A single-satellite autonomous task scheduling method and system for non-time-sensitive mobile target tracking

Abstract

The present invention provides a single-satellite autonomous task scheduling method and system for non-time-sensitive mobile target tracking, which collects non-time-sensitive mobile target task sets; constructs a single-satellite autonomous task scheduling optimization model; solves the single-satellite autonomous task scheduling optimization model to obtain a single-satellite autonomous task scheduling plan and then outputs it. Tasks are regarded as nodes, and the single-satellite autonomous task scheduling model is converted into a time-pose adjacency graph under various transient states. The graph attention network model is used to predict the next task node, and a scheduling solution sequence is obtained through step-by-step prediction. The method of solving using the graph attention network model improves the solving efficiency, so that a scheduling plan can be quickly given.

Description

Technical Field

[0001] The present invention belongs to the field of satellite mission planning, and in particular relates to a single-satellite autonomous mission scheduling method and system for non-time-sensitive mobile target tracking. Background Art

[0002] Non-time-sensitive moving targets are a type of moving targets with slow motion characteristics. Typical non-time-sensitive targets include ships on the sea surface and ground vehicles. The single-satellite autonomous task scheduling problem (SSATSP) for non-time-sensitive moving target tracking is a type of tracking problem for dynamic targets, whose target states are constantly changing. In order to respond to the uncertainty of the target, combined with the on-board image recognition capability, it is necessary to quickly autonomously schedule and generate a scheduling plan on the satellite.

[0003] The new generation of agile satellites has flexible three-axis attitude maneuvering capabilities, which expands the observation window of the target; the traditional "ground planning + on-board execution" control mode has a serious lag in responding to mobile targets. The enhancement of on-board computing capabilities and the development of AI technology provide the premise for on-board autonomous intelligence and bring opportunities to solve the lag. Agile capabilities bring constraints of time-dependent characteristics, making satellite planning and scheduling more complicated; timely response to mobile targets depends on fast, high-quality and refined autonomous task scheduling.

[0004] The existing technology uses heuristic algorithms with poor solution quality, and the meta-heuristic algorithm takes a long time to solve, which cannot achieve timely response to the dynamic uncertainty characteristics of mobile targets. At the same time, the operation and control constraints involved on the satellite are complicated and need to accurately avoid constraint conflicts. Rapid, high-quality and refined solution of this scheduling problem is the key to achieving timely response to non-time-sensitive mobile targets. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to quickly and high-quality track and schedule non-time-sensitive mobile targets, and a single-satellite autonomous task scheduling method and system for non-time-sensitive mobile target tracking are proposed.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A single-satellite autonomous task scheduling method for non-time-sensitive mobile target tracking includes the following steps:

[0008] Step 1: Collect non-time-sensitive mobile target task sets;

[0009] Step 2: construct a single-satellite autonomous task scheduling optimization model according to the non-time-sensitive task set;

[0010] Step 3: Solving the single-satellite autonomous task scheduling optimization model to obtain a single-satellite autonomous task scheduling solution;

[0011] Step 4: Output the single-satellite autonomous task scheduling plan.

[0012] Furthermore, the single-satellite autonomous task scheduling optimization model is:

[0013] Objective function:

[0014] Constraints:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] Among them, formula 1 represents the maximum priority sum of completed tasks, x ij represents a 0-1 decision variable, x ij =1 means that task i is the immediate predecessor of task j in the scheduling sequence solution; pri j represents the priority of task j, n tsk Indicates the number of tasks to be planned;

[0026] Formula 2 represents the hard time window constraint of the satellite, that is, the mission must observe within the satellite visible time window; wb i represents the start time of the satellite's visible time window for mission i, wb0 = 0, tb i represents the start time of satellite observation of mission i, tb0=0 represents the initial state time of satellite, te i represents the end time of satellite observation of mission i, te0=0 represents the initial state of the satellite; we i Indicates the end time of the satellite's visible time window for mission i, wb0 = 0;

[0027] Formula 3 represents the equation relationship between the imaging start, end and duration of the task; ct i Indicates the continuous observation time of the satellite on mission i, ct0 = 0;

[0028] Equation 4 represents the time-dependent transition time constraint between tasks; trans(ea i ,ba j ) represents the posture transition time between task i and task j;

[0029] Formula 5 represents the satellite's storage constraint, that is, the power consumption in a single orbit cannot exceed the satellite's remaining power threshold; uie represents the power consumed per unit time during imaging, unit unit, ute represents the power consumed per unit time during attitude maneuvers, unit unit; ζ represents the ratio of the satellite's remaining power threshold;

[0030] Egymax represents the maximum total amount of electricity of the satellite, in unit;

[0031] ea i represents the satellite's final observation attitude for mission i, which is determined by the side swing θ i ,te i , pitch φ i ,te i With yaw ψ i ,te i The three-axis attitude vector is composed of ea0 and ba. i represents the initial observation attitude of the satellite for mission i, which is determined by the side swing θ i ,tb i , pitch φ i ,tb i With yaw ψ i ,tb i The three-axis attitude vector is composed of ba0, which is the initial attitude of the satellite;

[0032] Formula 6 indicates that each task has at most one immediate predecessor task;

[0033] Formula 7 indicates that each task has at most one subsequent task;

[0034] Formula 8 indicates that a task cannot be its immediate predecessor or immediate successor and can only be completed once; Formula 9 indicates the value range of the decision variable;

[0035] Formula 10 shows the calculation method of the shortest posture transition time between two postures, ρi jrepresents the attitude conversion angle of the satellite between mission i and mission j; b0, b1, b2, b3, b4, a0, a1, a2, a3, a4, z0, z1, z2, z3 are preset constants;

[0036] Formula 11 shows the calculation method of the conversion angle between two postures, θ i , t represents the satellite's observed side swing angle for mission i at time t, te i represents the end time of satellite observation of mission i, te0=0 represents the initial state of the satellite, φ i , t represents the satellite's observed elevation angle of mission i at time t, ψ i , t represents the observed yaw angle of the satellite for mission i at time t.

