High-dynamic air formation cooperative behavior identification method and related device

Abstract

This invention relates to the technical field of military intelligence and computer vision, and discloses a method and related equipment for recognizing collaborative tactics in beyond-line-of-sight high-dynamic dual-aircraft formations. The method includes: constructing a spatiotemporal graph that fuses intra-frame spatial connectivity and inter-frame temporal correlation based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between friendly and enemy aircraft in continuous frame data; inputting feature vectors into a two-layer cascaded spatiotemporal feature extraction network, where each layer sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution extracts collaborative features of the dynamic space of the dual-aircraft nodes within the frame, and the parallel temporal feature fusion extracts local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuses them; outputting the recognition result of the collaborative tactical type performed by the dual-aircraft formation. Through the above method, this invention can comprehensively improve the accuracy of tactical recognition.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of military intelligence and computer vision, specifically to a method and related equipment for recognizing the cooperative behavior of high-dynamic aerial formations. Background Technology

[0002] Against the backdrop of the accelerating pace of intelligent air combat, dual-aircraft formations, as typical cooperative combat units, exhibit greater dynamism and complexity in their tactical behavior, with longer time spans and more complex spatial interactions in their tactical sequences. To address this change, air combat command systems must possess the ability to rapidly understand and adapt to cooperative tactical patterns, thereby achieving efficient decision support in the ever-changing battlefield environment. Beyond-visual-range (BVR) dual-aircraft formation cooperative tactical recognition (DFCTR) is a key technology for air combat decision-making, which can be transformed into pattern recognition of spatiotemporal data. Air combat scenarios themselves are characterized by high dynamism, strong adversarial nature, complex behavior, and strategic deception, making it extremely difficult to theoretically construct an analytical mapping relationship between tactical intentions and temporal state characteristics. In reality, tactical targets are often implicit in continuous time-series states; this temporal information typically includes target maneuver trajectories, flight paths, and related spatial coordinates. However, in complex tactical contexts, especially when deceptive intentions are involved, relying solely on the temporal evolution of aircraft position information for identification often fails to achieve the desired results.

[0003] Existing identification methods struggle to effectively characterize the highly dynamic spatiotemporal features of tactical data, specifically in the following two ways: 1. Existing methods employ fixed graph topologies, failing to dynamically adjust node connection weights based on the characteristics of different tactics, such as high-level offensives (weak spatial correlation between two aircraft) and decoy attacks (strong spatial correlation between two aircraft), resulting in limited accuracy in spatial feature extraction; 2. When processing long-term tactical data, existing methods suffer from information loss in the fusion of local and global features, making it difficult to simultaneously consider short-term maneuver details (such as sudden speed changes during the tactical initiation phase) and long-term tactical trends (such as the encirclement trajectory of the entire pincer attack), hindering the accurate capture of tactical spatiotemporal features and requiring improvement in identification accuracy. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide a method and related equipment for recognizing highly dynamic aerial formation cooperative behavior, in order to solve the problems existing in the prior art.

[0005] According to one aspect of the present invention, a beyond-visual-range high-dynamic dual-aircraft formation cooperative tactical identification method is provided, the method comprising:

[0006] The spatiotemporal graph construction steps are as follows: acquire continuous frame data of the two-aircraft formations of both sides in the air combat process; based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between the enemy and friendly aircraft in the continuous frame data, construct a spatiotemporal graph that integrates intra-frame spatial connectivity and inter-frame temporal correlation.

[0007] The spatiotemporal feature extraction step involves confirming the feature vector based on the spatiotemporal graph, and inputting the feature vector into a two-layer cascaded spatiotemporal feature extraction network for feature extraction. Each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution is used to generate a sample-level dynamic adjacency matrix based on data-driven methods to extract the collaborative features of the dynamic space of the two-machine nodes within the frame. The parallel temporal feature fusion is used to extract local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuse them.

[0008] The tactical identification step classifies the features extracted by the spatiotemporal feature extraction network via a two-layer concatenation and outputs the identification result of the cooperative tactical type executed by the dual-aircraft formation.

[0009] In one alternative approach, the spatiotemporal graph includes a node set V and an edge set E, wherein the node set V... In each frame, a fighter jet is represented as a node, and T represents the total number of frames. This represents the node information of the enemy's i-th fighter jet in frame t. The node information includes the three-dimensional coordinates of the enemy's i-th fighter jet, as well as the distance between the enemy's i-th fighter jet and our first fighter jet and the distance between the enemy's i-th fighter jet and our second fighter jet.

[0010] The edge set E includes the intra-frame spatial edge set. and inter-frame temporal edge set The intra-frame spatial edge represents the positional connection between two enemy aircraft within the same frame, and the inter-frame temporal edge represents the temporal connection between the same enemy aircraft in two adjacent frames.

[0011] In one alternative approach, the data-driven generation of the sample-level dynamic adjacency matrix includes:

[0012] The feature vectors are embedded using two preset embedding functions to obtain two embedding feature matrices;

[0013] The similarity matrix is ​​obtained by multiplying the two embedded feature matrices. The similarity matrix Each element in Represents a node and nodes Similarity;

[0014] The parameterized coefficients are learned through a gating mechanism. The parameterized coefficients Used to adjust the similarity matrix Importance at different layers;

[0015] The similarity matrix With the parameterized coefficients Multiplication yields the sample-level dynamic adjacency matrix. .

[0016] In an alternative approach, the adaptive spatial graph convolution is implemented using the following formula:

[0017] ,

[0018] in, The feature vector is input to the adaptive spatial graph convolution. The output vector of the adaptive spatial graph convolution. The size is Where C is the batch size, T is the number of channels, V is the number of frames, k is the number of nodes, and K is the spatial graph convolution kernel size. It is a 1*1 convolution.

