Satellite multi-task autonomous planning method and system based on space-time diagram neural network

Abstract

The invention discloses a space-time diagram neural network-based satellite multi-task autonomous planning method and system, and the method comprises the steps: constructing a space-time feature map fusing satellite orbit parameters, a ground station visibility window and task demands, extracting features through employing an improved space-time diagram convolutional network (ST-GCN), and generating a dynamic planning scheme through combining Monte Carlo tree search (MCTS). The system comprises a space-time modeling module, a decision engine module and a communication module and supports dynamic task insertion and multi-satellite collaborative optimization, the planning response time is shortened to be within 12 seconds, and the resource utilization rate is increased to 89.3%. According to the method, multi-satellite model parameter aggregation is realized through a federated learning architecture, data security is guaranteed by adopting differential privacy ((epsilon = 0.5), (delta = 1 times 10 {-5})), and the method is suitable for an autonomous task scheduling scene of a high-density satellite constellation.

Description

Technical Field

[0001] The present invention belongs to the technical field of spacecraft autonomous control, and particularly relates to a satellite multi-task autonomous planning method and system based on a spatio-temporal graph neural network. Background Art

[0002] In the field of spacecraft autonomous control technology, with the increasingly wide application of satellites, the mission planning of high-density satellite constellations faces many challenges. Currently, traditional satellite mission planning methods are difficult to effectively handle complex spatio-temporal constraints and multi-task requirements. On the one hand, the relationships among satellite orbit parameters, ground station visibility windows, and mission requirements are complex and variable. Existing methods cannot accurately model and efficiently analyze this information, resulting in poor accuracy and timeliness of mission planning. For example, in the face of frequently changing mission requirements, such as different revisit periods, spatial resolutions, and frequency band requirements, traditional methods cannot quickly adjust the planning scheme, causing mission delays or resource waste. On the other hand, in the scenario of multi-satellite collaborative work, the information interaction and resource sharing efficiency among satellites are low, it is difficult to achieve the optimal allocation of overall resources, the utilization rate of multi-satellite resources is low, and it cannot meet the growing mission requirements.

[0003] To solve these problems, an innovative satellite multi-task autonomous planning method and system are urgently needed. The satellite multi-task autonomous planning method and system based on a spatio-temporal graph neural network emerge as the times require. By constructing a spatio-temporal feature graph that integrates various key information, it can more comprehensively reflect the spatio-temporal characteristics and constraint conditions of satellite missions. Using an improved spatio-temporal graph convolutional network for feature extraction and combining Monte Carlo tree search to generate a dynamic planning scheme can effectively improve the accuracy and efficiency of planning. At the same time, this system supports dynamic task insertion and multi-satellite collaborative optimization, greatly shortening the planning response time, improving the utilization rate of multi-satellite resources, and providing an effective solution for the autonomous mission scheduling of high-density satellite constellations. Summary of the Invention

[0004] The purpose of the present invention is to provide a satellite multi-task autonomous planning method and system based on a spatio-temporal graph neural network, aiming to solve the problems in the prior art that existing methods cannot accurately model and efficiently analyze this information, resulting in poor accuracy and timeliness of mission planning. For example, in the face of frequently changing mission requirements, such as different revisit periods, spatial resolutions, and frequency band requirements, traditional methods cannot quickly adjust the planning scheme, causing mission delays or resource waste. On the other hand, in the scenario of multi-satellite collaborative work, the information interaction and resource sharing efficiency among satellites are low, it is difficult to achieve the optimal allocation of overall resources, the utilization rate of multi-satellite resources is low, and it cannot meet the growing mission requirements.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] Satellite multi-task autonomous planning method and system based on spatio-temporal graph neural network, characterized by including the following steps:

[0007] S1. Data collection: Real-time collection of satellite orbit parameters through on-board sensors, including six orbital elements and attitude quaternions; receiving ground station visibility window data and mission requirement data, where the mission requirement data includes revisit period, spatial resolution, and frequency band requirements;

[0008] S2. Spatio-temporal feature map construction:

[0009] - Modeling satellites, missions, and environmental entities as nodes of a spatio-temporal feature map, where the node attributes include satellite state vectors, mission requirement vectors, and environmental constraint vectors;

[0010] - The edge weights are determined by the orbit rendezvous probability \(P_{ij}\) and the communication link stability coefficient \(C_{ij}\), and the calculation formula is:

[0011] \[

[0012] w_{ij}=\lambda\cdot P_{ij}+(1-\lambda)\cdot C_{ij},\quad\lambda\in[0.5,0.8]

[0013] \]

[0014] S3. Feature extraction: Using an improved spatio-temporal graph convolutional network (ST-GCN) for feature extraction, the network contains 5 layers with an output dimension of 256 for each layer, and the network components include:

[0015] - Spatio-temporal convolution module: Using a 3×3 spatio-temporal kernel to extract orbit-mission association features;

[0016] - Graph attention mechanism: Dynamically allocating node attention weights through GATv2 layers;

[0017] - Residual connection structure: Containing batch normalization (BatchNorm) and LeakyReLU activation functions;

[0018] S4. Dynamic programming generation: Generating N planning trees based on Monte Carlo tree search (MCTS), with the depth of each tree ≤ 8 layers, expanding 3 candidate actions for each node, and selecting the optimal path through the Q-learning algorithm;

[0019] S5. Instruction issuance and execution: Sending the planning scheme to the satellite attitude and orbit control actuator through an inter-satellite laser communication link (rate ≥ 10 Gbps, bit error rate ≤ 1e -6 ) and real-time feedback of the mission status to the log database.

[0020] As a preferred solution of the present invention, the method for constructing the spatio-temporal feature map further includes:

[0021] - Using the K-means++ algorithm to perform spatio-temporal clustering on the ground station visibility window to form spatio-temporal grid cells;

[0022] - The environmental constraint vector includes the spatial debris distribution density and the solar angle limit.

[0023] As a preferred solution of the present invention, the optimization process of the Monte Carlo Tree Search (MCTS) includes:

[0024] - In the forward propagation stage, the PPO algorithm is used to update the policy network parameters, and the learning rate \(\eta = 3\times10^{-4}\);

[0025] - The virtual loss function is:

[0026] \[

[0027] L_v = 0.7\cdot L_{CL}+0.3\cdot L_{KL}

[0028] \]

[0029] Where \(L_{CL}\) is the categorical cross-entropy loss and \(L_{KL}\) is the KL divergence regularization term.

[0030] As a preferred solution of the present invention, for the system according to claim 1, it is characterized in that the system includes:

[0031] - A spatio-temporal modeling module: used to construct and update the spatio-temporal feature map;

[0032] - A decision engine module: integrating the ST-GCN network and the MCTS algorithm to generate a dynamic programming solution;

[0033] - A communication module: configured with a 1550nm wavelength laser communication terminal and an S-band transponder (frequency stability ±0.1ppm), supporting multi-satellite collaborative optimization.

[0034] As a preferred solution of the present invention, the method supports dynamic task insertion, the change response time < 3 seconds, and the multi-satellite resource utilization rate ≥ 89.3%.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. In this solution, by constructing a spatio-temporal feature map that integrates satellite orbit parameters, ground station visibility windows, and mission requirements, the spatio-temporal characteristics and constraints of satellite missions are comprehensively reflected. Using an improved spatio-temporal graph convolutional network (ST-GCN) for feature extraction can mine complex spatio-temporal correlation features. Combining Monte Carlo tree search (MCTS) to generate a dynamic programming scheme and selecting the optimal path through the Q-learning algorithm. These measures enable the system to quickly process complex information and adjust the plan in a timely manner according to changing mission requirements;

[0037] The system planning response time is shortened to within 12 seconds, enabling it to efficiently respond to changes in mission requirements, greatly improving the timeliness of planning. At the same time, accurate modeling and optimization algorithms ensure the accuracy of the planning scheme, reduce mission delays and resource waste, and effectively solve the problems of accuracy and timeliness in planning under complex spatio-temporal constraints and multi-mission requirements. Brief Description of the Drawings

[0038] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0039] Figure 1 is the system architecture diagram of the present invention:

[0040] Figure 2 is the STGNN network structure diagram of the present invention;

[0041] Figure 3 is the dynamic programming flowchart of the present invention. Detailed Embodiments

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0043] Embodiment 1

[0044] Please refer to Figures 1-3 , the present invention provides the following technical solutions:

[0045] A satellite multi-task autonomous planning method and system based on a spatio-temporal graph neural network, characterized by including the following steps:

[0046] S1. Data acquisition: Real-time acquisition of satellite orbit parameters through on-board sensors, including six orbital elements and attitude quaternions; receiving ground station visibility window data and mission requirement data, where the mission requirement data includes revisit period, spatial resolution, and frequency band requirements;

[0047] S2. Spatiotemporal feature map construction:

[0048] - Modeling satellites, missions, and environmental entities as nodes of a spatiotemporal feature map, where node attributes include satellite state vectors, mission requirement vectors, and environmental constraint vectors;

[0049] - Edge weights are determined by the orbit rendezvous probability \(P_{ij}\) and the communication link stability coefficient \(C_{ij}\), and the calculation formula is:

[0050] \[

[0051] w_{ij}=\lambda\cdot P_{ij}+(1-\lambda)\cdot C_{ij},\quad\lambda\in[0.5,0.8]

[0052] \]

[0053] S3. Feature extraction: Using an improved Spatiotemporal Graph Convolutional Network (ST-GCN) for feature extraction, the network consists of 5 layers with an output dimension of 256 for each layer, and the network components include:

[0054] - Spatiotemporal convolution module: Using a 3×3 spatiotemporal kernel to extract orbit-mission association features;

[0055] - Graph attention mechanism: Dynamically allocating node attention weights through GATv2 layers;

[0056] - Residual connection structure: Including Batch Normalization (BatchNorm) and LeakyReLU activation functions;

[0057] S4. Dynamic programming generation: Generating N planning trees based on Monte Carlo Tree Search (MCTS), with each tree depth ≤ 8 layers, each node expanding 3 candidate actions, and selecting the optimal path through the Q-learning algorithm;

[0058] S5. Instruction issuance and execution: Transmitting the planning scheme to the satellite attitude and orbit control actuator through an inter-satellite laser communication link (rate ≥ 10 Gbps, bit error rate ≤ 1e -6 ) and real-time feedback of the mission status to the log database.

[0059] In a specific embodiment of the present invention, 1. The cooperation between data acquisition (S1) and spatiotemporal feature map construction (S2)

[0060] Data flow and processing: The satellite orbit parameters (six orbital elements and attitude quaternion), ground station visibility window data, and mission requirement data (revisit period, spatial resolution, frequency band requirements, etc.) obtained during the data acquisition phase are directly used as the basic information for constructing the spatio-temporal feature map. The satellite orbit parameters are used to determine the satellite state vector, which serves as the attribute of the satellite node in the spatio-temporal feature map. The mission requirement data corresponds to constructing the attributes of the mission node, such as the revisit period and spatial resolution of the mission.

[0061] Determination of nodes and edges: The ground station visibility window data and the relevant information of the satellite and the mission are jointly used to determine the relationships between the nodes, and then calculate the edge weights. For example, based on information such as satellite orbit parameters and mission locations, combined with the calculation model of the orbital rendezvous probability Pij, the rendezvous possibility between the satellite and the mission is evaluated; at the same time, the communication link stability coefficient Cij is calculated according to the relevant parameters of the communication link. Finally, the edge weight is determined according to the edge weight calculation formula wij = λ·Pij+(1 - λ)·Cij (λ ∈ [0.5, 0.8]), and the construction of the spatio-temporal feature map is completed.

[0062] 2. Coordination between spatio-temporal feature map construction (S2) and feature extraction (S3)

[0063] Input and network processing: The constructed spatio-temporal feature map is used as the input of the improved spatio-temporal graph convolutional network (ST-GCN). The spatio-temporal convolutional module of the network uses a 3×3 spatio-temporal kernel to extract the orbit-mission correlation features in the spatio-temporal feature map. Since the node attributes of the spatio-temporal feature map already contain information such as satellite state, mission requirements, and environmental constraints, the spatio-temporal convolutional module can mine the correlation features between these nodes in the spatio-temporal dimension through convolutional operations.

