Unmanned aerial vehicle game decision-making system based on high-precision trajectory prediction

Abstract

The invention discloses an unmanned aerial vehicle game decision-making system based on high-precision trajectory prediction. The unmanned aerial vehicle game decision-making system comprises a data acquisition subsystem, a flight trajectory database and an upper computer subsystem, the data acquisition subsystem acquires flight data of the unmanned aerial vehicle in real time through a positioning radar, an infrared sensor and a scanning gyroscope, and then performs graph structuring processing on the flight data to obtain the flight data after graph structuring processing and stores the flight data in a flight track database; the flight path database is used for providing flight data subjected to graph structured processing for the upper computer subsystem; the upper computer subsystem is used for outputting flight data at the next moment corresponding to the flight data subjected to graph structured processing, and then visually presenting and storing the flight data in a flight track database; the upper computer subsystem comprises a data set preprocessing module, a VGAE embedding generation module, a graph neural network modeling module, a track prediction module and a result display module which are connected in sequence. According to the method, graph structuring and graph embedding representation are innovatively introduced, the multi-level dependency relationship in an unmanned aerial vehicle cluster is fully captured, the precision and applicability of unmanned aerial vehicle game decision flight trajectory prediction are remarkably improved, and finally, a game decision mechanism is combined, so that the unmanned aerial vehicle game decision flight trajectory prediction precision is improved. And an efficient and reliable game decision-making solution is provided for a cooperative task of unmanned aerial vehicle cluster game decision-making.

Description

Technical Field

[0001] The present invention belongs to the fields of unmanned aerial vehicle game decision-making, graph structure learning and data-driven, and in particular relates to a unmanned aerial vehicle game decision-making system based on high-precision trajectory prediction. Background Art

[0002] With the widespread application of drones in logistics, monitoring and other fields, the demand for system automation control and efficient decision-making is increasing. In a complex and changing environment, different participants (such as drones, intelligent agents, etc.) need to make real-time strategic decisions during the game to optimize their own benefits and complete tasks collaboratively. Traditional decision-making methods usually rely on preset rules or physical models, which are difficult to deal with the complex dependencies in multi-agent collaborative games, and the decision-making accuracy and efficiency are low when facing a dynamically changing environment.

[0003] In recent years, with the rapid development of deep learning and data-driven technologies, data-based game decision-making methods have gradually attracted attention. In particular, graph neural networks can effectively capture the spatiotemporal dependencies between multiple agents by modeling the interactions between game participants. The state data of game participants and their interactions can be naturally modeled as a graph structure, where the nodes in the graph represent the strategic state of each agent at different time steps, and the edges represent the mutual game relationship between the participants. However, traditional graph neural networks have certain limitations when processing high-dimensional game data, and it is difficult to efficiently extract potential strategy features, especially when dealing with the timing characteristics and dynamic changes in the game.

[0004] To solve this problem, variational autoencoders provide a new method to extract potential strategy features from complex high-dimensional game data. By embedding the agent's temporal strategy data in a low-dimensional manner and combining it with graph neural networks to model the game relationship and strategy evolution between multiple agents, the accuracy and real-time performance of decision-making can be significantly improved. Therefore, the game decision-making system based on variational graph autoencoders and graph neural networks has become an effective solution that can handle complex temporal relationships in multi-agent collaborative games, achieve accurate predictions of future strategy evolution, and adapt to decision-making needs in complex dynamic scenarios. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the prior art and provide a drone game decision-making system based on high-precision trajectory prediction.

[0006] The objective of the present invention is achieved through the following technical solutions: a UAV game decision system based on high-precision trajectory prediction, the UAV game decision system comprising: a data acquisition subsystem, a flight trajectory database and a host computer subsystem; the data acquisition subsystem collects the flight data of the UAV in real time through a positioning radar, an infrared sensor and a scanning gyroscope, and then performs graph-structured processing on the flight data to obtain the flight data after graph-structured processing and stores it in the flight trajectory database; the flight trajectory database is used to provide the host computer subsystem with the flight data after graph-structured processing; the host computer subsystem is used to output the flight data of the next moment corresponding to the flight data after graph-structured processing, and then performs visual presentation and stores it in the flight trajectory database; the host computer subsystem comprises a data set preprocessing module, a VGAE embedding generation module, a graph neural network modeling module, a trajectory prediction module and a result display module connected in sequence.

