Large-scale satellite task dynamic constraint conflict resolution method and system based on heterogeneous graph learning

Abstract

The invention provides a large-scale satellite task dynamic constraint conflict resolution method and system based on heterogeneous graph learning, and the method comprises the steps: carrying out the mathematical description of a task planning problem, and designing an optimization objective function; carrying out characteristic analysis and mathematical expression on the constraint; constructing a satellite task static constraint conflict graph model for the static constraint; constructing a satellite task dynamic constraint conflict graph model for the dynamic constraint; solving the static constraint conflict graph model; solving the dynamic constraint conflict graph model to obtain a satellite task selection probability; and performing decision optimization on the satellite task to finally obtain a satellite task planning result. According to the method, efficient scheduling and allocation decision-making of the multi-satellite large-scale intensive tasks are realized, the problem of efficient calculation of constraint investigation and conflict resolution in the multi-satellite task planning process under the intensive tasks is solved, and the conflict resolution frequency in the multi-satellite task planning solving process is greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing satellite mission planning, and in particular relates to a large-scale satellite mission dynamic constraint conflict resolution method and system based on heterogeneous graph learning. Background Art

[0002] Remote sensing satellite mission planning is widely considered to be a non-deterministic polynomial (NP) hard problem with complex constraints. In order to solve this problem, domestic and foreign scholars have conducted a lot of research.

[0003] Exact algorithms and heuristic algorithms for solving satellite mission planning problems: In the early days of solving multi-satellite mission planning problems, the algorithms commonly used were usually divided into exact algorithms and heuristic algorithms. Exact methods, such as branch and bound and mixed integer linear programming (MILP), have deterministic solution steps and can solve problems in polynomial time in most cases; however, for larger-scale problems, it is actually impossible to establish a deterministic model. Heuristic methods can speed up the solution process by finding a satisfactory solution, but these methods usually rely on specific heuristic strategies, which are effective for optimizing small-scale problems, but difficult to solve large-scale problems.

[0004] Metaheuristic algorithms for solving satellite mission planning problems: Metaheuristic algorithms guided by a set of search strategies are considered to be an advanced solution framework for combinatorial optimization problems, which are independent of the heuristic strategies of specific problems. Typical metaheuristic algorithms include genetic algorithms (GA), particle swarm optimization algorithms (PSO), taboo search, simulated annealing algorithms (SA), etc. They can provide sufficiently high-quality optimization solutions for most NP problems. For decades, they have been widely used in engineering fields such as satellite mission planning and scheduling, and vehicle path optimization. Although metaheuristic algorithms can obtain better calculation results, they still inevitably fall into local optimality and sometimes cannot provide satisfactory convergence performance. In terms of computational efficiency, they are often limited by the checking and conflict resolution process of complex constraints. The task-intensive multi-satellite mission planning problem requires the solution algorithm to have the ability to handle more complex constraint processing and conflict resolution.

[0005] Deep reinforcement learning algorithms for solving satellite mission planning problems: In recent years, algorithms that combine deep learning, reinforcement learning, and other neural networks have been proposed to solve multi-satellite mission planning problems. Deep reinforcement learning (DRL) has shown strong capabilities and potential in solving multi-satellite mission planning and scheduling problems. However, on the one hand, the use of deep reinforcement learning algorithms to solve multi-satellite mission planning problems is still in its infancy, and most research is still at the stage of solving single-satellite problems; on the other hand, as the number of missions increases, some deep reinforcement learning algorithms are still inferior to baseline algorithms such as genetic algorithms in terms of solution quality; in addition, whether these methods are applicable to large-scale multi-satellite mission planning problems with tens of thousands of missions has not been verified and discussed.

[0006] The above research can effectively solve combinatorial optimization problems such as satellite mission planning to a certain extent. However, when faced with the solution of satellite mission planning problems for large-scale intensive tasks, task constraint checking and conflict resolution in a huge solution space become important factors affecting the efficiency of algorithm solutions. However, very few studies focus on the important link of constraint checking and conflict resolution in the satellite mission planning process. There are still the following challenges in finding an efficient solution to the satellite mission planning problem for large-scale intensive tasks:

[0007] Due to the working characteristics of satellites, the single-satellite mission planning problem itself is highly complex. The constraints between satellite resource capabilities and mission requirements are difficult to characterize and describe, and simple greedy construction algorithms or heuristic strategies are difficult to obtain high-quality solutions.

[0008] Large-scale multi-satellite mission planning problems often lead to an explosive increase in problem size and task density. However, the effectiveness of most algorithms in solving multi-satellite mission planning problems under large-scale tasks has not been verified or discussed. Therefore, it is urgent to propose an efficient solution algorithm for multi-satellite mission planning problems suitable for large-scale intensive tasks.