[0037] Furthermore, the method for solving the single-satellite autonomous task scheduling optimization model in step 3 is: regard each target task in the single-satellite autonomous task scheduling optimization model as a graph node, regard the relationship between tasks as a weighted edge relationship in the graph, construct multiple transient time-pose adjacency graph models, solve the time-pose adjacency graph model, and obtain the solution of the single-satellite autonomous task scheduling optimization model.

[0038] Furthermore, the method for constructing the time-pose adjacency graph model is to use the timeline of the satellite operation as the horizontal axis, the side swing attitude angle at the time of passing the top as the vertical axis, and the point (0,0) where the passing time and the side swing are both 0 as the reference reference point. In the coordinate system, each task is regarded as a node in the graph, and the coordinate corresponding to each node is It is composed of the satellite overpass time and the corresponding side swing angle. Each node has the attribute of the task corresponding to the node. The edge weight between node i and node j is composed of the attitude conversion time trans(ea i ,ba j )express.

[0039] Furthermore, a method for solving the time-pose adjacency graph model is to decide the next task node based on the transient time-pose adjacency graph using a graph attention network model structure, and after obtaining the next task node, update the transient time-pose adjacency graph to make a decision on the next task node, and so on, to obtain a scheduling solution sequence.

[0040] Furthermore, the method of using the graph attention network model structure to decide the next task node is:

[0041] Collect graph features in the current state, including node feature vectors and edge feature matrices;

[0042] Normalizing the node feature vector and edge feature matrix;

[0043] The normalized node feature vector and edge feature matrix are input into the graph attention network structure to obtain the node selection probability of the next task node, and the node with the largest probability is selected as a node on the solution sequence.

[0044] Furthermore, each node in the node feature vector has 8 feature attributes: pri i , ct i , wb i , we i , Each edge has five characteristic attributes: ij ,

[0045] Among them, pri i Represents the priority attribute of node i;

[0046] represents the overhead time of the task corresponding to node i;

[0047] represents the overhead lateral swing angle of the task corresponding to node i;

[0048] ct i represents the continuous observation time of the task corresponding to node i;

[0049] wb i Indicates the start time of the visible time window of the task corresponding to node i;

[0050] we i Indicates the end time of the visible time window of the task corresponding to node i;

[0051] An identifier indicating whether the task corresponding to node i is the task to be selected in the current state;

[0052] An identifier indicating whether the task corresponding to node i is the last task in the currently scheduled sequence;

[0053] d ij Represents the distance between the two nodes associated with the edge in the time-pose coordinate;

[0054] Indicates whether the neighbor node j is the first node in ascending order of distance from node i under the time-pose coordinate;

[0055] Indicates whether the neighbor node j is the node identifier within the fifth in ascending order of distance from node i under the time-pose coordinate;

[0056] Indicates whether the neighbor node j is the node identifier within the tenth in ascending order of distance from node i under the time-pose coordinate;

[0057] Indicates whether the neighbor node j is a node identifier within twenty points of the node i in ascending order under the time coordinate.

[0058] Furthermore, the graph attention network structure is a 9-layer deep learning network, the first four layers are embedding layers, all of which adopt single-layer GAT networks, and the last five layers are all fully connected layers. The Mask mechanism is used at the network terminal to filter the tasks corresponding to the nodes that violate the constraints.

[0059] Furthermore, the method for training the graph attention network model is a network model training method based on the PPO strategy. The PPO strategy refers to an Actor-Critic learning framework, in which the Actor network is responsible for generating actions and interacting with the environment, and the Critic network is responsible for evaluating the Actor network performance and performing error feedback. Finally, each learning iteration will update the Critic network by copying the Actor network, and the two will continuously interact and improve to achieve the optimization of network parameters.

[0060] Furthermore, the node feature vector and edge feature matrix F1=8, F2=5,the transmission process in the network layer is:

[0061] (1) Embedded layer network transmission (l∈[1,4]), l represents the network layer identifier l∈[1,9]∧l∈N + ;

[0062] The node feature data v is transmitted in the embedded layer network using equations (12), (13) and (14) in sequence, where The conditions shown in formula (15) are met, and the ReLU function is used between layers for activation. The edge feature data E is transmitted in the embedding layer network using formula (16), where The conditions shown in equations (17) and (18) are met, and the ReLU function is also used for activation between layers.

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] F3 is the dimension of the graph attention network structure without the intermediate network layer structure at both ends. Training parameter vector for nodes in layer l; is the edge training parameter vector of the lth layer;

[0071] is the graph attention weight, is the graph attention weight of the lth layer;

[0072] (2) Intermediate layer and hidden layer network transmission, l∈[5,8]

[0073] The middle layer and hidden layer are all fully connected layers, and the dimensions of input and output are both F3. The data transmission adopts the method shown in formula (19), and the ReLU function is also used between layers for activation;

[0074]

[0075] (3) Output layer network transmission (l = 9)

[0076] The output layer is also a fully connected layer, the output dimension is 1, and its data transmission adopts the method shown in formula (20);

[0077]

[0078] The present invention also provides a single-satellite autonomous task scheduling system for non-time-sensitive mobile target tracking, comprising the following modules:

[0079] Acquisition module: used to collect non-time-sensitive mobile target task sets;

[0080] An optimization model building module is used to build a single-satellite autonomous task scheduling optimization model according to the non-time-sensitive task set;

[0081] Solving module: used for solving the single-satellite autonomous task scheduling optimization model to obtain a single-satellite autonomous task scheduling solution;

[0082] Output module: output the single-satellite autonomous task scheduling plan.