[0007] Furthermore, the data acquisition subsystem collects the flight data of the UAV in real time through positioning radar, infrared sensor and scanning gyroscope, specifically:

[0008] The data acquisition subsystem collects the flight status of the UAV through positioning radar, infrared sensor and scanning gyroscope, and obtains the flight data H of the UAV at time step t t : Where N represents the flight data H t Contains the flight feature vectors of N UAVs; represents the flight feature vector of the i-th UAV at time step t, i=1,2,…,i,…,N;

[0009] The flight characteristics vector for Among them, P i (t) represents the coordinate position vector of the positioning radar of the ith UAV at time step t; v i (t) represents the flight speed vector of the ith UAV at time step t; a i (t) represents the aircraft attitude vector of the i-th UAV at time step t;

[0010] Coordinate position vector P i (t) is P i (t) = [x i (t),y i (t),z i (t)], where x i (t), y i (t) and z i (t) represents the three-dimensional coordinates of the i-th UAV at time step t;

[0011] Flight velocity vector v i (t) is v i (t) = [v x,i (t),v y,i (t),v z,i (t)], where v x,i (t) represents the velocity of the ith UAV on the x-axis at time step t, v y,i (t) represents the velocity of the ith UAV on the y-axis at time step t, v z,i (t) represents the velocity of the i-th UAV on the z-axis at time step t;

[0012] Aircraft attitude vector a i (t) is Among them, ω i (t) represents the roll angle of the i-th UAV at time step t, θ i (t) represents the pitch angle of the i-th UAV at time step t, represents the yaw angle of the i-th UAV at time step t.

[0013] Furthermore, the data acquisition subsystem performs graph-structured processing on the flight data, obtains the graph-structured flight data and stores it in the flight trajectory database, which specifically includes the following sub-steps:

[0014] (a.1) The data acquisition subsystem first collects the flight data H t Construct edge matrix E t : Among them, e (i,t),(j,t) represents the trajectory dependency between the i-th UAV and the j-th UAV at the same time step t, j = 1, 2, …, j, …, N;

[0015] Then according to the matrix E t Get the weight matrix W t :

[0016] Among them, w (i,t),(j,t) represents the weight between the i-th UAV and the j-th UAV at the same time step t;

[0017] If the trajectory dependency between the i-th UAV and the j-th UAV at the same time step t exists, then the edge e (i,t),(j,t) Weight setting w (i,t),(j,t) is 1, otherwise the edge e (i,t),(j,t) The weight setting w (i,t),(j,t) is 0, the judgment formula for whether there is a dependency relationship is as follows:

[0018]

[0019] Among them, ||·|| 2 represents norm2 function; ∈ is the preset threshold;

[0020] (a.2) Then according to the edge matrix E t and the weight matrix W t , get the adjacency matrix A t :

[0021] Among them, a (i,t),(j,t) It represents the adjacency relationship between the i-th UAV and the j-th UAV at the same time step t. The calculation formula is as follows:

[0022]

[0023] (a.3) The flight data H t and the adjacency matrix A t The flight data G after graph structure processing t =(H t ,A t ) is stored in the flight trajectory database.

[0024] Furthermore, the host computer subsystem outputs the flight data at the next moment corresponding to the flight data after the graph structured processing, and then performs visual presentation and stores it in the flight trajectory database, which specifically includes the following sub-steps:

[0025] (c.1) The host computer subsystem first converts the flight data G after graph structure processing t =(H t ,A t ) is input into the data set preprocessing module for preprocessing, and the normalized graph-structured flight data G is output. t ′=(H t ′,A t ′);

[0026] (c.2) The normalized graph-structured flight data G t ′=(H t ′,A t ′) is input into the VGAE embedding generation module, and the embedding generation model in the VGAE embedding generation module is optimized and trained to obtain the optimized embedding generation model; and the normalized graph-structured flight data G is trained by the optimized embedding generation model. t ′=(H t ′,A t ′) to obtain the optimized potential embedding matrix Z t And the optimized reconstructed adjacency matrix

[0027] (c.3) Then the optimized latent embedding matrix And the optimized reconstructed adjacency matrix Input to the graph neural network modeling module, and output the node feature matrix by updating the graph neural network parameters of the L layer

[0028] (c.4) Then the node feature matrix Input to the game decision trajectory prediction module, and the nonlinear neural network layer predicts the flight characteristic data of the next moment

[0029] (c.5) The flight characteristics data of the next moment will be predicted The game decision result display module visualizes the result and stores it in the flight trajectory database.