[0019] In one alternative approach, the parallel temporal feature fusion includes a local branch and a global branch, wherein the local branch is used to extract local short-time maneuver detail features, and the global branch is used to extract global long-time features;

[0020] The output vector of the adaptive spatial graph convolution is input in parallel to the local branch and the global branch, and the output results of the local branch and the global branch are weighted and fused through a learnable weighted fusion mechanism.

[0021] In one alternative approach, the local branch is a temporal convolutional network comprising one-dimensional convolution and dilated convolution, and the global branch is a multi-layered stacked Transformer encoder, each of which includes a multi-head self-attention and feedforward network.

[0022] According to another aspect of the present invention, a beyond-line-of-sight high-dynamic dual-aircraft formation cooperative tactical identification device is provided, the device comprising:

[0023] The spatiotemporal graph construction module is used to acquire continuous frame data of the two-aircraft formations of both sides in the air combat process. Based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between the enemy and friendly aircraft in the continuous frame data, a spatiotemporal graph that integrates intra-frame spatial connectivity and inter-frame temporal correlation is constructed.

[0024] The spatiotemporal feature extraction module is used to confirm feature vectors based on the spatiotemporal graph, and input the feature vectors into a two-layer cascaded spatiotemporal feature extraction network for feature extraction. Each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution is used to generate a sample-level dynamic adjacency matrix based on data-driven generation to extract the collaborative features of the dynamic space of the two-machine nodes within the frame. The parallel temporal feature fusion is used to extract local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuse them.

[0025] The tactical identification module is used to classify the features extracted by the spatiotemporal feature extraction network via a two-layer concatenation, and output the identification result of the cooperative tactical type executed by the dual-aircraft formation.

[0026] In one alternative approach, the spatiotemporal graph includes a node set V and an edge set E, wherein the node set V... In each frame, a fighter jet is represented as a node, and T represents the total number of frames. This indicates that the enemy's i-th fighter jet is in the... The node information of the frame includes the three-dimensional coordinates of the enemy's i-th fighter jet, as well as the distance between the enemy's i-th fighter jet and our first fighter jet and the distance between our second fighter jet and our first fighter jet.

[0027] The edge set E includes the intra-frame spatial edge set. and inter-frame temporal edge set The intra-frame spatial edge represents the positional connection between two enemy aircraft within the same frame, and the inter-frame temporal edge represents the temporal connection between the same enemy aircraft in two adjacent frames.

[0028] According to another aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform the method described above.

[0029] According to another aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the method described above.

[0030] This invention first constructs a spatiotemporal graph based on five-dimensional features composed of the three-dimensional coordinates of the two machines and the distance between them and the enemy / ally. This spatiotemporal graph integrates intra-frame spatial connectivity and inter-frame temporal correlation. Then, a two-layer cascaded spatiotemporal feature extraction network is constructed. Each network integrates adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution generates a sample-level dynamic adjacency matrix driven by data, replacing the traditional fixed topology and accurately matching the spatial correlation differences of different tactics. The parallel temporal feature fusion uses local branches to model short-term maneuver details and global branches to capture global long-term dependencies. Learnable weighted fusion achieves complementarity between local and global temporal features. Finally, the tactical type identification result is output in the classification output layer. This invention, based on a spatiotemporal graph of spatial dynamic modeling and temporal complementary fusion, utilizes adaptive spatial graph convolution to capture the dynamic interaction relationships between the two machine nodes. It can adjust the connection weights of the two machine nodes in real time according to the spatial correlation characteristics of different tactics, improving the targeting and accuracy of spatial feature expression. By achieving collaborative extraction of local and global temporal features in parallel, it comprehensively improves the accuracy, expressiveness, and generalization performance of tactical identification.

[0031] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0032] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0033] Figure 1 The diagram shows a network block diagram of the beyond-line-of-sight high dynamic dual-aircraft formation cooperative tactical identification method provided in an embodiment of the present invention.

[0034] Figure 2 A flowchart illustrating the beyond-line-of-sight, high-dynamic dual-aircraft formation cooperative tactical identification method provided in an embodiment of the present invention is shown.

[0035] Figure 3 A schematic diagram of a dual-machine encirclement and cross-attack is shown;

[0036] Figure 4 A schematic diagram of the spatiotemporal feature extraction network structure is shown.

[0037] Figure 5 This diagram illustrates the structure of the beyond-line-of-sight high-dynamic dual-aircraft formation cooperative tactical identification device provided in an embodiment of the present invention.

[0038] Figure 6A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0039] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0040] In the process of modern intelligent air combat, the tactical behavior of two-aircraft formations exhibits high dynamism and complexity. Traditional spatiotemporal graph convolutional networks suffer from limitations in recognizing the cooperative tactical characteristics of beyond-visual-range (BVR) two-aircraft formations due to their fixed spatial topology and insufficient long-term dependency modeling, making it difficult to accurately capture tactical spatiotemporal features. Therefore, this invention proposes a high-dynamic beyond-visual-range (BVR) two-aircraft formation cooperative tactical recognition method, such as... Figure 1 As shown, firstly, a spatiotemporal graph is constructed. Based on the five-dimensional features composed of the dual-machine three-dimensional coordinates and enemy-friendly distance, a spatiotemporal graph fusing intra-frame spatial connectivity and inter-frame temporal correlation is built. Secondly, a two-layer cascaded spatiotemporal feature extraction network is constructed. Each spatiotemporal feature extraction network integrates adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution generates a sample-level dynamic adjacency matrix driven by data, replacing the traditional fixed topology, accurately matching the spatial correlation differences of different tactics, and introducing a gating mechanism to adjust the importance of different layers of dynamic graphs. The parallel temporal feature fusion uses a TCN branch to model short-term maneuver details and a Transformer branch to capture global long-term dependencies. Learnable weighted fusion is used to achieve complementarity between local and global temporal features. Finally, in the classification output layer, after global average pooling, a fully connected layer, and a classifier, the tactical type identification result is output.