[0083] By adopting the above technical solution, the present invention has the following beneficial effects:

[0084] The present invention provides a single-satellite autonomous task scheduling method and system for non-time-sensitive mobile target tracking. The single-satellite autonomous task scheduling model is constructed, and then the tasks are regarded as nodes, and the single-satellite autonomous task scheduling model is converted into a time-pose adjacency graph model under various transient states. The graph attention network model is used to predict the next task node, and the scheduling solution sequence is obtained through step-by-step prediction. In order to use the graph attention network model, the node feature vector and the edge feature matrix are also extracted, and the graph attention network model is trained using the PPO strategy. The method of solving using the graph attention network model improves the solution efficiency, so that a scheduling solution can be quickly given. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a flow chart of the system of the present invention;

[0086] Figure 2 It is a schematic diagram of the adjacency graph model based on time-pose coordinates;

[0087] Figure 3 To construct a sequential solution for the SSATSP problem based on the transient time-pose adjacency graph;

[0088] Figure 4 Schematic diagram of edge feature matrix and node feature vector;

[0089] Figure 5 This is a schematic diagram of the GAT network structure for the single-satellite autonomous task scheduling problem;

[0090] Figure 6 A framework for solving the SSATSP problem. DETAILED DESCRIPTION

[0091] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0092] The single-satellite autonomous task scheduling problem for non-time-sensitive mobile target tracking belongs to a class of combinatorial optimization problems under complex constraints. Its purpose is to maximize the total benefit of satellite observation tasks by quickly selecting imaging tasks and scheduling task imaging time to form a task scheduling sequence as the satellite's imaging plan. The task comes from the task assigned to a single satellite by multi-satellite collaborative tasks, which contains multi-attribute information such as task priority, imaging duration, visible time window and attitude. In order to quickly respond to the dynamic uncertainty of mobile targets, it is necessary to achieve fast and efficient scheduling of assigned tasks. Therefore, it is the best choice to solve the single-satellite autonomous task scheduling problem (SSATSP) problem by using the advantages of fast speed and good effect of machine learning methods. Since multiple tasks can be regarded as multiple graph nodes and the relationship between tasks can be regarded as weighted edge relationships in the graph, SSATSP can be converted into a graph problem and solved in combination with a graph neural network, but the key is to build a suitable and reasonable graph model. The present invention models the SSATSP problem for non-time-sensitive mobile target tracking as a class of combinatorial optimization problems with time-dependent constraint characteristics and then introduces a time-pose adjacency graph model to form a graphical model description of the problem. In order to construct a suitable model, the present invention makes the following assumptions:

[0093] 1) The satellite carries visible light payloads and SAR payloads. Regardless of the type of payload, the platform has agile three-axis attitude maneuvering capabilities and can form an observation window for the target.

[0094] 2) Only one payload can be working on the same satellite at the same time, and the SAR payload can only be in one imaging mode at any time.

[0095] 3) Each orbit of the satellite is a scheduling cycle. During the operation of the satellite, the satellite can maintain a certain charging rate. Under the premise of ensuring a certain power utilization ratio, it can ensure that the power is restored to its original state in the next orbit.

[0096] 4) Ignore satellite storage constraints. After the satellite has image recognition capabilities, it can choose not to send back worthless images without reference information and directly erase them. At the same time, for moving targets, their situation description information and observation thumbnails can be extracted and sent back. The amount of data is much smaller than the traditional high-resolution data images sent back. The satellite can use the measurement and control window during the transit to directly send back effective information and release storage. Therefore, the storage constraint will no longer be a tight constraint.

[0097] 5) Satellite imaging and attitude maneuvers will consume electricity, and the electricity consumption is proportional to the imaging time and attitude maneuvering time respectively.

[0098] 6) The satellite uses uniform ground speed imaging to image the target. The observation time of the mission is proportional to the length of the generated strip (grid side length) of the mission, which has been taken into account when the mission is generated.

[0099] 7) Ignore the impact of the satellite’s observation angle of the target on the imaging resolution. Since the accuracy of on-board image recognition depends on the imaging resolution, but the impact of the imaging angle on the imaging resolution is less than that of the payload working mode, and in the mission preprocessing, the satellite’s imaging observation angle of the target can be limited by window clipping to meet the requirements of on-board image recognition, this study does not consider the impact of the imaging observation angle on the imaging resolution.

[0100] Figures 1 to 6 A specific embodiment of a single-satellite autonomous task scheduling method for non-time-sensitive mobile target tracking according to the present invention is shown. Figure 1 As shown, the following steps are included:

[0101] Step 1: Collect non-time-sensitive mobile target task sets;

[0102] Step 2: construct a single-satellite autonomous task scheduling optimization model according to the non-time-sensitive mobile target task set;

[0103] Objective function:

[0104] Constraints:

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113]

[0114]

[0115] Among them, formula 1 represents the maximum priority sum of completed tasks, x ij represents a 0-1 decision variable, x ij=1 means that task i is the immediate predecessor of task j in the scheduling sequence solution; pri j represents the priority of task j, n tsk Indicates the number of tasks to be planned;

[0116] Formula 2 represents the hard time window constraint of the satellite, that is, the mission must observe within the satellite visible time window; wb i represents the start time of the satellite's visible time window for mission i, wb0 = 0, tb i represents the start time of satellite observation of mission i, tb0=0 represents the initial state time of satellite, te i represents the end time of satellite observation of mission i, te0=0 represents the initial state of the satellite; we i Indicates the end time of the satellite's visible time window for mission i, wb0 = 0;