[0030] Furthermore, the step (c.1) specifically includes the following sub-steps:

[0031] (c.1.1) Flight data G after graph structure processing t =(H t ,A t ) in the flight data H t Normalize and get the normalized flight data H t ′: H t ′=[h t ' 1 ,h t ' 2 ,…,h t ' i ,…,h t ' N ];

[0032] For each UAV’s flight feature vector at time step t Normalize and get the normalized flight feature vector h t ' i :

[0033]

[0034] in, Represents the flight feature vector The mean of all features in ; Represents the flight feature vector The standard deviation of all features in ;

[0035] (c.1.2) Flight data G after graph structure processing t =(H t ,A t ) in the adjacency matrix At Normalize and get the normalized adjacency matrix A t ′: Among them, D t Represents the adjacency matrix A t The degree matrix of

[0036] After normalization, the normalized graph-structured flight data G is obtained. t ′=(H t ′,A t ′).

[0037] Furthermore, the step (c.2) specifically includes the following sub-steps:

[0038] (c.2.1) The normalized graph-structured flight data G t ′=(H t ′,A t ′) is input into a VGAE embedding generation module, wherein the VGAE embedding generation module comprises an embedding generation model composed of a decoder and an encoder;

[0039] Normalized flight data H t Each flight feature vector h in ′ t ' i After being processed by the encoder, the mean μ is obtained t ' i and variance σ t ' i :

[0040] μ t ' i ,log(σ t ' i ) 2 =Encoder(h t ' i );

[0041] Where, Encoder(·) represents the encoder;

[0042] Then the mean μ t ' i and variance σ t ' i Construct the flight feature vector h t ' i The corresponding latent embedding vector

[0043] For all flight feature vectors h t ' i After the above processing, we get the potential embedding matrix Z t :

[0044] (c.2.2) Use the decoder to pass the latent embedding matrix Z t Construct the reconstructed adjacency matrix

[0045] in, represents the potential similarity between the i-th UAV and the j-th UAV at the same time step t, and the calculation formula is as follows:

[0046]

[0047] Where Decoder(·) represents a decoder, through which the potential similarity is calculated using an inner product operation or a nonlinear activation function;

[0048] (c.2.3) Constructing a total loss function L to optimize the model parameters of the embedding generation model, and obtaining an optimized embedding generation model;

[0049] The total loss function L is composed of the reconstruction error loss L recon And KL divergence loss function L KL composition;

[0050] The reconstruction error loss L recon for Among them, a′ (i,t),(j,t) Represents the normalized adjacency matrix A t Any adjacency relation in ′;

[0051] The KL divergence loss function L KL for Where Tr(·) represents the trace function of the matrix, i.e., the sum of the diagonal elements of the variance matrix; det(·) represents the determinant of the matrix; d is the dimension of the potential embedding vector;

[0052] (c.2.4) Then the normalized graph-structured flight data G t ′=(H t ′,A t ′) is input into the optimized embedding generation model, and steps (c.2.1)-step (c.2.2) are repeated to obtain the optimized potential embedding matrix And the optimized reconstructed adjacency matrix

[0053] Furthermore, the step (c.3) specifically includes the following sub-steps:

[0054] (c.3.1) The optimized latent embedding matrix for in, represents the optimized latent embedding vector of the i-th drone;

[0055] The optimized reconstructed adjacency matrix for in, represents the optimized potential similarity between the i-th UAV and the j-th UAV at the same time step t;

[0056] (c.3.2) Get the set of all adjacent drones of the i-th drone at the same time step t in, represents the neighboring drones of any i-th drone at the same time step t, E i represents the number of all neighboring drones of the i-th drone at the same time step t;

[0057] Optimized latent embedding vector for the i-th drone Update the parameters of the graph neural network of layer L;

[0058] First, any optimized latent embedding vector Node features as output of the first layer of graph neural network

[0059] Then the node features output by the l-th layer of graph neural network for Among them, W l is the learnable parameter matrix of the l-th layer graph neural network; σ(·) is the nonlinear Sigmoid activation function;

[0060] After updating the L-layer graph neural network, the node features output by the L-layer graph neural network are obtained.