[0041] Figure 2 A flowchart illustrating the beyond-line-of-sight, high-dynamic dual-aircraft formation cooperative tactical identification method provided in an embodiment of the present invention is shown, as follows: Figure 2 As shown, the method includes the following steps:

[0042] Step 10, Spatiotemporal graph construction step: Obtain continuous frame data of the two-aircraft formations of both sides in the air combat process. Based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between the enemy and friendly aircraft in the continuous frame data, construct a spatiotemporal graph that integrates intra-frame spatial connectivity and inter-frame temporal correlation.

[0043] The input to the spatiotemporal graph consists of a series of consecutive frames, each containing multidimensional state information and implicit features of several nodes. For example... Figure 1As shown, continuous frame data is obtained by extracting radar data from the fighter jets, duplicate data is detected and processed, and data fusion is performed to finally obtain time-series data. To fully characterize the motion characteristics of the two-aircraft formation in cooperative tactics, each fighter jet is taken as a node, and five-dimensional features of each node are extracted from the time-series data. The five-dimensional features include: the three-dimensional coordinates of each enemy fighter jet, the distance of each enemy fighter jet to the first friendly fighter jet, and the distance to the second friendly fighter jet. Based on the five-dimensional features, a spatiotemporal map of the two-aircraft formation containing intra-frame and inter-frame connectivity relationships is constructed.

[0044] The spatiotemporal graph includes a set of nodes V and a set of edges. The set of nodes In each frame, a fighter jet is represented as a node, and T represents the total number of frames. This represents the node information of the enemy's i-th fighter jet in frame t. The node information includes the three-dimensional coordinates of the enemy's i-th fighter jet, as well as the distance between the enemy's i-th fighter jet and our first fighter jet and the distance between the enemy's i-th fighter jet and our second fighter jet.

[0045] The edge set E includes the intra-frame spatial edge set. and inter-frame temporal edge set The intra-frame spatial edge represents the positional connection between two enemy aircraft within the same frame, and the inter-frame temporal edge represents the temporal connection between the same enemy aircraft in two adjacent frames.

[0046] ,

[0047] ,

[0048] in, This indicates the connection relationship between the positions of two enemy fighters in frame t. This represents the connection relationship between the i-th fighter jet at frame t and the adjacent frame t+1.

[0049] In multi-machine interaction relationships, such as Figure 3 As shown, taking two typical tactics, a two-aircraft encirclement and a cross-attack, as examples, if modeling is based solely on unilateral data, the geometric trajectories of both tactics exhibit a shrinking distance between the two aircraft over time, making it difficult to distinguish them using positional information. However, at the tactical semantic level, a two-aircraft encirclement manifests as a coordinated suppression of a target by two aircraft, while a cross-attack represents a separate attack after the trajectories of the two aircraft intersect. Ignoring the interaction between friend and foe, relying solely on coordinate evolution characteristics will make it difficult for the model to recognize these semantic differences.

[0050] This invention introduces dynamic distance features between friendly and enemy aircraft during the spatiotemporal map construction stage to enhance the ability to characterize multi-aircraft interactions. If the initial coordinate system of the dataset is the LLA coordinate system, it needs to be converted to the ENU coordinate system for easier subsequent processing and to unify the dimensions. Specifically, the LLA coordinate system needs to be converted to the ECEF coordinate system first, and then the ECEF coordinate system needs to be converted to the ENU coordinate system.

[0051] After coordinate system transformation, node information From the original Convert to .in, Let these represent the longitude, latitude, and altitude of the i-th fighter jet in the LLA coordinate system at time t, respectively. Let x, y, and z represent the x, y, and z coordinates of the i-th fighter jet in the ENU coordinate system at time t, respectively. These represent the distances between the enemy's i-th fighter and our first and second fighters at time t, respectively. This data is obtained directly from radar.

[0052] By introducing the dynamic range of enemy and friendly aircraft This significantly enhances the ability to depict multi-machine interaction relationships, such as... Figure 3 As shown, in the case of a two-aircraft encirclement, while the distance between the two enemy aircraft shortens, the distance between them and one of our aircraft also decreases simultaneously; while in the case of a cross attack, the distance between the two enemy aircraft and our two aircraft shows a trend of one decreasing and the other increasing. This difference in characteristics significantly improves the discriminability of the spatiotemporal graph at the tactical semantic level, enabling the model to accurately distinguish different cooperative behavior patterns even when the trajectories are similar.

[0053] Step 20, Spatiotemporal Feature Extraction Step: Based on the spatiotemporal graph, the feature vector is confirmed, and the feature vector is input into a two-layer cascaded spatiotemporal feature extraction network for feature extraction. Each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution is used to generate a sample-level dynamic adjacency matrix based on data-driven methods to extract the collaborative features of the dynamic space of the two-machine nodes within the frame. The parallel temporal feature fusion is used to extract local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuse them.

[0054] See also Figure 1 and Figure 4 Based on the spatiotemporal graph, the node information of the fighter jet is... Mapped to feature vectors Then, the feature vector is input into a two-layer cascaded spatiotemporal feature extraction network for feature extraction, such as... Figure 4As shown, each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. Figure 4 In this embodiment, spatial feature extraction AGCN-T is an adaptive spatial graph convolution, and its input is the feature vector. The output is Adaptive spatial graph convolution automatically adjusts based on the specific air combat scenario (i.e., the sample) currently in progress, precisely focusing on the most critical dual-aircraft interaction relationship in the current tactic. This automatic adjustment is based on the real-time observed data itself, i.e., data-driven.