[0117] Formula 3 represents the equation relationship between the imaging start, end and duration of the task; ct i Indicates the continuous observation time of the satellite on mission i, ct0 = 0;

[0118] Equation 4 represents the time-dependent transition time constraint between tasks; trans(ea i ,ba j ) represents the posture transition time between task i and task j;

[0119] Formula 5 represents the satellite's storage constraint, that is, the power consumption in a single orbit cannot exceed the satellite's remaining power threshold; uie represents the power consumed per unit time during imaging, unit unit, ute represents the power consumed per unit time during attitude maneuvers, unit unit; ζ represents the ratio of the satellite's remaining power threshold;

[0120] Egymax represents the maximum total amount of electricity of the satellite, in unit;

[0121] ea i represents the satellite's final observation attitude for mission i, which is determined by the side swing θ i ,te i , pitch φ i ,te i With yaw ψ i ,te i The three-axis attitude vector is composed of ea0 and ba. i represents the initial observation attitude of the satellite for mission i, which is determined by the side swing θ i ,tb i , pitch φ i ,tb i With yaw ψ i ,tb iThe three-axis attitude vector is composed of ba0, which is the initial attitude of the satellite;

[0122] Formula 6 indicates that each task has at most one immediate predecessor task;

[0123] Formula 7 indicates that each task has at most one subsequent task;

[0124] Formula 8 indicates that a task cannot be its immediate predecessor or immediate successor and can only be completed once; Formula 9 indicates the value range of the decision variable;

[0125] Formula 10 shows the calculation method of the shortest posture transition time between two postures, ρ ij represents the attitude conversion angle of the satellite between mission i and mission j; b0, b1, b2, b3, b4, a0, a1, a2, a3, a4, z0, z1, z2, z3 are preset constants;

[0126] Formula 11 shows the calculation method of the conversion angle between two postures, θ i , t represents the satellite's observed side swing angle for mission i at time t, te i represents the end time of satellite observation of mission i, te0=0 represents the initial state of the satellite, φ i , t represents the satellite's observed elevation angle of mission i at time t, ψ i , t represents the observed yaw angle of the satellite for mission i at time t.

[0127] Step 3: Solve the single-star autonomous task scheduling optimization model to obtain a single-star autonomous task scheduling solution. The solution method in this embodiment is to regard each target task in the single-star autonomous task scheduling optimization model as a graph node, and the relationship between tasks as a weighted edge relationship in the graph, construct multiple transient time pose adjacency graph models, and solve the time pose adjacency graph model to obtain a solution to the single-star autonomous task scheduling optimization model. The method for solving the time pose adjacency graph model is to use the graph attention network model structure to decide the next task node according to the transient time pose adjacency graph, and after obtaining the next task node, update the transient time pose adjacency graph to make a decision on the next task node, and so on, to obtain a scheduling solution sequence.

[0128] In this embodiment, the satellite has an observation window for each transit mission. When the satellite passes directly above the target (overpass moment), the satellite's observation pitch angle of the target is 0. The construction method of the time-pose adjacency graph model is to use the satellite's timeline as the horizontal axis, the side swing attitude angle at the overpass moment as the vertical axis, and the point (0,0) where both the overpass time and the overpass side swing are 0 as the reference reference point, such as Figure 2 As shown in the coordinate system, each task is regarded as a node in the graph, and the coordinates corresponding to each node are It is composed of the satellite overpass time and the corresponding side swing angle. Each node has the attribute of the task corresponding to the node. The edge weight between node i and node j is composed of the attitude conversion time trans(ea i ,ba j ) is represented by . After constructing the time-pose adjacency graph model, the optimization goal of the single-satellite autonomous scheduling optimization problem is transformed into finding a path starting from virtual node 0 that satisfies all constraints and maximizes the priority of all nodes on the path.

[0129] Since the temporal pose adjacency graph model is static, it cannot well describe the time-dependent conversion time constraint characteristics in the problem. Therefore, the single-satellite autonomous scheduling optimization problem is transformed into a temporal pose adjacency graph under multiple transient states. At each transient point in time, the time of the decided nodes in the temporal pose adjacency graph model is fixed. The task nodes that need to be selected are decided according to the transient temporal pose adjacency graph, and the sequential solution of the problem is constructed in sequence. Figure 3 As shown, in each decision, the next task node can be decided according to the instantaneous time-pose adjacency graph and the optimal decision network. After the decision of the next task node is obtained, the next instantaneous time-pose adjacency graph is updated to make a decision on the next task node, and so on. In this way, the optimal scheduling sequence solution {0, 2, 3, 5, 6, 7} of the problem can be constructed.

[0130] In this embodiment, the method of using the graph attention network model structure to decide the next task node is:

[0131] 1) Collect graph features in the current state, including node feature vectors and edge feature matrices, as shown in FIG4 .

[0132] In order to use the graph attention network model for solving, it is necessary to extract the feature vector of the time-pose adjacency graph model. In this embodiment, the node feature vector V and the edge feature matrix E are extracted. The node feature vector contains all the nodes, and each node has 8 attribute features pri i , ct i , wb i , we i , Edge features refer to the relationship between two nodes. The edge feature matrix E has five attribute features d ij ,

[0133] 2) Normalize the node feature vector and edge feature matrix. In this embodiment, in order to enhance the generalization ability of the graph attention network model and avoid the weakening or failure of the network decision effect caused by data distribution differences, the node feature vector and edge feature matrix are normalized. Each element in the normalized edge feature matrix E satisfies The time-position adjacency graph is a complete graph without self-loops. The E matrix does not have diagonal elements. Each node has independent node features. All node features can construct a node feature vector v. The elements of each dimension of the node feature vector satisfy Node attribute feature pri i represents the priority attribute of node i; the priority attribute characterizes the importance of the task and is also a measure of the maximum benefit of system optimization. Normalizing it yields pri max The highest priority.

