Data-driven large-scale space object surveillance daily scheduling method

Abstract

The application relates to a data-driven large-scale space target monitoring daily scheduling method. The method comprises the following steps: establishing an SGORJS mathematical model containing a task allocation and a central scheduling target function and constraints. Taking the information as input, task allocation is performed through a graph convolutional neural network to determine the scheduling center to which an on-orbit target belongs, and an initial task allocation scheme and corresponding allocation probability prior information are output. An HCO algorithm is designed to plan the task execution sequence of each scheduling center, and a central scheduling scheme containing task priority, resource allocation and time window planning is obtained. According to the ASAPS optimization strategy, the initial task allocation scheme and the central scheduling scheme are optimized in combination with the prior information, and the optimized scheme is output to form a final scheduling scheme to guide the on-orbit target observation and scheduling execution. The method can improve the optimization effect and calculation efficiency of the scheduling scheme.

Description

Technical Field

[0001] This application relates to the field of space situational awareness, and in particular to a data-driven daily scheduling method for large-scale space target surveillance. Background Technology

[0002] With the rapid development of space technology, the number of orbiting targets (RSOs) continues to surge, and the space environment is becoming increasingly complex, placing higher demands on the accuracy, comprehensiveness, and response speed of space situational awareness. To maintain the stability of the space environment and enhance space awareness capabilities, the Space Observation Network (SSN) has become the core infrastructure for observing, identifying, and tracking RSOs. Based on their deployment locations, SSN observation resources are divided into two categories: space-based observation resources and ground-based observation resources. Space-based observation resources are deployed on space platforms, are not limited by airspace or the curvature of the Earth, can overcome atmospheric interference, and can observe targets on the other side of the Earth. Ground-based observation resources are deployed on the ground, are technologically mature, have lower costs, and are not constrained by volume or mass; the two types of resources exhibit significant observational complementarity.

[0003] However, in existing technologies, space-based and ground-based observation resources have long operated independently, belonging to different resource scheduling centers, lacking an effective collaborative scheduling mechanism. This leads to the following prominent technical problems: First, the independent scheduling mode cannot fully leverage the complementary advantages of the two types of resources, making it difficult to balance observation coverage, time window utilization, and observation accuracy. Second, when facing multiple concurrent tasks or emergency observation scenarios, the decentralized scheduling system has low response efficiency and cannot meet the needs of rapid and accurate observation. Third, SSN scheduling requires the coordination of multiple resources such as observation, telemetry, and data transmission. The operational constraints and action coupling relationships of different resources are complex, and traditional scheduling methods have not formed a systematic joint modeling and solution framework, resulting in limited optimization space and low computational efficiency for scheduling schemes.

[0004] Meanwhile, orbiting space stations (RSOs) follow a two-body motion law, resulting in relatively stable orbital parameters. The visible time windows for RSO observations can be obtained through orbit propagators and assessment of observation resource capabilities. The core objective of space-based observation (SSN) mission scheduling is to select the optimal window from the visible time windows to complete RSO observations and correct its orbital parameters under various constraints. The scheduling process encompasses key stages such as observation request submission, scheduling plan generation, and observation data processing. Although research has been conducted on SSN resource scheduling, existing results largely focus on optimizing the scheduling of single ground-based resources, neglecting the joint scheduling requirements of space-based and ground-based observation resources. This fails to address the efficient scheduling problem under multi-resource coupling and complex constraints, making it difficult to adapt to the practical application scenarios of large-scale RSO observations.

[0005] Therefore, how to construct a scheduling model that takes into account the coordination of space-based and ground-based resources, design efficient solution algorithms, break down the barriers between the independent operation of the two types of resources, and improve the optimization effect and computational efficiency of the scheduling scheme under the premise of satisfying multiple constraints has become a key technical problem that urgently needs to be solved in the field of space target surveillance. Summary of the Invention

[0006] Therefore, it is necessary to provide a data-driven daily scheduling method for large-scale space target surveillance that can improve the optimization effect and computational efficiency of the scheduling scheme, in order to address the above-mentioned technical problems.

[0007] A data-driven daily scheduling method for large-scale space target surveillance, the method comprising: To acquire resource constraint information of each scheduling center in the space surveillance network, orbital parameters and observation requirements of on-orbit targets, and to establish a mathematical model of the SSNSGORJS problem, the mathematical model includes the objective function and constraints of task allocation and center scheduling. Using resource constraint information, orbital parameters of on-orbit targets, and observation requirements as inputs, a graph convolutional neural network is used for task allocation, determining the scheduling center to which each on-orbit target belongs, outputting an initial task allocation scheme, and using the allocation probability corresponding to the initial task allocation scheme as prior information; The HCO algorithm is designed to plan the task execution sequence of each scheduling center, resulting in a central scheduling scheme. The central scheduling scheme includes the task execution priority, resource allocation strategy, and time window planning of each scheduling center. Based on the ASAPS optimization strategy and prior information, the initial task allocation scheme and the central scheduling scheme are jointly optimized, and the optimized task allocation scheme and central scheduling scheme are output. The optimized task allocation scheme is integrated with the central scheduling scheme to form the final scheduling scheme, which is used to guide the on-orbit target observation and scheduling execution of the space surveillance network.