[0055] Furthermore, the data-driven generation of sample-level dynamic adjacency matrices includes:

[0056] The feature vectors are embedded using two preset embedding functions to obtain two embedding feature matrices;

[0057] The similarity matrix is ​​obtained by multiplying the two embedded feature matrices. The similarity matrix Each element in Represents a node and nodes Similarity;

[0058] The parameterized coefficients are learned through a gating mechanism. The parameterized coefficients Used to adjust the similarity matrix Importance at different layers;

[0059] The similarity matrix With the parameterized coefficients Multiplication yields the sample-level dynamic adjacency matrix. .

[0060] In adaptive spatial graph convolution, the learnable embedding function is a key component for generating the adaptive adjacency matrix. The embedding function maps the feature vectors of nodes to an embedding space, obtaining embedded features, which is equivalent to interpreting the feature vectors of nodes from different perspectives. Then, the data-driven adaptive adjacency matrix is ​​generated by calculating the similarity between the embedded features. These are two independent embedding functions, capable of learning asymmetric node relationships. The two independent embedding spaces provide stronger expressive power. Multiplying the two embedding features yields the similarity matrix. Similarity matrix It involves using a normalized Gaussian function to construct a matrix for specific air combat scenarios (i.e., samples) to calculate the similarity between two nodes. Based on the specific location, trajectory, and other features of the nodes in the sample, it determines the connection and strength between the nodes, thus ensuring... It can generate unique connectivity information for each sample, more accurately reflecting the actual situation of the sample.

[0061] Specifically, for feature vectors ,pass Two embedding functions are embedded into the feature space. ,in, The number of channels in the embedding space, T is the number of frames, and N is the batch size, determined by selection. The convolutional layer is used as the embedding function to reconstruct the feature maps of the two embeddings into embedding feature matrices respectively. :

[0062] ,

[0063] Multiplying these two embedding feature matrices yields the similarity matrix. Here, k represents the aircraft's index. In adaptive spatial graph convolution, the adjacency matrix is ​​typically divided into multiple subsets according to a certain strategy, and k is the index of these subsets. In this embodiment, the two aircraft are simply divided into two subsets, with each aircraft forming one subset, and k has a maximum value of 2. Similarity matrix Each element in Represents a node and nodes Similarity, similarity matrix The value is normalized to 0-1 and treated as an edge between two nodes:

[0064] , ,

[0065] Similarity matrix Each element in Represents a node and nodes The dynamic spatial coordination strength in the current frame and sample. If the two machines coordinate closely (such as in "luring and striking combined"), their feature similarity is high. The value is large; if the coordination is weak (such as "high-position strong attack"), The smaller the value, the more dynamic the relationship blueprint is generated, which is specific to the current tactical scenario.

[0066] In adaptive spatial graph convolution, if the number of input channels and output channels are inconsistent, a 1x1 convolution is inserted in the residual path to transform the input and match the output in the channel dimension. Learnable parameterized coefficients. The need for unique graphs varies across different layers; for example, deeper networks require processing more abstract features compared to shallower layers. This embodiment learns parameterized coefficients through a gating mechanism. ,pass Multiply by parameterization coefficients This allows for dynamic adjustment. The importance of different layers, parameterization coefficients Each layer is unique, and the data-driven training process is constantly learning and updating. For example, in deep networks, the model can learn to increase... More dependent The data-driven relationships revealed.

[0067] Existing methods, when dealing with highly dynamic spatiotemporal variations, employ fixed graph topologies, which cannot dynamically adjust node connection weights according to different tactical characteristics, thus limiting the accuracy of spatial feature extraction. This embodiment addresses this by using a data-driven, sample-level dynamic adjacency matrix. Based on the spatial correlation characteristics of different tactics, the connection weight of the dual-machine nodes can be adjusted in real time and dynamically to avoid spatial feature expression deviation caused by fixed topology structure, and significantly improve the accuracy of characterization of tactical spatial interaction mode.

[0068] Furthermore, the adaptive spatial graph convolution is implemented using the following formula: ,

[0069] in, The feature vector is input to the adaptive spatial graph convolution. The output vector of the adaptive spatial graph convolution. The size is For batch size, Where is the number of channels, T is the number of frames, V is the number of nodes, k is the number of layers, and K is the size of the spatial graph convolution kernel. It is a 1*1 convolution.

[0070] Obtaining the sample-level dynamic adjacency matrix Then, with the feature vector Perform matrix multiplication operations This completes the aggregation of general graph convolutions. In this aggregation operation, the features of each node are fused with the information of its neighboring nodes, weighted according to the dynamic cooperation strength. If the cooperation is strong... If the information fusion is deep, the extracted features will better reflect collaboration; if the collaboration is weak... If the fusion is shallow, the features will be more independent. At this point, the collaborative features of the intra-frame dual-machine node dynamic space have been initially extracted. Then, a 1*1 convolution operation is used to perform a linear transformation and channel-dimensional integration on the aggregated features, mapping them to a new feature space. Finally, a residual connection is inserted to obtain the output. It is a high-level representation rich in the collaborative features of the dynamic space of the two machine nodes within the frame. It is no longer the original coordinates, but encodes the spatial relationship semantic information of who is with whom and under what collaborative strength.

[0071] Furthermore, the parallel temporal feature fusion includes a local branch and a global branch. The local branch is used to extract local short-term maneuver detail features, and the global branch is used to extract global long-term features. The output vector of the adaptive spatial graph convolution is input to the local branch and the global branch in parallel, and the output results of the local branch and the global branch are weighted and fused through a learnable weighted fusion mechanism.

