is good.

[0098] (4) Model prediction error evaluation.

[0099] In addition to the above-mentioned intuitive observation-assisted qualitative analysis, the effectiveness of the model can also be evaluated quantitatively. Quantitative evaluation mainly includes aspects such as prediction error situation, fitting degree, and model stability. Among them, the prediction error situation is the focus of evaluation. The common idea is to use the samples in the validation set for testing, and observe the difference between its predicted value and the actual value, which is also called the residual. The smaller the residual, the better the model. The model is evaluated by whether the test error reaches the accuracy requirement to determine whether the model needs to be further improved. The coefficient of determination method can be used to measure the fitting effect, that is, the proportion of the samples that the model can explain. The calculation formula is shown in (6), where R 2 ranges from [0, 1], and the closer R 2 is to 1, the better the fitting effect, and the closer R 2 is to 0, the worse the effect.

[0100]

[0101] In the formula represents the sample mean; represents the residual; Denote the regression difference. Substitute the test set data of the present invention into the trained model to output the predicted value, and substitute the predicted value into formula (6) to obtain the determination coefficient R 2 = 0.78, which basically meets the requirements.

[0102] As can be seen from the above description, the embodiment of the present invention provides an intelligent air combat decision-making method for unmanned formations based on an optimized graph neural network, and proposes a simple and effective implementation scheme on the idea of improving the intelligence of unmanned formations through self-learning and training. After the unmanned formation makes a decision according to the model, certain cooperative tactics can emerge in the formation. The model fully considers the cumulative situation of individual aircraft, effectively combines the advantages in the spatial and temporal dimensions, and makes the model have a certain interpretability.

[0103] Embodiment III

[0104] Figure 10 is a schematic diagram of an intelligent air combat decision-making system for unmanned formations based on the optimized graph neural network provided by the embodiment of the present invention. As Figure 10 shown, the system includes: an acquisition module 10, a temporal feature extraction module 20, a spatial feature extraction module 30, and a decision-making module 40.

[0105] Specifically, the acquisition module 10 is used to acquire the simulated air combat confrontation data of the unmanned formation participating in the intelligent air combat confrontation.

[0106] The temporal feature extraction module 20 is used to input the simulated air combat confrontation data into the temporal convolutional neural network after dimensional change to obtain the temporal feature vector.

[0107] The spatial feature extraction module 30 is used to input the temporal feature vector into the trained graph neural network to obtain the predicted confrontation data of the unmanned formation.

[0108] The decision-making module 40 is used to determine the decision set of the unmanned formation participating in the intelligent air combat confrontation based on the predicted confrontation data.

[0109] Specifically, as Figure 10 shown, it further includes a training module 50, which is used for:

[0110] Construct a preset graph neural network based on the initial confrontation data of the unmanned formation participating in the intelligent air combat confrontation;

[0111] Train the preset graph neural network based on the preset training set to obtain the trained graph neural network.

[0112] Specifically, the decision-making module 40 is further used to determine the change amount of the state quantity of each unmanned aircraft in the unmanned formation participating in the intelligent air combat confrontation per unit time based on the predicted confrontation data and the current confrontation data of the unmanned formation participating in the intelligent air combat confrontation, and use the change amount as the decision set.

[0113] Preferably, the decision set includes: longitude change amount, latitude change amount, speed change amount, yaw angle change amount, pitch angle change amount, and roll angle change amount.

[0114] An embodiment of the present invention further provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the method in the first embodiment above is implemented.

[0115] An embodiment of the present invention further provides a computer-readable storage medium, where the computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method in the first embodiment above is implemented.

[0116] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

[0117] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An intelligent air combat decision-making method for unmanned formations based on optimized graph neural networks, characterized in that, Comprising: Obtaining simulated air combat confrontation data of a unmanned formation participating in intelligent air combat confrontation; After performing dimensionality change on the simulated air combat confrontation data, inputting it into a temporal convolutional neural network to obtain a temporal feature vector; Inputting the temporal feature vector into a trained graph neural network to obtain predicted confrontation data of the unmanned formation; Based on the predicted confrontation data, determining a decision set of the unmanned formation participating in intelligent air combat confrontation; Based on the predicted confrontation data, determining a decision set of the unmanned formation participating in intelligent air combat confrontation, including: Based on the predicted confrontation data and the current confrontation data of the unmanned formation participating in intelligent air combat confrontation, determining the change amount of the state quantity of each unmanned aerial vehicle in the unmanned formation participating in intelligent air combat confrontation per unit time, and taking the change amount as the decision set.

2. The method according to claim 1, wherein: The calculation formula of the temporal convolutional neural network includes: Among them, X TCN represents the timing feature vector, represents the value after linear variable expansion of the input variable, f(·) represents the activation function, d represents the dilation factor, k represents the convolutional kernel size, and s - d·i represents the past direction.

3. The method according to claim 1, wherein: The method further includes: Based on the initial confrontation data of the unmanned formation participating in intelligent air combat confrontation, constructing a preset graph neural network; Training the preset graph neural network based on a preset training set to obtain the trained graph neural network.

4. The method according to claim 3, wherein: The initial confrontation data includes: the longitude, latitude, yaw angle, pitch angle, roll angle, and speed of each unmanned aerial vehicle in the unmanned formation participating in intelligent air combat confrontation.

5. The method according to claim 3, characterized in that: The mathematical calculation formula of the preset graph neural network includes: Among them, W (k) , B (k) represent the weight parameters to be trained by the k-th layer of the graph neural network, J represents the number of nodes, and X TCN represents the time series feature vector, represents the prediction result of the j-th node in the k-th layer of the graph neural network; The training loss function of the preset graph neural network includes: Among them, N represents the total amount of data, P represents the prediction degree, and K represents the number of layers of the preset graph neural network. represents the predicted value, and y itjk represents the true value.

6. An unmanned formation intelligent air combat decision-making system based on an optimized graph neural network, characterized in that Comprising: An acquisition module, a temporal feature extraction module, a spatial feature extraction module, and a decision module; wherein, The acquisition module is used to obtain simulated air combat confrontation data of a unmanned formation participating in intelligent air combat confrontation; The temporal feature extraction module is used to perform dimensionality change on the simulated air combat confrontation data and then input it into a temporal convolutional neural network to obtain a temporal feature vector; The spatial feature extraction module is used to input the temporal feature vector into a trained graph neural network to obtain predicted confrontation data of the unmanned formation; The decision module is used to determine a decision set of the unmanned formation participating in intelligent air combat confrontation based on the predicted confrontation data; The decision module is further used to determine the change amount of the state quantity of each unmanned aerial vehicle in the unmanned formation participating in intelligent air combat confrontation per unit time based on the predicted confrontation data and the current confrontation data of the unmanned formation participating in intelligent air combat confrontation, and take the change amount as the decision set.

7. The system according to claim 6, characterized in that: It further includes a training module for: Based on the initial confrontation data of the unmanned formation participating in intelligent air combat confrontation, constructing a preset graph neural network; Training the preset graph neural network based on a preset training set to obtain the trained graph neural network.

8. An electronic device, characterized in that, Comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the computer program, it implements the method according to any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, and when the computer instructions are executed by the processor, they implement the method according to any one of claims 1-5.

Images