[0061] (c.3.2) For the optimized latent embedding matrix Z t Repeat step (c.3.2) for each optimized latent embedding vector in and output the node feature matrix

[0062] The beneficial effects of the present invention are: 1. The variational graph autoencoder is innovatively introduced to embed and generate the UAV trajectory, effectively extract the potential features of the UAV's multi-dimensional time series data, and at the same time obtain the optimized reconstructed adjacency matrix, thereby enhancing the prediction accuracy of the system in complex game tasks; 2. The collaborative relationship and spatiotemporal dependency between UAVs are modeled through the graph neural network. The algorithm has a high degree of optimization and can handle complex scenarios of multi-agent collaborative games to achieve high-precision trajectory prediction; 3. The system can identify changes in multiple game strategies in real time, has strong intelligence, and can adapt to game decision-making needs in complex dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a structural diagram of a UAV game decision-making system based on high-precision trajectory prediction;

[0064] In the figure, 1-data acquisition subsystem; 2-flight trajectory database; 3-host computer subsystem; 4-dataset preprocessing module; 5-VGAE embedding generation module; 6-graph neural network modeling module; 7-trajectory prediction module; 8-result display module. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical scheme and advantages of the present invention more clear, the present invention is further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0066] Example 1

[0067] like Figure 1 As shown, the present invention provides a UAV game decision system based on high-precision trajectory prediction, and the UAV game decision system includes: a data acquisition subsystem 1, a flight trajectory database and a host computer subsystem; the data acquisition subsystem 1 collects the flight data of the UAV in real time through a positioning radar, an infrared sensor and a scanning gyroscope, and then performs graph-structured processing on the flight data to obtain the flight data after graph-structured processing and stores it in the flight trajectory database 2; the flight trajectory database 2 is used to provide the flight data after graph-structured processing for the host computer subsystem 3; the host computer subsystem 3 is used to output the flight data of the next moment corresponding to the flight data after graph-structured processing, and then visualizes it and stores it in the flight trajectory database 2; the host computer subsystem 3 includes a data set preprocessing module 4, a VGAE embedding generation module 5, a graph neural network modeling module 6, a trajectory prediction module 7 and a result display module 8 connected in sequence.

[0068] The data acquisition subsystem 1 collects the flight data of the UAV in real time through positioning radar, infrared sensor and scanning gyroscope, specifically:

[0069] The data acquisition subsystem 1 collects the flight status of the UAV through positioning radar, infrared sensor and scanning gyroscope, and obtains the flight data H of the UAV at time step t t : Where N represents the flight data H t Contains the flight feature vectors of N UAVs; represents the flight feature vector of the i-th UAV at time step t, i = 1, 2,…, i,…, N.

[0070] The flight characteristics vector for Among them, P i (t) represents the coordinate position vector of the positioning radar of the ith UAV at time step t; v i (t) represents the flight speed vector of the ith UAV at time step t; a i (t) represents the aircraft attitude vector of the i-th UAV at time step t.

[0071] Coordinate position vector P i (t) is P i (t) = [x i (t),y i (t),z i (t)], where x i (t), y i (t) and z i (t) represents the three-dimensional coordinates of the i-th UAV at time step t;

[0072] Flight velocity vector v i (t) is v i (t) = [v x,i (t),v y,i (t),v z,i (t)], where v x,i (t) represents the velocity of the ith UAV on the x-axis at time step t, v y,i (t) represents the velocity of the ith UAV on the y-axis at time step t, v z,i (t) represents the velocity of the i-th UAV on the z-axis at time step t.

[0073] Aircraft attitude vector a i (t) is Among them, ω i (t) represents the roll angle of the i-th UAV at time step t, θ i (t) represents the pitch angle of the i-th UAV at time step t, represents the yaw angle of the i-th UAV at time step t.

[0074] The data acquisition subsystem 1 performs graph-structured processing on the flight data, obtains the graph-structured flight data and stores it in the flight trajectory database 2, which specifically includes the following sub-steps:

[0075] (a.1) The data acquisition subsystem 1 first collects the flight data H t Construct edge matrix Et : Among them, e (i,t),(j,t) represents the trajectory dependency between the i-th UAV and the j-th UAV at the same time step t, j = 1, 2, …, j, …, N.