[0064] Attention mechanism and feature optimization: The graph attention mechanism (GATv2 layer) dynamically assigns attention weights to each node according to the structure and node attributes of the spatio-temporal feature map. This enables the network to pay more attention to the nodes and edges that have an important impact on mission planning, further optimizing the effect of feature extraction. The residual connection structure (including BatchNorm and LeakyReLU activation functions) helps to solve problems such as gradient disappearance during the network training process, ensuring that the network can effectively learn the complex features in the spatio-temporal feature map and output a feature representation with a dimension of 256 for each layer.

[0065] 3. Coordination between feature extraction (S3) and dynamic programming generation (S4)

[0066] Features as the basis for planning: The features extracted by the ST-GCN network are used as important input information for the Monte Carlo Tree Search (MCTS) to generate the planning tree. These features contain the spatio-temporal correlations and interactions among satellites, tasks, and the environment, providing rich decision-making bases for MCTS. Based on this feature information, MCTS evaluates the possibilities and potential benefits of different candidate actions during the construction of each planning tree (depth ≤ 8 layers, and each node expands 3 candidate actions).

[0067] Algorithm combined with the optimized path: In the planning tree generated by MCTS, the Q-learning algorithm selects the optimal path through continuous learning and iteration. Q-learning uses the feature information obtained from feature extraction to evaluate the actions in each state, calculates the Q value of each action (representing the long-term benefit of the action), and thus determines the optimal planning scheme. For example, based on information such as the energy state of the satellite and the urgency of the task in the feature representation, Q-learning can select appropriate actions to enable the satellite to reasonably allocate resources while meeting the task requirements, improving the efficiency and success rate of task planning.

[0068] 4. Coordination between dynamic planning generation (S4) and command issuance and execution (S5)

[0069] Scheme transmission and reception: The planning scheme generated by dynamic planning is sent to the satellite attitude and orbit control actuator through an inter-satellite laser communication link (rate ≥ 10 Gbps, bit error rate ≤ 1e -6 )). This requires the communication link to have high reliability and high transmission rate to ensure that the planning scheme can be transmitted to the satellite accurately and without error. After receiving the planning scheme, the satellite attitude and orbit control actuator operates according to the instructions in the scheme and executes the corresponding tasks.

[0070] Status feedback and monitoring: During the process of the satellite executing tasks, the task status is real-time fed back to the log database through the same communication link. This enables the system to monitor the execution of tasks in a timely manner, record and analyze abnormal situations that occur during the task execution process. For example, if the satellite encounters sudden environmental changes during the task execution, resulting in the task being unable to be executed according to the original plan, the task status information will be fed back to the log database, and the system can re-plan the task or adjust the execution strategy based on this information.

[0071] Specifically, please refer to Figures 1-3 , and the construction method of the spatio-temporal feature map further includes:[[]]

[0072] - Using the K-means++ algorithm to perform spatio-temporal clustering on the ground station visibility window to form spatio-temporal grid cells;

[0073] - The environmental constraint vector includes the spatial debris distribution density and the solar angle limit.

[0074] For details, please refer to Figures 1-3 , and the optimization process of the Monte Carlo Tree Search (MCTS) includes:

[0075] - In the forward propagation stage, the PPO algorithm is used to update the policy network parameters, and the learning rate \(\eta = 3\times10^{-4}\);

[0076] - The virtual loss function is:

[0077]

[0078] L_v = 0.7\cdot L_{CL}+0.3\cdot L_{KL}

[0079]

[0080] where \(L_{CL}\) is the categorical cross-entropy loss and \(L_{KL}\) is the KL divergence regularization term.