[0134] Node attribute characteristics represents the time it takes for node i to pass the task. As one of the dimensions of the time-pose coordinate, the time can measure the attribute characteristics of the task. Normalizing it yields for:

[0135] T plan Indicates the duration of a single planning cycle.

[0136] Node attribute characteristics Represents the overhead lateral swing angle of the task corresponding to node i; as one of the time-pose coordinates, the overhead lateral swing can represent the target position and affect the posture conversion calculation.

[0137] θ max Indicates the satellite's maximum lateral attitude angle.

[0138] Node attribute feature ct i represents the continuous observation time of the task corresponding to node i; the continuous observation time can measure the resource consumption of the task and is included as one of the node characteristics. After normalization, it is expressed as ct i ∈[ct min , ct max ], ct min ,ct max are the shortest and longest continuous observation time of the task, respectively.

[0139] Node attribute features wb i Represents the visible time window start time of the task corresponding to node i; the visible window start time represents the time when the task can be observed, which is of great significance for the selection and scheduling of tasks. Its normalized expression is

[0140] Node attribute features i Represents the end time of the visible time window of the task corresponding to node i; the end time of the visible window represents the latest time when the task can be observed, which is of great significance for the selection and scheduling of tasks. After normalization, it is expressed as

[0141] Node attribute characteristics Indicates whether the task corresponding to node i is the task to be selected in the current state; for each graph state, the next selected task needs to be decided through the network. The influence of task selection lies in the ability to determine whether a task is among the candidate tasks. When the task corresponding to node i is among the candidate tasks, otherwise No normalization is required.

[0142] Node attribute characteristics Indicates whether the task corresponding to node i is the last task in the currently scheduled sequence; used to indicate whether the task corresponding to node i is the task corresponding to the current decision state. As an important state attribute of the problem, it plays a key role in promoting the construction of sequential solutions. Is it the last task in the currently scheduled sequence? If so, otherwise, No normalization is required.

[0143] Edge feature d ij Represents the distance between the two nodes associated with the edge in the time-pose coordinates; to characterize the influence of this feature on the time constraint of time-dependent posture transition, for feature d ij Manhattan distance is also used for calculation.

[0144]

[0145] Among them, tw max is the maximum time window length of the satellite to the target, is the maximum elevation angle of the satellite, for feature d ij Normalization can obtain normalized features as follows:

[0146]

[0147] Among them, θ max Satellite maximum sway angle, T plan The mission planning period is usually set to the duration of a single satellite orbit.

[0148] Indicates whether the neighbor node j is the first node in ascending order of distance from node i under the time coordinate; that is, for k∈N i , N i is the set of all neighbor node numbers of node i, for all d ik Sort in ascending order, if d ij If it is ranked first, otherwise, No normalization is required.

[0149] Indicates whether the neighbor node j is the node identifier within the fifth node in ascending order of distance from node i under the time coordinate; that is, for k∈N i , N i is the set of all neighbor node numbers of node i, for all d ik Sort in ascending order, if d ij In the top five, otherwise, No normalization is required.

[0150] Indicates whether the neighbor node j is the node identifier within the tenth node in ascending order of distance from node i under the time coordinate; that is, for k∈N i , N i is the set of all neighbor node numbers of node i, for all d ik Sort in ascending order, if d ij In the top ten, otherwise, No normalization is required.

[0151] Indicates whether the neighbor node j is a node identifier within 20 in ascending order of distance from node i under the time coordinate. That is, for k∈N i , N i is the set of all neighbor node numbers of node i, for all d ik Sort in ascending order, if d ij In the top 20, otherwise, No normalization is required.

[0152] 3) Input the normalized node feature vector and edge feature matrix into the graph attention network structure to obtain the node selection probability of the next task node, and select the node with the maximum probability as a node on the solution sequence. Figure 5As shown in the figure, the graph features, namely the node feature vector and the edge feature matrix, extracted from the temporal adjacency graph in the current state are input into the graph attention network structure, and the Mask mechanism is used to avoid the impossible actions at the network terminal. Finally, the score of the action is output to obtain the selection probability of the node, and then the node with the largest probability is used as a node on the solution sequence.

[0153] from Figure 5 It can be seen that the graph attention network structure is a 9-layer deep learning network, which is a typical deep learning network. The first four layers are embedding layers, all of which use single-layer GAT networks, and the last five layers are all fully connected layers. The Mask mechanism is used at the network terminal to screen the tasks corresponding to the nodes that violate the constraints. The edge and node attributes are weighted and the new node attribute features are updated through the attention mechanism. It should be noted that the updated node attribute dimension is inconsistent with the extracted input feature dimension. The last five layers are all fully connected layers, which are only responsible for updating the graph node features. Among them, the fifth layer is the middle layer, which is responsible for the conversion of the network dimension. The sixth to eighth layers are hidden layers, whose dimensions remain consistent, and the last layer is the output layer, whose output dimension is 1. The feature update of the entire network to the node is independent of the graph structure, which means that only the features are weighted, so the decision network learned is not affected by the scale of the graph nodes, and the number of network parameters obtained by training graphs of different node scales is equal, which highlights the advantages of the graph attention network in generalization ability.