[0076] Then according to the matrix E t Get the weight matrix W t :

[0077] Among them, w (i,t),(j,t) represents the weight between the i-th UAV and the j-th UAV at the same time step t.

[0078] If the trajectory dependency between the i-th UAV and the j-th UAV at the same time step t exists, then the edge e (i,t),(j,t) Weight setting w (i,t),(j,t) is 1, otherwise the side e (i,t),(j,t) The weight setting w (i,t),(j,t) is 0, the formula is as follows:

[0079]

[0080] Among them, ||·|| 2 represents norm2 function; ∈ is the preset threshold.

[0081] (a.2) Then according to the edge matrix E t and the weight matrix W t , get the adjacency matrix A t :

[0082] Among them, a (i,t),(j,t) It represents the adjacency relationship between the i-th UAV and the j-th UAV at the same time step t. The calculation formula is as follows:

[0083]

[0084] (a.3) The flight data H t and the adjacency matrix A t The flight data G after graph structure processing t =(H t ,A t ) is stored in the flight trajectory database 2.

[0085] The host computer subsystem 3 outputs the flight data of the next moment corresponding to the flight data after the graph structure processing, and then visualizes and stores it in the flight trajectory database 2, which specifically includes the following sub-steps:

[0086] (c.1) The host computer subsystem 3 first converts the flight data G after graph structure processingt =(H t ,A t ) is input to the data set preprocessing module 4 for preprocessing, and the normalized graph-structured flight data G is output. t ′=(H t ′,A t ′).

[0087] The step (c.1) specifically includes the following sub-steps:

[0088] (c.1.1) In order to ensure the consistency of feature ranges and prevent data of different dimensions from affecting model training, the dataset preprocessing module performs normalization.

[0089] For the flight data G after graph structure processing t =(H t ,A t ) in the flight data H t Normalize and get the normalized flight data H t ′: H t ′=[h t ' 1 ,h t ' 2 ,…,h t ' i ,…,h t ' N ].

[0090] For each UAV’s flight feature vector at time step t Normalize and get the normalized flight feature vector h t ' i :

[0091]

[0092] in, Represents the flight feature vector The mean of all features in ; Represents the flight feature vector The standard deviation of all features in .

[0093] (c.1.2) is to make the adjacency matrix A t The interaction between nodes in the game can contribute fairly in the game optimization process. The dataset preprocessing module can make a fair contribution to the adjacency matrix A. t Normalize.

[0094] For the flight data G after graph structure processing t =(H t ,A t ) in the adjacency matrix At Normalize and get the normalized adjacency matrix A t ′: Among them, D t Represents the adjacency matrix A t The degree matrix of .

[0095] After normalization, the normalized graph-structured flight data G is obtained. t ′=(H t ′,A t ′).

[0096] (c.2) The normalized graph-structured flight data G t ′=(H t ′,A t ′) is input to the VGAE embedding generation module 5, and the embedding generation model in the VGAE embedding generation module 5 is optimized and trained to obtain the optimized embedding generation model; and the normalized graph-structured flight data G is trained by the optimized embedding generation model t ′=(H t ′,A t ′) to obtain the optimized potential embedding matrix And the optimized reconstructed adjacency matrix

[0097] The step (c.2) specifically includes the following sub-steps:

[0098] (c.2.1) The normalized graph-structured flight data G t ′=(H t ′,A t ′) is input into the VGAE embedding generation module 5, and the VGAE embedding generation module 5 includes an embedding generation model composed of a decoder and an encoder.

[0099] Normalized flight data H t Each flight feature vector h in ′ t ' i After being processed by the encoder, the mean μ is obtained t ' i and variance σ t ' i :

[0100] μ t ' i ,log(σ t ' i ) 2 =Encoder(h t ' i );

[0101] Wherein, Encoder(·) represents an encoder.

[0102] Then the mean μ t ' i and variance σ t ' i Construct the flight feature vector h t ' i The corresponding latent embedding vector ε is a noise vector sampled from a standard normal distribution, ensuring the randomness of the latent embedding vector.