[0009] Mission constraint checking and conflict resolution methods are important factors affecting the computational efficiency of satellite mission planning algorithms. Most current studies have not yet conducted theoretical research and discussion on the essential relationship of constraints in satellite mission planning. Summary of the invention

[0010] Therefore, the present invention proposes a large-scale satellite mission dynamic constraint conflict resolution method based on heterogeneous graph learning. By introducing the powerful representation ability of heterogeneous graphs for the relationship between unstructured and multi-type data, a satellite mission dynamic constraint conflict graph model based on heterogeneous graphs is constructed, and the temporal change characteristics of complex dynamic constraints of satellites are converted into spatial topological characteristics of multi-type nodes and edges under the heterogeneous graph; and through the satellite mission dynamic constraint conflict resolution strategy learning and multi-satellite collaborative task planning solution method design, the number of conflict resolution times in the multi-satellite task planning solution process is greatly reduced, and efficient scheduling and allocation decisions for multi-satellite large-scale intensive tasks are achieved.

[0011] A method for resolving dynamic constraint conflicts of large-scale satellite missions based on heterogeneous graph learning, including:

[0012] Mathematically describe the large-scale satellite mission planning problem and design the optimization objective function; analyze the characteristics and mathematically express the constraints in the large-scale satellite mission planning problem, and divide the constraints into static constraints and dynamic constraints according to their different characteristics;

[0013] For static constraints, a satellite mission static constraint conflict graph model is constructed; for dynamic constraints, a satellite mission dynamic constraint conflict graph model is constructed;

[0014] Input satellite missions and their original information, solve the static constraint conflict graph model, eliminate the static constraint conflicts between tasks, and obtain a set of satellite tasks with reduced static constraint conflicts; for the set of satellite tasks with reduced static constraint conflicts, solve the dynamic constraint conflict graph model of satellite tasks constituted by them, eliminate the dynamic constraint conflicts between tasks, and obtain the probability of satellite mission selection; based on the satellite mission selection probability after dynamic constraint conflicts are reduced, optimize the decision of satellite missions, and finally obtain the satellite mission planning result.

[0015] Furthermore, the static constraint is an inherent attribute of the task, which is independent of the decision sequence, and the constraint association between tasks does not change with the combination of tasks; the dynamic constraint is related to the decision sequence, and the constraint association relationship between tasks changes dynamically with the combination of tasks; the static constraint conflict graph model of the satellite task is specifically:

[0016] Consider tasks as nodes, V = {task1,task2,...,task k ,...,task K}, taking the constraint conflicts between tasks as edges, E = {conflict ij |i,j=1,2,...,K,i≠j}.

[0017] Furthermore, the satellite dynamic constraint conflict graph model is specifically:

[0018] Construct a dynamic constraint conflict heterogeneous graph for satellite missions, which is mathematically described as Where V and E represent the set of nodes and edges respectively, F and R represent the type set of nodes and edges respectively, satisfying |F|+|R|>2, and each node v∈V is associated with a node type mapping function Each edge e∈E is associated with an edge type mapping function φ:E→R.

[0019] In the satellite mission dynamic constraint conflict heterogeneous graph, the nodes are satellite missions and dynamic constraint critical conflict group nodes, the edges are the connection edges between energy constraint critical conflict group nodes and satellite mission nodes, and the connection edges between storage constraint critical conflict group nodes and satellite mission nodes. The dynamic constraint critical conflict group nodes are divided into energy constraint critical conflict group nodes and storage constraint critical conflict group nodes.

[0020] Furthermore, the dynamic constraint critical conflict group node is: the satellite task node connected by the dynamic constraint critical conflict group node just exceeds the satellite energy and storage limits, that is, dynamic constraint conflicts such as energy and storage occur. If one task is deleted from these tasks, the satellite energy and storage constraints will be just met.

[0021] Furthermore, the method for acquiring the dynamic constraint critical conflict clique node adopts a dynamic constraint critical conflict clique node generation algorithm based on a sampling strategy, and the specific steps are as follows:

[0022] (1) Import all satellite mission sets and determine the number of dynamic constraint critical conflict clusters to be sampled;

[0023] (2) Record the current number of samples as i and initialize the conflict group C i ;

[0024] (3) Randomly select a task from the satellite task set and add it to the conflict group C i middle;

[0025] (4) Conflict Group C i Perform dynamic constraint checking on the tasks in . If the constraints are met, go to step (3); if not, go to step (5);

[0026] (5) Determine whether the current conflict group is a critical conflict group. If so, save the current critical conflict group minC i , and go to step (7); otherwise, go to step (6);

[0027] (6) Randomly delete a task from the current conflict group and go to step (5);

[0028] (7) Determine whether the number of conflict clusters meets the sampling number requirement. If so, output the set of all critical conflict clusters, where each critical conflict cluster corresponds to a dynamic constraint critical conflict cluster node. Otherwise, i = i + 1, go to step (2) and continue sampling.

[0029] Furthermore, the solution to the satellite mission dynamic constraint conflict graph model is as follows:

[0030] Input and output element design: The input elements include the task constraint heterogeneous graph, task node characteristics, and constraint conflict meta-path set; the output element is the probability set of the satellite task being selected for execution;

[0031] Heterogeneous graph attention network structure design: The heterogeneous graph attention network structure includes two-layer attention mechanisms: node attention and meta-path attention. Node attention mainly represents the importance of each task node, while meta-path attention mainly represents the contribution of different categories of dynamic constraint critical conflict groups to the conflict correlation between tasks.

[0032] Conflict resolution strategy learning and training: The conflict strategy learning and training process adopts an unsupervised training method, and the loss function adopts a quadratic unconstrained optimization function that is improved from discrete to continuous.