[0072] In beyond-visual-range (BVR) dual-aircraft formation cooperative tactical recognition, the tactical time series typically manifests as a complex dynamic process involving both short-term maneuvers and long-term coordination. To balance short-term local features and long-term global features in the spatiotemporal graph of dual-aircraft formation tactical actions, this invention constructs a parallel temporal feature fusion mechanism. Local branches extract local short-term maneuver detail features, while global branches extract global long-term features. The outputs of the local and global branches are then weighted and fused using a learnable weighted fusion mechanism. This achieves complementary modeling of short-term action details and global coordination trends, effectively improving the model's recognition performance under complex tactical modes.

[0073] Furthermore, the local branch is a temporal convolutional network, which includes one-dimensional convolution and dilated convolution, and the global branch is a multi-layer stacked Transformer encoder, each of which includes a multi-head self-attention and feedforward network.

[0074] While a standalone temporal convolutional network can capture local temporal dependencies, it struggles to model global collaborative relationships across time periods. Conversely, directly connecting a temporal convolutional network to a Transformer encoder enhances global perception but weakens the model's sensitivity to short-term action details. This is particularly problematic in datasets requiring beyond-line-of-sight dual-aircraft formation tactical recognition, where the number of nodes is minimal (only two), spatial relationships are simple, and temporal variations are significant. Using a concatenated structure in such cases can lead to spatial feature redundancy, insufficient temporal dependency modeling, and overfitting or underfitting. This embodiment utilizes a temporal convolutional network with one-dimensional convolution and dilated convolution to efficiently extract local dynamic features in the temporal dimension, focusing on short-term action details and transient maneuver changes within the spatiotemporal graph. The multi-layered stacked Transformer encoder, with each component comprising a multi-head self-attention and feedforward network, allows for parallel attention to the relationships between different temporal segments, capturing long-range dependencies and global collaborative patterns. Finally, a learnable weighted fusion mechanism integrates the parallel outputs of the two branches proportionally, achieving a balance between local and global features. The fused output is then fed into a classification head for tactical discrimination.

[0075] like Figure 1 As shown, the spatiotemporal feature extraction network consists of a two-layer, cascaded structure. The effectiveness of deep learning models, especially those processing spatiotemporal sequence data, often depends on the depth of the layers. Shallow networks typically learn low-level, local features, while deep networks can combine these low-level features into more complex, global abstract representations. This hierarchical learning is crucial in the tactical identification task of two-aircraft formations. The first layer of the spatiotemporal feature extraction network receives raw feature vectors based on the spatiotemporal graph. These feature vectors contain basic motion information (coordinates) and simple interaction information (range between friendly and enemy aircraft). This layer primarily captures basic spatial relationships and short-term patterns. For example, it can learn the basic spatial relationship of whether the two aircraft are close or far apart at a given moment, as well as simple maneuvers (such as turning and acceleration) of the aircraft within several consecutive frames. The second layer of the spatiotemporal feature extraction network receives the features output from the first layer, which have undergone preliminary refinement. These features are no longer raw data but rather representations containing certain basic tactical elements. At this layer, based on the features extracted in the first layer, more complex tactical coordination patterns can be learned. For example, it combines basic patterns such as "continuous approach" and "change in height difference" to identify overall tactical intentions that require long-term observation, such as "high-level attack" or "pincer attack".

[0076] Furthermore, the first-layer adaptive spatial graph convolution primarily focuses on geometric spatial relationships, such as calculating the basic coordination strength between the two aircraft based on their real-time positions. The parallel temporal feature fusion of the first layer emphasizes extracting local temporal features. The second-layer adaptive spatial graph convolution learns more abstract tactical-level coordination relationships. For example, even if the two aircraft are geographically far apart, if their feature change patterns are highly synchronized, they can be assigned high connection weights in deeper layers. The parallel temporal feature fusion of the second layer focuses on understanding long-term tactical logic, such as connecting a series of continuous maneuvers to understand the complete execution process of the entire luring-the-enemy-in-depth or pincer attack tactic.

[0077] Step 30, tactical identification step, classifies the features extracted by the spatiotemporal feature extraction network via the two-layer concatenation, and outputs the identification result of the cooperative tactical type executed by the dual-aircraft formation.

[0078] like Figure 1 As shown, in tactical identification, the output of the two-layer cascaded spatiotemporal feature extraction network is processed by global average pooling, fully connected layers and classifiers to output the tactical type identification result. The tactical type identification result is output in the form of a probability distribution, and the tactic with the highest probability is the final tactic identified.

[0079] This invention first constructs a spatiotemporal graph based on five-dimensional features composed of the three-dimensional coordinates of the two machines and the distance between them and the enemy / ally. This spatiotemporal graph integrates intra-frame spatial connectivity and inter-frame temporal correlation. Then, a two-layer cascaded spatiotemporal feature extraction network is constructed. Each network integrates adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution generates a sample-level dynamic adjacency matrix driven by data, replacing the traditional fixed topology and accurately matching the spatial correlation differences of different tactics. The parallel temporal feature fusion uses local branches to model short-term maneuver details and global branches to capture global long-term dependencies. Learnable weighted fusion achieves complementarity between local and global temporal features. Finally, the tactical type identification result is output in the classification output layer. This invention, based on a spatiotemporal graph of spatial dynamic modeling and temporal complementary fusion, utilizes adaptive spatial graph convolution to capture the dynamic interaction relationships between the two machine nodes. It can adjust the connection weights of the two machine nodes in real time according to the spatial correlation characteristics of different tactics, improving the targeting and accuracy of spatial feature expression. By achieving collaborative extraction of local and global temporal features in parallel, it comprehensively improves the accuracy, expressiveness, and generalization performance of tactical identification.