[0103] For all flight feature vectors h t ' i After the above processing, we get the potential embedding matrix Z t :

[0104] (c.2.2) Use the decoder to pass the latent embedding matrix Z t Construct the reconstructed adjacency matrix

[0105] in, represents the potential similarity between the i-th UAV and the j-th UAV at the same time step t, and the calculation formula is as follows:

[0106]

[0107] Where Decoder(·) represents a decoder, and the potential similarity is calculated by the decoder using an inner product operation or a nonlinear activation function.

[0108] (c.2.3) Construct the total loss function L to optimize the model parameters of the embedding generation model to obtain the optimized embedding generation model.

[0109] The total loss function L is composed of the reconstruction error loss L recon And KL divergence loss function L KL composition.

[0110] The reconstruction error loss L recon for Among them, a′ (i,t),(j,t) Represents the normalized adjacency matrix A t Any adjacency relationship in ′.

[0111] The KL divergence loss function L KL for Where Tr(·) represents the trace function of the matrix, i.e., the sum of the diagonal elements of the variance matrix; det(·) represents the determinant of the matrix; and d is the dimension of the potential embedding vector.

[0112] (c.2.4) Then the normalized graph-structured flight data G t ′=(H t ′,A t ′) is input into the optimized embedding generation model, and steps (c.2.1)-step (c.2.2) are repeated to obtain the optimized potential embedding matrix And the optimized reconstructed adjacency matrix A t .

[0113] (c.3) Then the optimized latent embedding matrix And the optimized reconstructed adjacency matrix Input to the graph neural network modeling module 6, and output the node feature matrix by updating the graph neural network parameters of the L layer

[0114] The step (c.3) specifically includes the following sub-steps:

[0115] (c.3.1) The optimized latent embedding matrix for in, represents the optimized latent embedding vector of the i-th drone;

[0116] The optimized reconstructed adjacency matrix for in, represents the optimized potential similarity between the i-th UAV and the j-th UAV at the same time step t.

[0117] (c.3.2) Get the set of all adjacent drones of the i-th drone at the same time step t in, represents the neighboring drones of any i-th drone at the same time step t, E i represents the number of all neighboring drones of the ith drone at the same time step t.

[0118] Optimized latent embedding vector for the i-th drone Update the L-layer graph neural network.

[0119] First, any optimized latent embedding vector Node features as output of the first layer of graph neural network

[0120] Then the node features output by the l-th layer of graph neural network Among them, W lis the learnable parameter matrix of the l-th layer graph neural network; σ(·) is the nonlinear Sigmoid activation function.

[0121] After updating the L-layer graph neural network, the node features output by the L-layer graph neural network are obtained.

[0122] (c.3.2) For the optimized latent embedding matrix Repeat step (c.3.2) for each optimized latent embedding vector in and output the node feature matrix

[0123] (c.4) Then the node feature matrix Input to the game decision trajectory prediction module 7, and the nonlinear neural network layer predicts the flight characteristic data of the next moment

[0124] The flight characteristic data at the next moment includes the future coordinate position vector of each drone at the next moment t+1 Flight speed vector and the aircraft attitude vector Among them, f p (·),f v (·) and f a (·) are nonlinear functions used for predicting coordinate position, flight speed and aircraft attitude, respectively, which are usually implemented by neural networks.

[0125] (c.5) The predicted flight data for the next moment is then The game decision result display module 8 performs visual presentation and stores it in the flight trajectory database 2 for subsequent system evaluation, prediction and further game analysis.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A UAV game decision system based on high-precision trajectory prediction, characterized in that: The UAV game decision system includes: a data acquisition subsystem, a flight trajectory database and a host computer subsystem; the data acquisition subsystem collects the flight data of the UAV in real time through a positioning radar, an infrared sensor and a scanning gyroscope, and then performs graph-structured processing on the flight data to obtain the flight data after graph-structured processing and stores it in the flight trajectory database; the flight trajectory database is used to provide the host computer subsystem with the flight data after graph-structured processing; the host computer subsystem is used to output the flight data of the next moment corresponding to the flight data after graph-structured processing, and then visualizes it and stores it in the flight trajectory database; the host computer subsystem includes a data set preprocessing module, a VGAE embedding generation module, a graph neural network modeling module, a trajectory prediction module and a result display module that are connected in sequence.