[0033] Furthermore, the input and output elements are specifically:

[0034] Input elements include: Satellite mission dynamic constraint conflict heterogeneous graph Satellite mission node characteristics The satellite mission node features include satellite number, observation target number, mission number, observation window start time, observation window end time, duration, satellite attitude angle, imaging energy consumption, storage usage, mission benefit, etc.; dynamic constraint conflict metapath set metaPath = {Φ1, Φ2, ..., Φ m ,...,Φ M}; conflicting metapath Φ m It refers to the collection of various nodes through the "satellite mission task i - Dynamic constraint critical conflict group node of category m m -Satellite mission task j "A path connected by relationships;

[0035] The output element is the probability set of each satellite mission being selected for execution, that is, probs = {p1, p2, ..., p i ,...,p N}.

[0036] Furthermore, the specific steps of conflict resolution strategy learning and training are as follows:

[0037] (1) Based on the satellite mission probability values ​​output by the heterogeneous graph attention network, determine whether they are greater than the set threshold one by one. If so, add them to the execution task node set and go to step (3); otherwise, add them to the deleted task node set and go to step (2);

[0038] (2) Based on the deleted task node, dissolve the critical conflict group to which it belongs, and update the critical conflict group node set, and go to step (3);

[0039] (3) Calculate the penalty matrix P: The penalty matrix P is obtained based on the current critical conflict group node set, and the satellite task nodes connected to the critical conflict group nodes are obtained. The number of edges associated with these satellite task nodes and the critical conflict group nodes is counted. The penalty mechanism can adopt strategies such as single point-single penalty, single point-cumulative penalty, multiple points-single penalty, and multiple points-cumulative penalty to obtain the task penalty matrix.

[0040] (4) Calculate the loss function loss = X T (R+P)X, where matrix R is the reward matrix of the task, such as the benefit or priority of each task; P is the penalty matrix; X is the selection probability matrix of the task;

[0041] (5) Determine whether the maximum number of iterations has been reached or whether the loss function has converged. If so, go to (6). Otherwise, update the parameters of the heterogeneous graph attention network and continue training.

[0042] (6) Output the satellite mission selection probability.

[0043] Furthermore, the decision optimization of satellite missions is as follows:

[0044] (1) Initialize the satellite mission planning decision result vector and initialize the conflict resolution statistics count to 0;

[0045] (2) According to the satellite mission selection probability obtained from the satellite mission constraint conflict graph model and conflict resolution, the satellite mission set is sorted according to the probability value;

[0046] (3) Check the constraints of the selected task set. If all constraints are met, go to step (6); if the constraints are not met, go to step (4);

[0047] (4) In the current set of tasks that have constraint conflicts, delete the task with the smallest probability value;

[0048] (5) Update the task planning decision vector and update the conflict resolution statistics count = count + 1, and then go to step (3);

[0049] (6) Output the task planning decision result vector and the number of conflict resolutions.

[0050] A system for implementing the above-mentioned large-scale satellite mission planning method based on dynamic constraint conflict graph learning includes: a satellite mission input module, a static constraint conflict graph model construction module, a dynamic constraint conflict graph model construction module, a static constraint conflict resolution module, a dynamic constraint conflict resolution module, a satellite mission decision optimization module, and a satellite mission planning result output module.

[0051] Beneficial effects:

[0052] In view of the problems of large-scale solution space of multi-satellite mission planning under dense mission scenarios, complex and changeable task constraint relationships, long conflict detection and resolution time, and low solution efficiency, the present invention proposes a large-scale satellite mission dynamic constraint conflict resolution method based on heterogeneous graph learning. Through the construction of multi-satellite mission constraint conflict graph model, satellite mission dynamic constraint conflict resolution strategy learning based on heterogeneous graph attention network, and multi-satellite collaborative task planning solution method design, efficient scheduling and allocation decision-making of multi-satellite large-scale dense tasks are achieved. The problem of efficient calculation of constraint detection and conflict resolution in the process of multi-satellite mission planning under dense tasks is solved. Through the construction and learning of heterogeneous graphs, the common knowledge of complex dynamic constraint conflict relationships in different mission scenarios is extracted, which greatly reduces the number of conflict resolutions in the process of multi-satellite mission planning solution, and provides important theoretical and technical support for further improving the intelligence and efficiency of future satellite management and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The overall process of the dynamic constraint conflict resolution method for large-scale satellite missions based on heterogeneous graph learning;

[0054] Figure 2 This is an example of a static constraint and a dynamic constraint conflict graph;

[0055] Figure 3 Dynamically constrain time-varying characteristics for satellite missions;

[0056] Figure 4 Generate algorithmic flow for dynamic constrained critical conflict cliques based on sampling strategy;

[0057] Figure 5 The overall process of the algorithm;

[0058] Figure 6 Design of attention network structure and conflict resolution learning and training for heterogeneous graphs;

[0059] Figure 7 Planning algorithm flow for multi-satellite collaborative missions based on probabilistic greedy strategy;

[0060] Figure 8 It is the simulation training time curve;