[0080] For example, this embodiment of the invention uses the combat data records of the red and blue sides in a certain simulation system as the research object. The experiment records the entire process of beyond-visual-range air combat and close-range dogfighting between two fighter jets from each side after takeoff. Based on the data, a two-aircraft formation tactical space is established, including 10 tactical types such as pincer attack, combined decoy attack, cross attack, high-altitude attack, two-aircraft dispersal, luring the enemy in, and two-aircraft encirclement. These are shown in Table 1 below.

[0081] Tactical types Feature Description Pincer attack When the two aircraft detect an enemy aircraft, they will first spread out and then engage the enemy aircraft in a pincer attack. Inducing and striking The decoy aircraft is locked on and evades, while the attack aircraft attacks the enemy aircraft from the flanks. Cross attack Two aircraft attack enemy aircraft in a cross-attack formation. High-level attack Upon spotting the enemy aircraft, the lead aircraft quickly climbed to a higher altitude and launched a long-range strike. Dual-machine evacuation The two machines can attack from a horizontal or vertical distance. Luring the enemy deep into the territory The decoy aircraft is locked on and will lure the enemy aircraft away to achieve the desired decoy effect. Double-machine encirclement Two aircraft attack the same target from different directions (sides). Horizontal line The two aircraft advance in a formation that maintains a basically parallel arrangement in the lateral direction. column The two aircraft proceeded in formation, arranged in a front-to-back sequence. echelon The two aircraft proceeded in a formation with a certain degree of stagger in altitude and front-to-back position. Table 1

[0082] Using annotation software, 10 tactical types were labeled based on prominent tactical features, generating a total of 3375 samples across 10 tactical types.

[0083] To standardize the input format for network training, this example proposes a sample cleaning and processing algorithm. All samples are standardized to a length of 300 frames. While preserving tactical dynamic features, the algorithm performs frame padding, frame deletion, and filtering of samples. The specific process is as follows:

[0084] (1) Sample screening: Count the number of frames n for each sample. If n < 50, the information is insufficient, and the sample is directly removed to reduce training noise.

[0085] (2) Frame interpolation (50≤n<300): Calculate the number of frames to be interpolated Δ=300-n, insert new frames at uniform intervals, and generate new frames by linear interpolation of adjacent frames to maintain data continuity and smoothness.

[0086] (3) Frame deletion processing (n>300): Calculate the number of frames exceeding Δ=n-300, select deletion positions at equal intervals, effectively compress the data length, and retain the timing characteristics of the original tactical actions.

[0087] Through the above processing, a unified and complete dataset was constructed, providing high-quality and standardized input samples for model training.

[0088] The evaluation metrics of this invention mainly include overall recognition accuracy, number of network parameters (Params), single-frame recognition speed (Speed1F), and overall recognition speed over 300 frames (Speed300F).

[0089] Overall recognition accuracy represents the proportion of samples correctly classified by the network out of all samples, and is an evaluation metric for assessing the overall performance of the network.

[0090] Network parameters: used to measure the complexity of a network and the computational resources required. A smaller number of parameters usually means that the network is more lightweight and easier to deploy on platforms with limited computing power.

[0091] Single-frame recognition speed (Speed1F) and overall recognition speed of 300 frames (Speed300F): These two metrics reflect the network's response speed when processing single-frame and continuous multi-frame data, respectively, and are directly related to the system's real-time performance and actual efficiency in application scenarios.

[0092] In the example, stochastic gradient descent (SGD) was used for training with a learning rate of 0.005. The learning rate decreased by a factor of 10 every 10 training iterations, for a total of 200 iterations. The batch size of the training and test sets was set to 64 in each iteration. Finally, the highest accuracy on the test set during the 200 iterations was selected as the network's accuracy, and the precision and recall for tactical identification were calculated. The experimental results are shown in Table 2 below. Table 2 includes the traditional spatiotemporal graph convolutional network STGCN (Baseline), the convolution-based temporal network RGB-CNN, the LSTM and GRU models based on recurrent neural structures, and the BVR-CTRN of this invention.

[0093] Network methods Recognition accuracy Params Speed1F Speed300F STGCN (Baseline) 83.51% 1.77M 3.89ms 19.69ms RGB-CNN 65.45% 12.06M / 300.00ms LSTM

[70] 72.63% 5.26M 7.33ms 34.67ms GRU

[71] 77.28% 3.33M 5.21ms 30.20ms BVR-CTRN(Ours) 90.43% 5.26M 7.20ms 88.89ms Table 2

[0094] The tactical recognition accuracy of this invention reaches over 90.43%, which is significantly higher than the recognition accuracy of existing technologies. It has stronger spatiotemporal feature modeling and collaborative behavior discrimination capabilities in complex tactical scenarios, verifying the effectiveness of the adaptive spatial graph convolution and parallel temporal feature fusion mechanism.

[0095] Figure 5A schematic diagram of the structure of a beyond-line-of-sight, high-dynamic dual-aircraft formation cooperative tactical identification device according to an embodiment of the present invention is shown. Figure 5 As shown, the device includes:

[0096] The spatiotemporal graph construction module 501 is used to acquire continuous frame data of the two-aircraft formations of both sides in the air combat process. Based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between the enemy and friendly aircraft in the continuous frame data, a spatiotemporal graph that integrates intra-frame spatial connectivity and inter-frame temporal correlation is constructed.

[0097] The spatiotemporal feature extraction module 502 is used to confirm the feature vector based on the spatiotemporal graph, and input the feature vector into a two-layer cascaded spatiotemporal feature extraction network for feature extraction. Each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution is used to generate a sample-level dynamic adjacency matrix based on data-driven generation to extract the collaborative features of the dynamic space of the two-machine nodes within the frame. The parallel temporal feature fusion is used to extract local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuse them.