[0154] For the input node feature vector and edge feature matrix F1=8, F2=5. In order to avoid complicating the network, the GAT network structure adopted by the present invention keeps the dimensions of the intermediate network layer structures at both ends consistent, unified as F3. The transmission process of the extracted feature data in the network layer is as follows:

[0155] (1) Embedded layer network transmission (l∈[1,4]), l represents the network layer identifier l∈[1,9]∧l∈N + ;

[0156] The node feature data v is transmitted in the embedded layer network using equations (12), (13) and (14) in sequence, where is the node training parameter vector of the lth layer, which satisfies the conditions shown in formula (15), and the ReLU function is used for activation between layers. The edge feature data E is transmitted in the embedding layer network using formula (16), where is the edge training parameter vector of the lth layer, satisfying the conditions shown in equations (17) and (18),

[0157] The ReLU function is also used between layers for activation.

[0158]

[0159]

[0160]

[0161]

[0162]

[0163]

[0164]

[0165] F3 is the dimension of the graph attention network structure without the intermediate network layer structure at both ends;

[0166] is the graph attention weight, is the graph attention weight of the lth layer;

[0167] (2) Intermediate layer and hidden layer network transmission (l∈[5,8])

[0168] The middle layer and hidden layer are all fully connected layers, and the dimensions of input and output are both F3. The data transmission adopts the method shown in formula (19), and the ReLU function is also used between layers for activation;

[0169]

[0170] (3) Output layer network transmission (l = 9)

[0171] The output layer is also a fully connected layer, the output dimension is 1, and its data transmission adopts the method shown in formula (20);

[0172]

[0173] In this embodiment, the Mask mechanism is used at the network terminal to filter the tasks corresponding to the nodes that violate the constraints, so as to avoid infeasible action selections. That is, for the tasks corresponding to the nodes that violate the constraints, the Mask mechanism is used to control their output scores to zero. i , when deciding the next node, if the task corresponding to node i violates the constraint, then m i =0, otherwise m i = 1. If the final output node i's score is The Mask mechanism can be implemented in sequence through the methods shown in equations (21), (22), (23) and (24).

[0174]

[0175]

[0176]

[0177]

[0178] In this embodiment, the method for training the graph attention network model is a network model training method based on the PPO strategy. The PPO strategy refers to an Actor-Critic learning framework, in which the Actor network is responsible for generating actions and interacting with the environment, and the Critic network is responsible for evaluating the performance of the Actor network and performing error feedback. Finally, each learning iteration will update the Critic network by copying the Actor network, and the two will continuously interact and improve to achieve the optimization of network parameters. For the SSATSP problem of non-time-sensitive mobile target tracking, there is currently a lack of standard supervised learning data, and the mining of the essential laws of the problem can only be achieved through unsupervised learning. Reinforcement learning (RL), as a popular machine learning method in the current computing era, is a typical unsupervised learning method that can complete data self-generation, self-collection, and model self-training and self-improvement. It is very suitable for solving the SSATSP problem of non-time-sensitive mobile target tracking for serialized modeling. The present invention adopts an efficient reinforcement learning method to realize the training of the GAT decision network, obtain the optimal decision network to achieve effective solution to the problem.

[0179] like Figure 6 The SSATSP problem-solving framework shown in the figure covers the entire process from problem instance acquisition, i.e. task set acquisition, modeling, i.e. building a model based on the task set, feature extraction, i.e. extracting task features, network design, network model training, network-based solution to solution output, with the focus on network-based solution and network model training. Network-based problem solving is a process of sequential solution construction, which realizes the construction of scheduling sequence solution by cyclically calling the optimal decision network. This process is consistent with the process of collecting learning data in each episode during network training.

[0180] In this embodiment, when the graph attention network model is trained, the input is the graph feature vector of the time-pose adjacency graph in each transient state, that is, the edge feature matrix and the node feature vector {E, v}. The state of the graph attention network refers to the feature vector set of the time-pose adjacency graph, and each element in the set corresponds to a state, that is, the graph feature vector in each transient state, that is, the edge and node features {E, v} of the extracted time-pose adjacency graph. Actions refer to all selectable decision objects in a certain state. In the present invention, an action corresponds to a node in the time-pose adjacency graph, and is also a task in the satellite mission planning. State transition refers to the process in which a decision to select an action in a certain state causes the state of the problem to change. In the present invention, the problem directly transfers to the next state after selecting a task. Reward refers to the change in systematic benefits brought about by the decision to select an action in a certain state. From the perspective of single-step decision-making, the reward of the action is the priority benefit of the task corresponding to the selected node, which can be multiplied by the discount factor. Strategy refers to the method of selecting an action after the score of the node to be selected is calculated through the decision network. There are two action selection strategies in the present invention. In the network training process, sampling selection based on output probability or roulette selection is adopted. When the network is used for solving, the greedy rule is adopted to select the node with the largest output probability score.

[0181] Step 4: Output the single-satellite autonomous task scheduling plan.

[0182] The present invention also provides a single-satellite autonomous task scheduling system for non-time-sensitive mobile target tracking, comprising the following modules:

[0183] Acquisition module: used to collect non-time-sensitive mobile target task sets;

[0184] An optimization model building module is used to build a single-satellite autonomous task scheduling optimization model according to the non-time-sensitive task set;

[0185] Solving module: used for solving the single-satellite autonomous task scheduling optimization model to obtain a single-satellite autonomous task scheduling solution;

[0186] Output module: output the single-satellite autonomous task scheduling plan.

[0187] The present invention detects the effect of training and analyzes the convergence of training from the analysis experiment of the training process. When it reaches 1000 episodes, the change of network benefits begins to flatten and is basically stable in the future. This shows that the GAT network converges quickly under the training of the PPO algorithm and the training effect is significant. After 40,000 episodes of training, the network basically converges stably and can obtain a decision network that tends to be optimal.