[0061] Fig. 9 Probability distribution bar graphs for simulation task selection, (a) for 1000 task nodes, (b) for 2000 task nodes, (c) for 3000 task nodes, (d) for 4000 task nodes, and (e) for 5000 task nodes;

[0062] Fig.10 The following are the average planning indicator statistics under different task scales: (a) is the average number of executed tasks, (b) is the average total revenue of executed tasks, (c) is the average number of task conflicts resolved during the planning process, and (d) is the average planning calculation time. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is further described below in conjunction with the embodiments, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0064] The following is a specific implementation method of a large-scale satellite mission dynamic constraint conflict resolution method based on heterogeneous graph learning. The overall process is as follows: Figure 1 As shown, the following steps are included:

[0065] Mathematically describe the large-scale satellite mission planning problem and design the optimization objective function; analyze the characteristics and mathematically express the constraints in the large-scale satellite mission planning problem, and divide the constraints into static constraints and dynamic constraints according to their different characteristics; Figure 2 This is an example of a static constraint and a dynamic constraint conflict graph.

[0066] The static constraints are inherent attributes of the tasks and have nothing to do with the decision sequence. The constraint association between tasks does not change with the combination of tasks. For example, the time constraints of tasks in the satellite mission planning process. Specifically, the time constraint means that a satellite can only perform one observation task in the same time period (time window). If the time windows of different observation tasks of a satellite overlap, it means that a time conflict occurs.

[0067] Dynamic constraints are related to the decision sequence, and the constraint association between tasks changes dynamically with the combination of tasks. For example, satellite storage constraints and energy constraints in the satellite mission planning process. Specifically, "storage constraints" means that due to the limited satellite storage space, the amount of data for all tasks cannot exceed the satellite storage capacity limit; "energy constraints" means that for each satellite, the total energy consumption of the tasks cannot exceed the satellite's battery energy limit. Figure 3 Description of time-varying characteristics of dynamic constraints for satellite missions.

[0068] For static constraints, (such as mission time constraints), the non-time-varying characteristics of static constraints are considered and a static constraint conflict graph model for satellite missions is constructed.

[0069] The satellite mission static constraint conflict graph model is specifically:

[0070] Consider tasks as nodes, V = {task1,task2,...,task k ,...,task K}, taking the constraint conflicts between tasks as edges, E = {conflict ij |i,j=1,2,...,K,i≠j}.

[0071] The above satellite task static constraint conflict graph model is a satellite task static constraint conflict graph model based on a homogeneous graph. A homogeneous graph refers to a graph with only one node type and edge type, which is mathematically described as G = (V, E). For the satellite task static constraint conflict graph, all satellite tasks are one type of nodes, that is, V = {task1, task2, ..., task k ,...,task K}, the static constraint conflict between satellite tasks (such as time conflict) is a type of edge, that is, E = {conflict ij |i,j=1,2,...,K,i≠j}.

[0072] Satellite mission static constraint conflict graph example based on homogeneous graph Figure 2 As shown in (a), task nodes connected by edges indicate that the tasks conflict with each other and cannot be executed simultaneously. Therefore, the resolution of satellite static constraint conflicts is transformed into the problem of finding independent set nodes in the graph, that is, finding a set of nodes that are not connected to each other in the graph. The representation and learning of static constraint conflict relations can be achieved by using neural network algorithms such as graph convolutional networks (GCN), graph sample and aggregate networks (GraphSAGE), and graph attention networks (GAT) that process homogeneous graphs.

[0073] For dynamic constraints (such as energy constraints, storage constraints, and attitude maneuver constraints), considering the higher complexity of dynamic constraints and the time-varying characteristics of decision sequences, a satellite mission dynamic constraint conflict graph model is constructed. The present invention constructs a task dynamic constraint conflict graph model based on a heterogeneous graph. Figure 2 As shown in (b), taking satellite energy, storage and other dynamic constraints as an example, a graph contains three types of nodes: satellite missions (red dots), energy constraint critical conflict cluster nodes (green squares), and storage constraint critical conflict cluster nodes (blue triangles), as well as two types of edges connecting storage and energy critical conflict cluster nodes with satellite mission nodes (green solid lines and blue dotted lines), respectively.

[0074] The satellite dynamic constraint conflict graph model is specifically:

[0075] Construct a dynamic constraint conflict heterogeneous graph for satellite missions, which is mathematically described as Where V and E represent the set of nodes and edges respectively, F and R represent the type set of nodes and edges respectively, satisfying |F|+|R|>2, and each node v∈V is associated with a node type mapping function Each edge e∈E is associated with an edge type mapping function φ:E→R.

[0076] In the satellite mission dynamic constraint conflict heterogeneous graph, the nodes are satellite missions and dynamic constraint critical conflict group nodes, the edges are the connection edges between energy constraint critical conflict group nodes and satellite mission nodes, and the connection edges between storage constraint critical conflict group nodes and satellite mission nodes. The dynamic constraint critical conflict group nodes are divided into energy constraint critical conflict group nodes and storage constraint critical conflict group nodes.

[0077] The above satellite dynamic constraint conflict graph model is a satellite mission dynamic constraint conflict graph based on a heterogeneous graph.