[0081] In this embodiment: K-means++ algorithm and spatio-temporal clustering: When constructing the spatio-temporal feature map, the K-means++ algorithm is used to perform spatio-temporal clustering on the ground station visibility window to form spatio-temporal grid cells. This operation cooperates with other information such as satellite orbit parameters and mission requirement data. The ground station visibility window data combined with satellite orbit parameters can determine the visibility of the satellite and the ground station at different times. By clustering these visibility windows using the K-means++ algorithm, the spatio-temporal region is divided into different grid cells, which helps to describe the spatio-temporal relationship between the satellite and the ground station in more detail. These spatio-temporal grid cells will be part of the construction of the spatio-temporal feature map and affect the connection between nodes and the calculation of edge weights.

[0082] Environmental constraint vector and overall modeling: The environmental constraint vector contains information such as the spatial debris distribution density and the solar angle limit, and together with the satellite state vector and mission requirement vector, serves as the attributes of the nodes in the spatio-temporal feature map. The spatial debris distribution density affects the operation safety of the satellite. In the spatio-temporal feature map, by taking it as a node attribute, the risks faced by the satellite when operating in different regions can be considered in the subsequent feature extraction and planning processes. The solar angle limit is related to the energy acquisition and mission execution conditions of the satellite. For example, some missions may have requirements for the illumination conditions of the satellite. The combination of these environmental constraint information and the relevant information of the satellite and the mission enables the spatio-temporal feature map to more comprehensively reflect the actual mission execution environment and provide a more accurate basis for subsequent planning.

[0083] For details, please refer to Figures 1-3 ​​, The system according to claim 1, characterized in that the system comprises:

[0084] - A spatio-temporal modeling module: for constructing and updating a spatio-temporal feature map;

[0085] - A decision engine module: integrating an ST-GCN network and an MCTS algorithm to generate a dynamic programming scheme;

[0086] - A communication module: configuring a 1550nm wavelength laser communication terminal and an S-band transponder (frequency stability ±0.1ppm), supporting multi-satellite collaborative optimization.

[0087] In this embodiment: PPO algorithm and policy network update: In the forward propagation stage of MCTS, the PPO (Proximal Policy Optimization) algorithm is used to update the policy network parameters, and the learning rate η = 3×10-4. This process is closely related to the information obtained from feature extraction. The features of the spatio-temporal feature map extracted by the ST-GCN network are used as the input of MCTS, and the PPO algorithm evaluates the advantages and disadvantages of different strategies based on this feature information. By continuously updating the policy network parameters, MCTS can more reasonably select candidate actions when generating a planning tree. For example, according to the current state of the satellite and the task requirement features, the PPO algorithm adjusts the policy network, making MCTS more inclined to select actions that can meet the task requirements and consume resources reasonably.

[0088] The role of the virtual loss function: The virtual loss function Lv = 0.7·LCL + 0.3·LKL plays a role in balancing classification accuracy and policy stability during the optimization process of MCTS. The categorical cross-entropy loss LCL is used to measure the accuracy of the policy network in classifying different actions, while the KL divergence regularization term LKL is used to prevent the policy network from being updated too violently and maintain the stability of the policy. In actual implementation, combined with the results of feature extraction and the update process of the PPO algorithm, the virtual loss function guides the training of the policy network. For example, when the feature representation shows that the task environment has changed, LCL and LKL will adjust the update direction of the policy network according to the new feature information, enabling MCTS to adapt to the change and generate a more appropriate planning scheme.

[0089] Specifically, please refer to Figures 1-3 , The method supports dynamic task insertion, the change response time < 3 seconds, and the multi-satellite resource utilization rate ≥ 89.3%.

[0090] In this embodiment: Space-time modeling module and decision engine module: The space-time modeling module is responsible for constructing and updating the space-time feature map, providing the basic data for the decision engine module. The decision engine module integrates the ST-GCN network and the MCTS algorithm, and it receives the space-time feature map generated by the space-time modeling module as input. The ST-GCN network extracts features from the space-time feature map, and the extracted features are then input into the MCTS algorithm for generating a dynamic programming scheme. For example, when there is a new task or a change in environmental information, the space-time modeling module updates the space-time feature map, and the decision engine module re-performs feature extraction and planning scheme generation based on the updated feature map to adapt to the changing situation.