[0188] In the feasibility analysis of the algorithm, the main test is to test whether the solution obtained by the GAT network decision is feasible and whether it violates the satellite's constraints through example testing. To achieve this goal, the experiment also conducts solution traceability analysis for examples with task scales of 40, 60, 80 and 100. The experiment conducts feasibility analysis from the visualization of the solution, including two visualization graphs: scheduling sequence graph and time-pose adjacency graph. The scheduling sequence graph uses the satellite running time as the x-axis and the task ID as the y-axis, and can display information such as the task window, priority attributes, and the sequence and duration of the scheduled task. In the task scheduling sequence graph, the node route map represented by the sequence cannot be crossed, because the satellite cannot perform two tasks at the same time, and the tasks cannot be seamlessly connected because there is attitude conversion time. The time-pose adjacency graph uses time as the x-axis and the overhead side swing angle as the y-axis. The tasks in the time-pose adjacency graph exist in a single coordinate and cannot display the window. Since the time-pose adjacency graph can only display coordinate information, it has a certain coordinate mapping relationship with the observation path in the real world, which can well display the satellite's observation path, and secondly, it can also display the amplitude of the satellite's attitude change. In the time-pose adjacency graph, the observation paths can overlap, because its time is only a static coordinate and cannot represent the task execution time. The time-pose adjacency graph can be used to determine whether the satellite has unreasonable large-angle maneuvers and lacks consideration for the tasks on the maneuvering path. Under the four scales, there is no overlap in the scheduling sequence route, each task has a certain execution time, and there are different attitude conversion times between tasks. In the scheduling results, the satellite tends to complete tasks with higher task priority. From the time-pose adjacency graph, it can be found that the observation paths executed by the satellite overlap. This is because the tasks have longer observation windows and can complete observations at non-vertex points. In addition, the satellite can take into account other tasks on the path during large attitude maneuvers to avoid unreasonable large-angle attitude maneuvers. From the perspective of the two graphs, it can be seen that the scheduling of the problem by the GAT network is feasible, and it can make good use of limited satellite resources to achieve the highest observation benefits. In addition, for the scheduling under the four scales, in the absence of GPU acceleration in a personal notebook, the scheduling time of the four scales is 0.65s, 1.16s, 1.98s and 3.12s respectively. For an NP-hard problem, the algorithm exhibits excellent scheduling performance in the presence of an exponentially explosive solution space, making it possible to achieve fully autonomous and automatic scheduling on board. The effectiveness of the algorithm will be further analyzed in the future.

[0189] The present invention constructs a single-satellite autonomous task scheduling model, then regards the tasks as nodes, converts the single-satellite autonomous task scheduling model into a time-pose adjacency graph model under various transient states, uses a graph attention network model to predict the nodes in the next state, and obtains a scheduling solution sequence through step-by-step prediction. In order to use the graph attention network model, the node feature vector and edge feature matrix are also extracted, and the graph attention network model is trained using the PPO strategy. The method of solving using the graph attention network model improves the solution efficiency, so that a scheduling solution can be quickly given.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A single-satellite autonomous task scheduling method for non-time-sensitive mobile target tracking, characterized in that: The following steps are involved: Step 1: Obtain a non-time-sensitive mobile target task set; Step 2: construct a single-satellite autonomous task scheduling optimization model according to the non-time-sensitive task set; Step 3: Solving the single-satellite autonomous task scheduling optimization model to obtain a single-satellite autonomous task scheduling solution; Step 4: Output the single-satellite autonomous task scheduling plan; The single-satellite autonomous task scheduling optimization model is: Objective function: (1) Constraints: （2） （3） （4） （5） （6） （7） （8） （9） （10） （11） Among them, formula 1 represents the maximum priority sum of completed tasks, represents a 0-1 decision variable, Represents the tasks in the scheduling sequence solution For the task The immediate tasks; Indicates the task The priority of Indicates the number of tasks to be planned; Formula 2 represents the hard time window constraint of the satellite, that is, the mission must observe within the satellite visible time window; Satellite mission The start time of the visible time window, , Satellite mission The start time of observation, Indicates the satellite initial state time, express Satellite pair mission The end observation time, Indicates the initial state of the satellite; express Satellite pair mission The visible time window end time, ; Equation 3 represents the equation relationship between the imaging start, end and duration of the task; Satellite mission The continuous observation time, ; Equation 4 represents the time dependency transition time constraint between tasks; Indicates the task With the task j The time between posture transitions; Formula 5 represents the satellite's storage constraint, that is, the power consumption in a single orbit cannot exceed the satellite's remaining power threshold; uie Indicates the amount of power consumed per unit time during imaging, in units of unit, ute It indicates the amount of electricity consumed per unit time when performing attitude maneuvers, in units of unit; ζ Indicates the satellite remaining power threshold ratio; Egy max represents the maximum total power of the satellite, in unit; Satellite mission The end observation posture is swung sideways ， , pitch , With Yaw , Three-axis attitude vector composition, is the initial attitude of the satellite, Satellite mission The starting observation posture is from side swing , , pitch , With Yaw , Three-axis attitude vector composition, is the initial attitude of the satellite; Formula 6 indicates that each task has at most one immediate predecessor task; Formula 7 indicates that each task has at most one subsequent task; Formula 8 indicates that a task cannot be its immediate predecessor or immediate successor and can only be completed once; Formula 9 indicates the value range of the decision variable; Formula 10 shows the calculation method of the shortest posture transition time between two postures. Indicates that the satellite is on mission With the task The angle of posture transition between 、 、 、 、 、 、 、 、 、 、 、 、 、 are preset constants; Formula 11 shows the calculation method of the conversion angle between two postures, , Indicates the satellite at time To the task The observed side swing angle, tei Satellite mission The end observation time, te 0 = 0 means the satellite is in the initial state. , Indicates the satellite at time The observed pitch angle for the mission, , Indicates the satellite at time To the task The observed yaw angle.