2. The UAV game decision system based on high-precision trajectory prediction according to claim 1 is characterized in that: The data acquisition subsystem collects the flight data of the UAV in real time through positioning radar, infrared sensor and scanning gyroscope, specifically: The data acquisition subsystem collects the flight status of the UAV through positioning radar, infrared sensor and scanning gyroscope, and obtains the flight data H of the UAV at time step t t : Where N represents the flight data H t Contains the flight feature vectors of N UAVs; represents the flight feature vector of the i-th UAV at time step t, i=1,2,...,i,…,N; The flight characteristics vector for Among them, P i (t) represents the coordinate position vector of the positioning radar of the ith UAV at time step t; v i (t) represents the flight speed vector of the ith UAV at time step t; a i (t) represents the aircraft attitude vector of the i-th UAV at time step t; Coordinate position vector P i (t) is P i (t) = [x i (t),y i (t),z i (t)], where x i (t), y i (t) and z i (t) represents the three-dimensional coordinates of the i-th UAV at time step t; Flight velocity vector v i (t) is v i (t) = [v x,i (t),v y,i (t),v z,i (t)], where v x,i (t) represents the velocity of the ith UAV on the x-axis at time step t, v y,i (t) represents the velocity of the ith UAV on the y-axis at time step t, v z,i (t) represents the velocity of the i-th UAV on the z-axis at time step t; Aircraft attitude vector a i (t) is Among them, ω i (t) represents the roll angle of the i-th UAV at time step t, θ i (t) represents the pitch angle of the ith UAV at time step t, represents the yaw angle of the i-th UAV at time step t.

3. The unmanned aerial vehicle game decision system based on high-precision trajectory prediction according to claim 2 is characterized in that: The data acquisition subsystem performs graph-structured processing on the flight data, obtains the graph-structured flight data and stores it in the flight trajectory database, which specifically includes the following sub-steps: (a.1) The data acquisition subsystem first collects the flight data H t Construct edge matrix E t : Among them, e (i,t),(j,t) represents the trajectory dependency between the i-th UAV and the j-th UAV at the same time step t, j = 1, 2, …, j, …, N; Then according to the matrix E t Get the weight matrix W t : Among them, w (i,t),(j,t) represents the weight between the i-th UAV and the j-th UAV at the same time step t; If the trajectory dependency between the i-th UAV and the j-th UAV at the same time step t exists, then the edge e (i,t),(j,t) Weight setting w (i,t),(j,t) is 1, otherwise the edge e (i,t),(j,t) The weight setting w (i,t),(j,t) is 0, the judgment formula for whether there is a dependency relationship is as follows: Among them, ||·|| 2 represents norm2 function; ∈ is the preset threshold; (a.2) Then according to the edge matrix E t and the weight matrix W t , get the adjacency matrix A t : Among them, a (i,t),(j,t) It represents the adjacency relationship between the i-th UAV and the j-th UAV at the same time step t. The calculation formula is as follows: (a.3) The flight data H t and the adjacency matrix A t The flight data G after graph structure processing t =(H t ,A t ) is stored in the flight trajectory database.

4. The drone game decision system based on high-precision trajectory prediction according to claim 3 is characterized in that: The host computer subsystem outputs the flight data of the next moment corresponding to the flight data after the graph structured processing, and then visualizes and stores it in the flight trajectory database, which specifically includes the following sub-steps: (c.1) The host computer subsystem first converts the flight data G after graph structure processing t =(H t ,A t ) is input into the data set preprocessing module for preprocessing, and the normalized graph-structured flight data G′ is output t =(H′ t ,A′ t ); (c.2) The normalized graph-structured flight data G′ t =(H′ t ,A′ t ) is input into the VGAE embedding generation module, and the embedding generation model in the VGAE embedding generation module is optimized and trained to obtain the optimized embedding generation model; and the normalized graph-structured flight data G′ is trained by the optimized embedding generation model t =(H′ t ,A′ t ) to obtain the optimized potential embedding matrix And the optimized reconstructed adjacency matrix (c.3) Then the optimized latent embedding matrix And the optimized reconstructed adjacency matrix Input to the graph neural network modeling module, and output the node feature matrix by updating the graph neural network parameters of the L layer (c.4) Then the node feature matrix Input to the game decision trajectory prediction module, and the nonlinear neural network layer predicts the flight characteristic data of the next moment (c.5) The flight characteristics data of the next moment will be predicted The game decision result display module visualizes the result and stores it in the flight trajectory database.