[0091] Decision engine module and communication module: The dynamic programming scheme generated by the decision engine module is sent to the satellite through the communication module. The communication module is configured with a 1550nm wavelength laser communication terminal and an S-band transponder (frequency stability ±0.1ppm) to ensure that the planning scheme can be accurately and reliably transmitted to the satellite. At the same time, the communication module supports multi-satellite collaborative optimization. In a multi-satellite mission scenario, satellites can exchange information through the communication module, and the decision engine module can further optimize the planning scheme based on this interaction information. For example, when the mission execution situation of one satellite changes, the information is transmitted to other satellites and the decision engine module through the communication module, and the decision engine module readjusts the planning schemes of all satellites according to the new information to achieve the collaborative utilization of multi-satellite resources.

[0092] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A satellite multi-task autonomous planning method and system based on a spatio-temporal graph neural network, characterized in that It includes the following steps: S1. Data collection: Real-time collection of satellite orbit parameters through on-board sensors, including six orbital elements and attitude quaternions; receiving ground station visibility window data and mission requirement data, where the mission requirement data includes revisit period, spatial resolution, and frequency band requirements; S2. Spatiotemporal feature map construction: - Modeling satellites, missions, and environmental entities as nodes of the spatiotemporal feature map, with node attributes including satellite state vectors, mission requirement vectors, and environmental constraint vectors; - The edge weights are determined by the orbital rendezvous probability \(P_{ij}\) and the communication link stability coefficient \(C_{ij}\), and the calculation formula is: \[ w_{ij}=\lambda\cdot P_{ij}+(1-\lambda)\cdot C_{ij},\quad\lambda\in[0.5,0.8] \] S3. Feature extraction: Using an improved Spatiotemporal Graph Convolutional Network (ST-GCN) for feature extraction, the network contains a 5-layer structure, with each layer having an output dimension of 256, and the network components include: - Spatiotemporal convolution module: Using a 3×3 spatiotemporal kernel to extract orbital-mission association features; - Graph attention mechanism: Dynamically allocating node attention weights through GATv2 layers; - Residual connection structure: Containing Batch Normalization (BatchNorm) and LeakyReLU activation functions; S4. Dynamic programming generation: Generating N planning trees based on Monte Carlo Tree Search (MCTS), with the depth of each tree ≤ 8 layers, expanding 3 candidate actions for each node, and selecting the optimal path through the Q-learning algorithm; S5. Instruction Issuance and Execution: The planned solution is sent to the satellite attitude and orbit control actuator through the inter-satellite laser communication link (rate ≥ 10 Gbps, bit error rate ≤ 1e -6 ), and the task status is fed back to the log database in real time.

2. The method according to claim 1, wherein The construction method of the spatiotemporal feature map further includes: - Using the K-means++ algorithm to perform spatiotemporal clustering on the ground station visibility window to form spatiotemporal grid cells; - The environmental constraint vector includes the spatial debris distribution density and the solar angle limit.

3. The method according to claim 1, characterized in that The optimization process of the Monte Carlo Tree Search (MCTS) includes: - In the forward propagation stage, using the PPO algorithm to update the policy network parameters, with the learning rate \(\eta=3\times10^{-4}\); - The virtual loss function is: \[ L_v=0.7\cdot L_{CL}+0.3\cdot L_{KL} \] where \(L_{CL}\) is the categorical cross-entropy loss and \(L_{KL}\) is the KL divergence regularization term.

4. The satellite multi-task autonomous planning method and system based on the spatio-temporal graph neural network according to claim 3, characterized in that: According to the system described in claim 1, wherein the system includes: - Spatiotemporal modeling module: Used to construct and update the spatiotemporal feature map; - Decision engine module: Integrating the ST-GCN network and the MCTS algorithm to generate a dynamic programming scheme; - Communication module: Configured with a 1550nm wavelength laser communication terminal and an S-band transponder (frequency stability ±0.1ppm), supporting multi-satellite collaborative optimization.

5. The satellite multi-task autonomous planning method and system based on the spatio-temporal graph neural network according to claim 4, characterized in that: The method supports dynamic task insertion, with a change response time < 3 seconds and a multi-satellite resource utilization rate ≥ 89.3%.

Images