2. The scheduling method according to claim 1, characterized in that: The method for solving the single-satellite autonomous task scheduling optimization model in step 3 is: regard each target task in the single-satellite autonomous task scheduling optimization model as a graph node, regard the relationship between tasks as a weighted edge relationship in the graph, construct multiple transient time-pose adjacency graph models, solve the time-pose adjacency graph model, and obtain the solution of the single-satellite autonomous task scheduling optimization model.

3. The scheduling method according to claim 2, characterized in that: The method for constructing the time-pose adjacency graph model is to use the timeline of the satellite operation as the horizontal axis, the side swing attitude angle at the time of passing the top as the vertical axis, and the point (0 , 0) is the reference point. In the coordinate system, each task is regarded as a node in the graph. The coordinates corresponding to each node are It is composed of the satellite passing time and the corresponding side swing angle. Each node has the attribute of the task corresponding to the node. With Node j The edge weights between tasks are determined by With the task j The time between attitude transitions express.

4. The scheduling method according to claim 2, characterized in that: The method for solving the time-pose adjacency graph model is to use the graph attention network model structure to decide the next task node based on the transient time-pose adjacency graph, and after obtaining the next task node, update the transient time-pose adjacency graph to make a decision on the next task node, and so on, to obtain a scheduling solution sequence.

5. The scheduling method according to claim 4, characterized in that: The method of using the graph attention network model structure to decide the next task node is: Collect graph features in the current state, including node feature vectors and edge feature matrices; Normalizing the node feature vector and edge feature matrix; The normalized node feature vector and edge feature matrix are input into the graph attention network structure to obtain the node selection probability of the next task node, and the node with the largest probability is selected as a node on the solution sequence.

6. The scheduling method according to claim 5, characterized in that: Each node in the node feature vector has 8 feature attributes: , , , , , , , ; Each edge has 5 characteristic attributes: , , , , ; in, Representation Node The priority attribute of Representation Node The overhead time of the corresponding task; Representation Node The overhead lateral swing angle corresponding to the task; Representation Node The continuous observation time of the corresponding task; Representation Node The start time of the visible time window of the corresponding task; Representation Node The end time of the visible time window of the corresponding task; Representation Node An indicator of whether the corresponding task is the task to be selected in the current state; Representation Node An indicator of whether the corresponding task is the last task in the currently scheduled sequence; Represents the distance between the two nodes associated with the edge in the time-pose coordinate; Represents neighbor nodes in time coordinates Is it a node? The node ID that is first in ascending distance order; Represents neighbor nodes in time coordinates Is it a node? The node identifiers within the fifth node in ascending distance order; Represents neighbor nodes in time coordinates Is it a node? The node identifiers within the tenth node in ascending order of distance; Represents neighbor nodes in time coordinates Is it a node? Node identifiers within twenty in ascending order.

7. The scheduling method according to claim 4, characterized in that: The graph attention network model structure is a 9-layer deep learning network. The first four layers are embedding layers, all of which use single-layer GAT networks, and the last five layers are all fully connected layers. The Mask mechanism is used at the network terminal to filter the tasks corresponding to the nodes that violate the constraints.

8. The scheduling method according to claim 7, characterized in that: The method for training the graph attention network model is a network model training method based on the PPO strategy. The PPO strategy refers to an Actor-Critic learning framework, in which the Actor network is responsible for generating actions and interacting with the environment, and the Critic network is responsible for evaluating the Actor network performance and performing error feedback. Finally, each learning iteration will update the Critic network by copying the Actor network, and the two will continuously interact and improve to achieve the optimization of network parameters.

9. The scheduling method according to claim 5, characterized in that: The node feature vector and edge feature matrix , F 1 = 8, F 2 = 5, the transmission process in the network layer is: (1) Embedding layer network transmission, , Indicates the network layer identifier. ; The node feature data v is transmitted in the embedded layer network using equations (12), (13) and (14) in sequence, where The conditions shown in formula (15) are met, and the ReLU function is used between layers for activation. The edge feature data E is transmitted in the embedding layer network using formula (16), where , Satisfying the conditions shown in equations (17) and (18), The ReLU function is also used between layers for activation. （12） （13） （14） （15） （16） （17） （18） F 3 is the dimension of the graph attention network structure without the intermediate network layer structure at both ends. For the l The node training parameter vector of the layer; , For the l The edge training parameter vector of the layer; is the graph attention weight, For the l The graph attention weights of the layers; (2) Intermediate layer and hidden layer network transmission, l ∈[5 , 8]; The middle layers and hidden layers are all fully connected layers, and the dimensions of input and output are F 3. Its data transmission adopts the method shown in formula (19), and the ReLU function is also used between layers for activation; （19） (3) Output layer network transmission, l = 9; The output layer is also a fully connected layer, the output dimension is 1, and its data transmission adopts the method shown in formula (20); 。 10. A single-satellite autonomous task scheduling system for non-time-sensitive mobile target tracking, using a single-satellite autonomous task scheduling method for non-time-sensitive mobile target tracking as claimed in any one of claims 1 to 9, characterized in that: Includes the following modules: Acquisition module: used to collect non-time-sensitive mobile target task sets; An optimization model building module is used to build a single-satellite autonomous task scheduling optimization model according to the non-time-sensitive task set; Solving module: used for solving the single-satellite autonomous task scheduling optimization model to obtain a single-satellite autonomous task scheduling solution; Output module: output the single-satellite autonomous task scheduling plan.

Images