[0078] A heterogeneous graph is a graph that contains multiple different types of nodes and edges. It is mathematically described as Where V and E represent the set of nodes and edges respectively, F and R represent the type set of nodes and edges respectively, satisfying |F|+|R|>2, and each node v∈V is associated with a node type mapping function Each edge e∈E is associated with an edge type mapping function φ:E→R.

[0079] Based on the basic principles of heterogeneous graphs and the characteristics of satellite dynamic constraints, heterogeneous graphs are used to describe the dynamic constraint conflict relationships of satellite missions. The temporal variation characteristics of complex dynamic constraints of satellites are converted into spatial topological characteristics of multiple types of nodes and edges under heterogeneous graphs.

[0080] The dynamic constraint critical conflict group node is: the satellite task node connected by the dynamic constraint critical conflict group node just exceeds the satellite energy and storage limits, that is, dynamic constraint conflicts such as energy and storage occur. If one task is deleted from these tasks, the satellite energy and storage constraints will be met.

[0081] How to obtain dynamic constraint critical conflict cluster nodes is the key to constructing a dynamic constraint heterogeneous graph for satellite missions. Since it is difficult to accurately describe the dynamic constraint relationship between task nodes and to traverse and obtain all dynamic constraint conflict clusters in polynomial time, the present invention adopts a dynamic constraint critical conflict cluster generation algorithm based on a sampling strategy. The basic process of the algorithm is as follows: Figure 4 shown.

[0082] The acquisition method of dynamic constraint critical conflict group nodes adopts a dynamic constraint critical conflict group node generation algorithm based on sampling strategy. The specific steps are as follows:

[0083] (1) Import all satellite mission sets and determine the number of dynamic constraint critical conflict clusters to be sampled;

[0084] (2) Record the current number of samples as i and initialize the conflict group C i ;

[0085] (3) Randomly select a task from the satellite task set and add it to the conflict group C i middle;

[0086] (4) Conflict Group C i Perform dynamic constraint checking on the tasks in . If the constraints are met, go to step (3); if not, go to step (5);

[0087] (5) Determine whether the current conflict group is a critical conflict group. If so, save the current critical conflict group minC i , and go to step (7); otherwise, go to step (6);

[0088] (6) Randomly delete a task from the current conflict group and go to step (5);

[0089] (7) Determine whether the number of conflict groups meets the sampling number requirement. If so, output all critical conflict groups, each of which corresponds to a dynamic constraint critical conflict group node; otherwise, i = i + 1, go to step (2) and continue sampling. Input the satellite mission and its original information, solve the static constraint conflict graph model, eliminate the static constraint conflicts between tasks, and obtain the satellite mission set with reduced static constraint conflicts; for the satellite mission set with reduced static constraint conflicts, solve the satellite mission dynamic constraint conflict graph model, eliminate the dynamic constraint conflicts between tasks, and obtain the satellite mission selection probability; based on the satellite mission selection probability after dynamic constraint conflicts are eliminated, optimize the satellite mission decision, and finally obtain the satellite mission planning result.

[0090] The overall process of satellite mission dynamic constraint conflict resolution strategy learning algorithm based on heterogeneous graph attention network is as follows: Figure 5 shown.

[0091] The solution to the satellite mission dynamic constraint conflict graph model is as follows:

[0092] Input and output element design: The input elements include the task constraint heterogeneous graph, task node characteristics, and constraint conflict meta-path set; the output element is the probability set of the satellite task being selected for execution;

[0093] Heterogeneous graph attention network structure design: The heterogeneous graph attention network structure includes two-layer attention mechanisms: node attention and meta-path attention. Node attention mainly represents the importance of each task node, and meta-path attention mainly represents the contribution of different categories of dynamic constraint critical conflict groups to the conflict correlation between tasks. Figure 6 As shown in (a), it includes two-layer attention mechanisms: node attention and meta-path attention. Node attention mainly represents the importance of each task node, and meta-path attention mainly represents the contribution degree of different categories of dynamic constraint critical conflict groups to the conflict correlation between tasks.

[0094] Conflict resolution strategy learning and training: The conflict strategy learning and training process adopts unsupervised training, and the loss function adopts the quadratic unconstrained optimization function (QUBO) improved from discrete to continuous; the basic process of conflict resolution strategy learning and training is as follows Figure 6 (b) as shown.

[0095] The input and output elements are specifically:

[0096] Input elements include: Satellite mission dynamic constraint conflict heterogeneous graph Satellite mission node characteristics The satellite mission node features include satellite number, observation target number, mission number, observation window start time, observation window end time, duration, satellite attitude angle, imaging energy consumption, storage usage, mission benefit, etc.; dynamic constraint conflict metapath set metaPath = {Φ1, Φ2, ..., Φ m ,...,Φ M}; conflicting metapath Φ m It refers to the collection of various nodes (satellite mission nodes, critical conflict group nodes) through the "satellite mission task i - Dynamic constraint critical conflict group node of category m m -Satellite mission task j "A path connected by relationships;

[0097] The output element is the probability set of each satellite mission being selected for execution, that is, probs = {p1, p2, ..., p i ,...,p N}.