[0098] The tactical identification module 503 is used to classify the features extracted by the spatiotemporal feature extraction network via a two-layer concatenation, and output the identification result of the cooperative tactical type executed by the dual-aircraft formation.

[0099] The embodiments of the beyond-visual-range high-dynamic dual-aircraft formation cooperative tactical identification device are basically the same as the embodiments of the above-described beyond-visual-range high-dynamic dual-aircraft formation cooperative tactical identification method, and can be referred to the above embodiments.

[0100] Figure 6 The diagram shows a structural schematic of an embodiment of the computer device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0101] like Figure 6 As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0102] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other computer devices, such as clients or other server network elements. The processor 402 executes program 410, specifically performing the relevant steps described above in the computer device embodiment.

[0103] Specifically, program 410 may include program code, which includes computer-executable instructions.

[0104] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0105] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0106] Specifically, program 410 can be called by processor 402 to cause the computer device to perform the following operations:

[0107] The spatiotemporal graph construction steps are as follows: acquire continuous frame data of the two-aircraft formations of both sides in the air combat process; based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between the enemy and friendly aircraft in the continuous frame data, construct a spatiotemporal graph that integrates intra-frame spatial connectivity and inter-frame temporal correlation.

[0108] The spatiotemporal feature extraction step involves confirming the feature vector based on the spatiotemporal graph, and inputting the feature vector into a two-layer cascaded spatiotemporal feature extraction network for feature extraction. Each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution is used to generate a sample-level dynamic adjacency matrix based on data-driven methods to extract the collaborative features of the dynamic space of the two-machine nodes within the frame. The parallel temporal feature fusion is used to extract local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuse them.

[0109] The tactical identification step classifies the features extracted by the spatiotemporal feature extraction network via a two-layer concatenation and outputs the identification result of the cooperative tactical type executed by the dual-aircraft formation.

[0110] In one alternative approach, the spatiotemporal graph includes a node set V and an edge set E, wherein the node set V... In each frame, a fighter jet is represented as a node, and T represents the total number of frames. This represents the node information of the enemy's i-th fighter jet in frame t. The node information includes the three-dimensional coordinates of the enemy's i-th fighter jet, as well as the distance between the enemy's i-th fighter jet and our first fighter jet and the distance between the enemy's i-th fighter jet and our second fighter jet.

[0111] The edge set E includes the intra-frame spatial edge set. and inter-frame temporal edge set The intra-frame spatial edge represents the positional connection between two enemy aircraft within the same frame, and the inter-frame temporal edge represents the temporal connection between the same enemy aircraft in two adjacent frames.

[0112] In one alternative approach, the data-driven generation of the sample-level dynamic adjacency matrix includes:

[0113] The feature vectors are embedded using two preset embedding functions to obtain two embedding feature matrices;

[0114] The similarity matrix is ​​obtained by multiplying the two embedded feature matrices. The similarity matrix Each element in Represents a node and nodes Similarity;

[0115] The parameterized coefficients are learned through a gating mechanism. The parameterized coefficients Used to adjust the similarity matrix Importance at different layers;

[0116] The similarity matrix With the parameterized coefficients Multiplication yields the sample-level dynamic adjacency matrix. .

[0117] In an alternative approach, the adaptive spatial graph convolution is implemented using the following formula:

[0118] ;

[0119] in, The feature vector is input to the adaptive spatial graph convolution. The output vector of the adaptive spatial graph convolution. The size is For batch size, Where is the number of channels, T is the number of frames, V is the number of nodes, k is the number of layers, and K is the size of the spatial graph convolution kernel. It is a 1*1 convolution.

[0120] In one alternative approach, the parallel temporal feature fusion includes a local branch and a global branch, wherein the local branch is used to extract local short-time maneuver detail features, and the global branch is used to extract global long-time features;

[0121] The output vector of the adaptive spatial graph convolution is input in parallel to the local branch and the global branch, and the output results of the local branch and the global branch are weighted and fused through a learnable weighted fusion mechanism.

[0122] In one alternative approach, the local branch is a temporal convolutional network comprising one-dimensional convolution and dilated convolution, and the global branch is a multi-layered stacked Transformer encoder, each of which includes a multi-head self-attention and feedforward network.

[0123] This invention provides a computer-readable storage medium storing at least one executable instruction that, when executed on a computer device, causes the computer device to perform any of the above-described method embodiments.

[0124] This invention provides a computer program that can be invoked by a processor to cause a computer device to execute any of the above-described method embodiments.

[0125] This invention provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions that, when executed on a computer, cause the computer to perform any of the above-described method embodiments.

[0126] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the content of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0127] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0128] Similarly, it should be understood that, in order to streamline the invention and aid in understanding one or more of the various aspects of the invention, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim.