5. The drone game decision system based on high-precision trajectory prediction according to claim 4 is characterized in that: The step (c.1) specifically includes the following sub-steps: (c.1.1) Flight data G after graph structure processing t =(H t ,A t ) in the flight data H t Normalize and get the normalized flight data H t ′: H t ′=[h t ' 1 ,h t ' 2 ,…,h t ' i ,…,h t ' N ]; For each UAV’s flight feature vector at time step t Normalize and get the normalized flight feature vector h t ' i : in, Represents the flight feature vector The mean of all features in ; Represents the flight feature vector The standard deviation of all features in ; (c.1.2) Flight data G after graph structure processing t =(H t ,A t ) in the adjacency matrix A t Normalize and get the normalized adjacency matrix A t ′: Among them, D t Represents the adjacency matrix A t The degree matrix of After normalization, the normalized graph-structured flight data G′ is obtained. t =(H′ t ,A′ t ).

6. The drone game decision system based on high-precision trajectory prediction according to claim 5 is characterized in that: The step (c.2) specifically includes the following sub-steps: (c.2.1) The normalized graph-structured flight data G′ t =(H′ t ,A′ t ) is input into a VGAE embedding generation module, wherein the VGAE embedding generation module includes an embedding generation model composed of a decoder and an encoder; Normalized flight data H t Each flight feature vector h in ′ t ' i After being processed by the encoder, the mean μ is obtained t ' i and variance σ t ' i : m t ′ i ,log(σ t ′ i ) 2 =Encoder(h t ′ i ); Where, Encoder(·) represents the encoder; Then the mean μ t ' i and variance σ t ' i Construct the flight feature vector h t ' i The corresponding latent embedding vector For all flight feature vectors h t ' i After the above processing, we get the potential embedding matrix Z t : (c.2.2) Use the decoder to pass the latent embedding matrix Z t Construct the reconstructed adjacency matrix in, represents the potential similarity between the i-th UAV and the j-th UAV at the same time step t, and the calculation formula is as follows: Where Decoder(·) represents a decoder, through which the potential similarity is calculated using an inner product operation or a nonlinear activation function; (c.2.3) Constructing a total loss function L to optimize the model parameters of the embedding generation model, and obtaining an optimized embedding generation model; The total loss function L is composed of the reconstruction error loss L recon And KL divergence loss function L KL composition; The reconstruction error loss L recon for Among them, a ( ' i,t),(j,t) Represents the normalized adjacency matrix A t Any adjacency relation in ′; The KL divergence loss function L KL for Where Tr(·) represents the trace function of the matrix, i.e., the sum of the diagonal elements of the variance matrix; det(·) represents the determinant of the matrix; d is the dimension of the potential embedding vector; (c.2.4) Then the normalized graph-structured flight data G′ t =(H′ t ,A′ t ) is input into the optimized embedding generation model, and steps (c.2.1)-step (c.2.2) are repeated to obtain the optimized potential embedding matrix And the optimized reconstructed adjacency matrix 7. The drone game decision system based on high-precision trajectory prediction according to claim 6 is characterized in that: The step (c.3) specifically includes the following sub-steps: (c.3.1) The optimized latent embedding matrix for in, represents the optimized latent embedding vector of the i-th drone; The optimized reconstructed adjacency matrix for in, represents the optimized potential similarity between the i-th UAV and the j-th UAV at the same time step t; (c.3.2) Get the set of all adjacent drones of the i-th drone at the same time step t in, represents the neighboring drones of any i-th drone at the same time step t, E i represents the number of all neighboring drones of the i-th drone at the same time step t; Optimized latent embedding vector for the i-th drone Update the parameters of the graph neural network of layer L; First, any optimized latent embedding vector Node features as output of the first layer of graph neural network Then the node features output by the l-th layer of graph neural network for Among them, W l is the learnable parameter matrix of the l-th layer graph neural network; σ(·) is the nonlinear Sigmoid activation function; After updating the L-layer graph neural network, the node features output by the L-layer graph neural network are obtained. (c.3.2) For the optimized latent embedding matrix Repeat step (c.3.2) for each optimized latent embedding vector in and output the node feature matrix

Images