[0098] The specific steps of conflict resolution strategy learning and training are as follows:

[0099] (1) Based on the satellite mission probability values ​​output by the heterogeneous graph attention network, determine whether they are greater than the set threshold one by one. If so, add them to the execution task node set and go to step (3); otherwise, add them to the deleted task node set and go to step (2);

[0100] (2) Based on the deleted task node, dissolve the dynamic constraint critical conflict group to which it belongs, and update the dynamic constraint critical conflict group node set, and go to step (3);

[0101] (3) Calculate the penalty matrix P: The penalty matrix P is obtained based on the current dynamic constraint critical conflict group node set, obtains the satellite task nodes connected to the dynamic constraint critical conflict group nodes by edges, and counts the number of edges associated with these satellite task nodes and the dynamic constraint critical conflict group nodes. The penalty mechanism can adopt strategies such as single point-single penalty, single point-cumulative penalty, multiple points-single penalty, and multiple points-cumulative penalty to obtain the task penalty matrix.

[0102] (4) Calculate the loss function loss = X T (R+P)X, where matrix R is the reward matrix of the task, such as the benefit or priority of each task; P is the penalty matrix; X is the selection probability matrix of the task;

[0103] (5) Determine whether the maximum number of iterations has been reached or whether the loss function has converged. If so, go to (6). Otherwise, update the parameters of the heterogeneous graph attention network and continue training.

[0104] (6) Output the satellite mission selection probability.

[0105] Based on the satellite mission selection probability after dynamic constraint conflict reduction, the satellite mission decision is optimized, and a multi-satellite collaborative mission planning solution method based on probabilistic greedy strategy is designed. By utilizing the satellite mission constraint conflict graph model and the constraint conflict resolution experience knowledge provided by the conflict resolution network module, the difficulty of conflict resolution and the time-consuming constraint checking in the decision-making process of large-scale intensive mission planning can be effectively reduced, the satellite mission planning results can be quickly obtained, and the efficiency of satellite mission planning solution can be further improved.

[0106] The overall process of the multi-satellite collaborative mission planning solution algorithm based on the probability greedy strategy is as follows: Figure 7 shown.

[0107] The decision optimization for satellite missions is as follows:

[0108] (1) Initialize the satellite mission planning decision result vector and initialize the conflict resolution statistics count to 0;

[0109] (2) According to the satellite mission selection probability obtained from the satellite mission constraint conflict graph model and conflict resolution, the satellite mission set is sorted according to the probability value;

[0110] (3) Check the constraints of the selected task set. If all constraints are met, go to step (6); if the constraints are not met, go to step (4);

[0111] (4) In the current set of tasks that have constraint conflicts, delete the task with the smallest probability value;

[0112] (5) Update the task planning decision vector and update the conflict resolution statistics count = count + 1, and then go to step (3);

[0113] (6) Output the task planning decision result vector and the number of conflict resolutions.

[0114] The present invention also proposes a system for implementing the above-mentioned large-scale satellite mission planning method based on dynamic constraint conflict graph learning, the system comprising: a satellite mission input module, a static constraint conflict graph model construction module, a dynamic constraint conflict graph model construction module, a static constraint conflict resolution module, a dynamic constraint conflict resolution module, a satellite mission decision optimization module, and a satellite mission planning result output module.

[0115] Simulation verification:

[0116] Taking actual satellite mission data as a reference, in simulation mission scenarios of different scales, including 10,000, 20,000, 30,000, 40,000 and 50,000 satellite missions respectively, the effectiveness of the large-scale satellite mission dynamic constraint conflict resolution method based on heterogeneous graph learning was verified.

[0117] The simulation verification results are as follows:

[0118] 1. Analysis of training results of dynamic constraint conflict resolution network model

[0119] The dynamic constraint conflict resolution graph network model is simulated and trained through mission data sets of different sizes ranging from 10,000 to 50,000, focusing on verifying the effectiveness of the dynamic constraint conflict resolution method for satellite missions. The core network modules use graph convolutional network (GCN), sampling aggregation graph network (SAGE), and heterogeneous graph attention network (HAN), and a comparative analysis is performed.

[0120] In terms of training time, the training time of different networks is as follows: Figure 8 As shown in Figure 2, as the number of task nodes increases, the training time of the model also gradually increases. Under the same task scale, the HAN model requires the shortest training time, while the GCN model takes the longest training time, and the SAGE model is between the two.

[0121] Fig. 9The probability distribution of tasks selected for execution at different task sizes is shown. The highly sparse arrangement in the histogram intuitively reflects the model's decision preference for task selection. The GCN model produces a dense distribution of task selection probabilities, with the selection probability of almost all tasks close to 1, showing aggressive and greedy behavior. In contrast, the SAGE model behaves more conservatively, with a sparse distribution of task selection probabilities, and the selection probability of most tasks is close to 0. The sparsity of the probability distribution of the HAN model is between the GCN and SAGE models.

[0122] In summary, through the analysis of the training process performance and training results, compared with the GCN and SAGE models, the HAN model proposed in this patent shows significant performance advantages in many aspects such as training loss function convergence, training time and task selection probability distribution.