[0129] Those skilled in the art will understand that modules in the computer device of the embodiments can be adaptively modified and placed in one or more computer devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or computer device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0130] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A beyond-visual-range, high-dynamic dual-aircraft formation cooperative tactical identification method, characterized in that, include: The spatiotemporal graph construction steps are as follows: acquire continuous frame data of the two-aircraft formations of both sides in the air combat process; based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between the enemy and friendly aircraft in the continuous frame data, construct a spatiotemporal graph that integrates intra-frame spatial connectivity and inter-frame temporal correlation. The spatiotemporal feature extraction step involves confirming the feature vector based on the spatiotemporal graph, and inputting the feature vector into a two-layer cascaded spatiotemporal feature extraction network for feature extraction. Each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution is used to generate a sample-level dynamic adjacency matrix based on data-driven methods to extract the collaborative features of the dynamic space of the two-machine nodes within the frame. The parallel temporal feature fusion is used to extract local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuse them. The tactical identification step classifies the features extracted by the spatiotemporal feature extraction network via a two-layer concatenation and outputs the identification result of the cooperative tactical type executed by the dual-aircraft formation. The data-driven generation of sample-level dynamic adjacency matrices includes: The feature vectors are embedded using two preset embedding functions to obtain two embedding feature matrices; The similarity matrix is ​​obtained by multiplying the two embedded feature matrices. The similarity matrix Each element in Represents a node and nodes Similarity; The parameterized coefficients are learned through a gating mechanism. The parameterized coefficients Used to adjust the similarity matrix Importance at different layers; The similarity matrix With the parameterized coefficients Multiplication yields the sample-level dynamic adjacency matrix. ; The adaptive spatial graph convolution is implemented using the following formula: ， in, The feature vector is input to the adaptive spatial graph convolution. The output vector of the adaptive spatial graph convolution. The size is N is the batch size, C is the number of channels, T is the number of frames, V is the number of nodes, k is the number of layers, and K is the spatial graph convolution kernel size. It is a 1*1 convolution; The parallel temporal feature fusion includes a local branch and a global branch. The local branch is used to extract local short-term maneuver detail features, and the global branch is used to extract global long-term features. The output vector of the adaptive spatial graph convolution is input in parallel to the local branch and the global branch, and the output results of the local branch and the global branch are weighted and fused through a learnable weighted fusion mechanism.

2. The method according to claim 1, characterized in that, The spatiotemporal graph includes a node set V and an edge set E, the node set V... In each frame, a fighter jet is represented as a node, and T represents the total number of frames. This represents the node information of the enemy's i-th fighter jet in frame t. The node information includes the three-dimensional coordinates of the enemy's i-th fighter jet, as well as the distance between the enemy's i-th fighter jet and our first fighter jet and the distance between the enemy's i-th fighter jet and our second fighter jet. The edge set E includes the intra-frame spatial edge set. and inter-frame temporal edge set The intra-frame spatial edge represents the positional connection between two enemy aircraft within the same frame, and the inter-frame temporal edge represents the temporal connection between the same enemy aircraft in two adjacent frames.

3. The method according to claim 1, characterized in that, The local branch is a temporal convolutional network, which includes one-dimensional convolution and dilated convolution. The global branch is a multi-layer stacked Transformer encoder, and each Transformer encoder includes a multi-head self-attention and feedforward network.

4. A beyond-line-of-sight, high-dynamic dual-aircraft formation cooperative tactical identification device, characterized in that, The device includes: The spatiotemporal graph construction module is used to acquire continuous frame data of the two-aircraft formations of both sides in the air combat process. Based on the three-dimensional spatial coordinates of each enemy aircraft and the dynamic distance between the enemy and friendly aircraft in the continuous frame data, a spatiotemporal graph that integrates intra-frame spatial connectivity and inter-frame temporal correlation is constructed. The spatiotemporal feature extraction module is used to confirm feature vectors based on the spatiotemporal graph, and input the feature vectors into a two-layer cascaded spatiotemporal feature extraction network for feature extraction. Each layer of the spatiotemporal feature extraction network sequentially performs adaptive spatial graph convolution and parallel temporal feature fusion. The adaptive spatial graph convolution is used to generate a sample-level dynamic adjacency matrix based on data-driven generation to extract the collaborative features of the dynamic space of the two-machine nodes within the frame. The parallel temporal feature fusion is used to extract local short-term maneuver detail features and global long-term features in parallel based on the collaborative features, and then fuse them. The tactical identification module is used to classify the features extracted by the spatiotemporal feature extraction network via a two-layer concatenated network and output the identification result of the cooperative tactical type executed by the dual-aircraft formation. The adaptive spatial graph convolution is specifically used for: The feature vectors are embedded using two preset embedding functions to obtain two embedding feature matrices; The similarity matrix is ​​obtained by multiplying the two embedded feature matrices. The similarity matrix Each element in Represents a node and nodes Similarity; The parameterized coefficients are learned through a gating mechanism. The parameterized coefficients Used to adjust the similarity matrix Importance at different layers; The similarity matrix With the parameterized coefficients Multiplication yields the sample-level dynamic adjacency matrix. ; The adaptive spatial graph convolution is implemented using the following formula: ， in, The feature vector is input to the adaptive spatial graph convolution. The output vector of the adaptive spatial graph convolution. The size is N is the batch size, C is the number of channels, T is the number of frames, V is the number of nodes, k is the number of layers, and K is the spatial graph convolution kernel size. It is a 1*1 convolution; The parallel temporal feature fusion includes a local branch and a global branch. The local branch is used to extract local short-term maneuver detail features, and the global branch is used to extract global long-term features. The output vector of the adaptive spatial graph convolution is input in parallel to the local branch and the global branch, and the output results of the local branch and the global branch are weighted and fused through a learnable weighted fusion mechanism.

5. The apparatus according to claim 4, characterized in that, The spatiotemporal graph includes a node set V and an edge set E, the node set V... In each frame, a fighter jet is represented as a node, and T represents the total number of frames. This represents the node information of the enemy's i-th fighter jet in frame t. The node information includes the three-dimensional coordinates of the enemy's i-th fighter jet, as well as the distance between the enemy's i-th fighter jet and our first fighter jet and the distance between the enemy's i-th fighter jet and our second fighter jet. The edge set E includes the intra-frame spatial edge set. and inter-frame temporal edge set The intra-frame spatial edge represents the positional connection between two enemy aircraft within the same frame, and the inter-frame temporal edge represents the temporal connection between the same enemy aircraft in two adjacent frames.

6. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the method as described in any one of claims 1-3.

7. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the method as described in any one of claims 1-3.

Images