[0123] 2. Analysis of the effectiveness of the dynamic constraint conflict resolution network model for solving large-scale satellite mission planning

[0124] To further verify the effectiveness of the proposed task conflict resolution method based on dynamic constraint heterogeneous graph (module B) for solving large-scale satellite mission planning problems. Based on test data sets of different task sizes (there are certain distribution differences with the training data sets), the task conflict resolution model based on static constraint homogeneous graph (module A) is first input to process static constraints, and a set of satellite tasks with reduced static constraint conflicts is obtained. Then, the task conflict resolution model based on dynamic constraint heterogeneous graph (module B) is input to process dynamic constraints. Finally, the task decision optimization method based on greedy strategy (module C) is used to perform planning and solve, and the final task planning result is obtained.

[0125] Through task data sets of different sizes ranging from 10,000 to 50,000, ablation analysis simulation experiments are carried out on each module. Based on the simulation results, the effectiveness of the dynamic constraint conflict resolution model in the task planning decision-making process is analyzed.

[0126] After the test data sets of different task sizes pass through modules A, B and C in turn, the task planning results are obtained. The average number of scheduled execution tasks is as follows: Fig.10 As shown in (a), the average reward for executing a task is Fig.10 (b) shows that the trends of the two are similar. The average number of task conflict resolutions and the average scheduling computation time are shown in Fig.10 (c) and Fig.10 Obviously, compared with modules A and C that only have the ability to handle static constraints, adding module B that has the ability to handle dynamic constraints can significantly reduce the pressure of task conflict resolution during constraint checking, as shown in (d). Fig.10(c) As shown. Specifically, different models of module B have different performance characteristics. For different task sizes, the HAN model reduces conflict resolution by an average of 73.48%, SAGE reduces by about 54.25%, and GCN reduces by about 8.92%. Therefore, the scheduling efficiency is improved to a certain extent while maintaining the quality of the solution, as shown in Figure 3 As shown in (a), (b), and (d), the average number of tasks planned to be executed and the average benefits of executed tasks under different task sizes remain at almost the same level. Fig.10 As shown in (d), as the task size increases, the effect of module B in reducing scheduling time consumption becomes more significant. Compared with the GCN and SAGE models, the HAN model obviously has obvious absolute advantages in task conflict resolution and computing time.

[0127] The above is only a preferred embodiment of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for resolving dynamic constraint conflicts of large-scale satellite missions based on heterogeneous graph learning, comprising: Provide mathematical description for large-scale satellite mission planning problems and design optimization objective functions; The constraints in large-scale satellite mission planning are analyzed and mathematically expressed, and are divided into static constraints and dynamic constraints according to their different characteristics. For static constraints, a satellite mission static constraint conflict graph model is constructed; for dynamic constraints, a satellite mission dynamic constraint conflict graph model is constructed; Input satellite missions and their original information, solve the static constraint conflict graph model, eliminate the static constraint conflicts between tasks, and obtain a set of satellite tasks with reduced static constraint conflicts; for the set of satellite tasks with reduced static constraint conflicts, solve the dynamic constraint conflict graph model of satellite tasks constituted by them, eliminate the dynamic constraint conflicts between tasks, and obtain the probability of satellite mission selection; based on the satellite mission selection probability after dynamic constraint conflicts are reduced, optimize the decision of satellite missions, and finally obtain the satellite mission planning result.

2. According to claim 1, a large-scale satellite mission dynamic constraint conflict resolution method based on heterogeneous graph learning is characterized in that: The static constraints are inherent attributes of the tasks and have nothing to do with the decision sequence. The constraint association between tasks does not change with the combination of tasks. The dynamic constraints are related to the decision sequence. The constraint association relationship between tasks changes dynamically with the combination of tasks. The satellite task static constraint conflict graph model is specifically: Consider tasks as nodes, V = {task1,task2,...,task k ,...,task K }, taking the constraint conflicts between tasks as edges, E = {conflict ij |i,j=1,2,...,K,i≠j}.

3. According to claim 1, a large-scale satellite mission dynamic constraint conflict resolution method based on heterogeneous graph learning is characterized in that: The satellite dynamic constraint conflict graph model is specifically: Construct a dynamic constraint conflict heterogeneous graph for satellite missions, which is mathematically described as Where V and E represent the set of nodes and edges respectively, F and R represent the type set of nodes and edges respectively, satisfying |F|+|R|>2, and each node v∈V is associated with a node type mapping function Each edge e∈E is associated with an edge type mapping function φ:E→R; In the satellite mission dynamic constraint conflict heterogeneous graph, the nodes are satellite missions and dynamic constraint critical conflict group nodes, the edges are the connection edges between energy constraint critical conflict group nodes and satellite mission nodes, and the connection edges between storage constraint critical conflict group nodes and satellite mission nodes. The dynamic constraint critical conflict group nodes are divided into energy constraint critical conflict group nodes and storage constraint critical conflict group nodes.

4. The method for resolving dynamic constraint conflicts of large-scale satellite missions based on heterogeneous graph learning according to claim 3 is characterized in that: The dynamically constrained critical conflict cluster node is: the satellite mission node connected by the dynamically constrained critical conflict cluster node just exceeds the satellite energy and storage capacity limit.

5. A large-scale satellite mission dynamic constraint conflict resolution method based on heterogeneous graph learning according to claim 3, characterized in that: The method for acquiring the dynamic constraint critical conflict clique node adopts a dynamic constraint critical conflict clique node generation algorithm based on a sampling strategy, and the specific steps are as follows: (1) Import all satellite mission sets and determine the number of dynamic constraint critical conflict clusters to be sampled; (2) Record the current number of samples as i and initialize the conflict group C i ; (3) Randomly select a task from the satellite task set and add it to the conflict group C i middle; (4) Conflict Group C i Perform dynamic constraint checking on the tasks in . If the constraints are met, go to step (3); if not, go to step (5); (5) Determine whether the current conflict group is a critical conflict group. If so, save the current critical conflict group minC i , and go to step (7); otherwise, go to step (6); (6) Randomly delete a task from the current conflict group and go to step (5); (7) Determine whether the number of conflict clusters meets the sampling number requirement. If so, output the set of all critical conflict clusters, where each critical conflict cluster corresponds to a dynamic constraint critical conflict cluster node. Otherwise, i = i + 1, go to step (2) and continue sampling.

6. The method for resolving dynamic constraint conflicts of large-scale satellite missions based on heterogeneous graph learning according to claim 1 is characterized in that: The solution to the satellite mission dynamic constraint conflict graph model is as follows: Input and output element design: The input elements include the task constraint heterogeneous graph, task node characteristics, and constraint conflict meta-path set; the output element is the probability set of the satellite task being selected for execution; Heterogeneous graph attention network structure design: The heterogeneous graph attention network structure includes two-layer attention mechanisms: node attention and meta-path attention. Node attention mainly represents the importance of each task node, while meta-path attention mainly represents the contribution of different categories of dynamic constraint critical conflict groups to the conflict correlation between tasks. Conflict resolution strategy learning and training: The conflict strategy learning and training process adopts an unsupervised training method, and the loss function adopts a quadratic unconstrained optimization function that is improved from discrete to continuous.

7. The method for resolving dynamic constraint conflicts of large-scale satellite missions based on heterogeneous graph learning according to claim 6 is characterized in that: The input and output elements are specifically: Input elements include: Satellite mission dynamic constraint conflict heterogeneous graph Satellite mission node characteristics The satellite mission node features include satellite number, observation target number, mission number, observation window start time, observation window end time, duration, satellite attitude angle, imaging energy consumption, storage usage, mission benefit, etc.; dynamic constraint conflict metapath set metaPath = {Φ1, Φ2, ..., Φ m ,...,Φ M }; conflicting metapath Φ m It refers to the collection of various nodes through the "satellite task i - Dynamic constraint critical conflict group node of category m m -Satellite mission task j "A path connected by relationships; The output element is the probability set of each satellite mission being selected for execution, that is, probs = {p1, p2, ..., p i ,...,p N }.

8. The method for resolving dynamic constraint conflicts of large-scale satellite missions based on heterogeneous graph learning according to claim 6 is characterized in that: The specific steps of conflict resolution strategy learning and training are as follows: (1) Based on the satellite mission probability values ​​output by the heterogeneous graph attention network, determine whether they are greater than the set threshold one by one. If so, add them to the execution task node set and go to step (3); otherwise, add them to the deleted task node set and go to step (2); (2) Based on the deleted task node, dissolve the critical conflict group to which it belongs, and update the critical conflict group node set, and go to step (3); (3) Calculate the penalty matrix P: The penalty matrix P is obtained based on the current critical conflict group node set, and the satellite task nodes connected to the critical conflict group nodes are obtained. The number of edges associated with these satellite task nodes and the critical conflict group nodes is counted. The penalty mechanism can adopt strategies such as single point-single penalty, single point-cumulative penalty, multiple points-single penalty, and multiple points-cumulative penalty to obtain the task penalty matrix. (4) Calculate the loss function loss = X T (R+P)X, where matrix R is the reward matrix of the task, such as the benefit or priority of each task; P is the penalty matrix; X is the selection probability matrix of the task; (5) Determine whether the maximum number of iterations has been reached or whether the loss function has converged. If so, go to (6). Otherwise, update the parameters of the heterogeneous graph attention network and continue training. (6) Output the satellite mission selection probability.

9. The method for resolving dynamic constraint conflicts of large-scale satellite missions based on heterogeneous graph learning according to claim 1, characterized in that: The decision optimization for satellite missions is as follows: (1) Initialize the satellite mission planning decision result vector and initialize the conflict resolution statistics count to 0; (2) According to the satellite mission selection probability obtained from the satellite mission constraint conflict graph model and conflict resolution, the satellite mission set is sorted according to the probability value; (3) Check the constraints of the selected task set. If all constraints are met, go to step (6); if the constraints are not met, go to step (4); (4) In the current set of tasks that have constraint conflicts, delete the task with the smallest probability value; (5) Update the task planning decision vector and update the conflict resolution statistics count = count + 1, and then go to step (3); (6) Output the task planning decision result vector and the number of conflict resolutions.

10. A system for implementing the large-scale satellite mission planning method based on dynamic constraint conflict graph learning according to any one of claims 1 to 9, characterized in that: The system includes: a satellite mission input module, a static constraint conflict graph model building module, a dynamic constraint conflict graph model building module, a static constraint conflict resolution module, a dynamic constraint conflict resolution module, a satellite mission decision optimization module, and a satellite mission planning result output module.

Images