[0008] The aforementioned data-driven daily scheduling method for large-scale space target surveillance first obtains information on the resource constraints of the scheduling center, the orbital parameters of on-orbit targets, and observation requirements, providing complete data support for modeling. The constructed SGORJS mathematical model simultaneously covers the objective functions and constraints of task allocation and center scheduling, accurately depicting the coupling relationship of multiple resources and avoiding the optimization limitations caused by information gaps or one-sided modeling in traditional models. Secondly, a graph convolutional neural network is used for task allocation. Leveraging its ability to learn relationships in the feature space, it quickly determines the scheduling center to which the target belongs. The output initial scheme and prior information on allocation probabilities lay an accurate foundation for subsequent optimization, reducing blind searches. The HCO algorithm is designed to efficiently solve task conflicts and resource coordination problems within a single scheduling center by hierarchically planning task execution priorities, resource allocation strategies, and time windows, improving local scheduling efficiency. Simultaneously, the ASAPS optimization strategy, combined with prior information, collaboratively optimizes the initial task allocation and center scheduling scheme. Through mechanisms such as adaptive neighborhood selection and tabu lists, it avoids local optima and accelerates convergence, further optimizing the overall scheme. Finally, by integrating and optimizing the two types of schemes, the solution forms the globally optimal final scheduling scheme. It exerts efforts throughout the entire process from data foundation, modeling accuracy, algorithm efficiency to global optimization, effectively breaking through the barriers of traditional independent scheduling and significantly improving the optimization effect and computational efficiency of the scheduling scheme. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a data-driven daily scheduling method for large-scale space target surveillance in one embodiment. Figure 2 This is a diagram illustrating a general process for space target tracking in one embodiment; Figure 3 This is a diagram of a space target tracking joint observation framework based on the Holonic organization model in one embodiment; Figure 4 Here is a flowchart of the DTJS method in one embodiment; Figure 5 Here is a diagram of the GCN network structure in one embodiment; Figure 6 Here is a flowchart of the HCO algorithm in one embodiment; Figure 7 Here is a flowchart of the ASAPS optimization strategy in one embodiment; Figure 8 This is a diagram showing random variations in one embodiment. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] In one embodiment, such as Figure 1 As shown, a data-driven daily scheduling method for large-scale space target surveillance is provided, including the following steps: Step 102: Obtain resource constraint information, orbital parameters and observation requirements of on-orbit targets from each scheduling center in the space surveillance network, and establish a mathematical model of the SSNSGORJS problem. The mathematical model includes the objective function and constraints of task allocation and center scheduling.

[0012] from Figure 2 It can be seen that completing a full RSO observation process requires collaboration across multiple stages, involving different resources, each with its own usage constraints. Different types of resources are mutually restrictive and influential. For example, space-based resource action commands need to be transmitted through telemetry and control resources, and space-based resources need to issue data transmission commands after observing the target. Therefore, the joint planning problem of space-ground and space-based resources in space target tracking not only requires the rational arrangement of observation plans for space and ground-based sensors but also considers telemetry, control, and data transmission plans. The coupling between different types of resources greatly increases the complexity of collaborative planning, making it difficult to directly model the problem using combinatorial optimization theory. Therefore, to improve the universality and practicality of the model, this application analyzes the hierarchical structure of the space target tracking system, then maps it to a Holonic organization model and performs formal modeling.

[0013] Currently, SSN's space-based and ground-based observation resources are executed separately (belonging to space-based and ground-based centers respectively), and their resource planning systems are independent and autonomous, belonging to a decentralized hierarchical structure. The advantage is a simple structure and relatively low difficulty in planning models and solutions. However, with the continuous development of equipment, completing space target observation tasks requires requests to different observation platforms, making user coordination difficult. Another approach is a fully centralized structure, where space / ground-based resources are managed uniformly by a single center. This structure offers strong equipment resource synergy and high processing efficiency. However, due to significant differences in the operation modes, observation modes, usage constraints, and data transmission between space-based and ground-based resources, unified modeling is difficult, making model solving challenging. The Holonic organization model is an organizational management model based on self-organization and distributed control. It emphasizes distributed decision-making and self-organization capabilities. Each Holon can make decisions based on local information without centralized command and control. This distributed decision-making and control mechanism can improve the organization's flexibility, adaptability, and efficiency. To overcome the shortcomings of decentralized / centralized hierarchical structures, this application proposes a joint observation framework for RSO tracking space-ground-based resources based on the Holonic organization model, such as... Figure 3 .according to Figure 3 As can be seen, this framework involves the control center coordinating the scheduling of space / ground-based resources, while the specific scheduling optimization of resources is completed by their respective space / ground-based centers. Space / ground-based centers delegate some control of their respective observation resources to the control center for unified management, yet maintain relative independence, allowing adjustments based on specific payload constraints. Therefore, the Holonic-based joint observation framework for space / ground-based resources can fully utilize existing resources, shielding the complex usage methods and constraint types of different types of observation resources, enabling the control center to coordinate the planning of space / ground-based resources. Through the Holonic framework, the entire RSO observation process achieves coordinated resource utilization, distributed task execution, and centralized data processing, thereby improving space situational awareness. This organizational structure provides flexibility and scalability, enabling the system to adapt to different mission requirements and resource changes.

[0014] The SSN (Solar-Signaled Network) space / ground-based resource joint planning involves numerous resources and elements, making unified modeling difficult. To facilitate standardized modeling of the problem, the following assumptions are made: 1) During the observation of RSOs, it is not allowed to be interrupted. The next observation task can only be carried out after the current observation task is completed. 2) Ground observation / control equipment is fixed on the Earth's surface and has an uninterrupted power supply; 3) Without considering the delay between ground transmission links, the "benefit" can be obtained as soon as the space-based resources transmit the observation data of the target. 4) Assume the scheduling cycle is in a static environment, i.e., do not consider the addition of targets or resource corruption; Space-based and ground-based centers are considered as individual units of Holon, complementing each other and working in tandem. Based on the Holon organizational model, a detailed analysis is conducted on the input information, decision variables, optimization objectives, and constraints in the planning problem, using... This indicates the input information. Represents decision variables, Describe the objective function. The constraints are then represented. The mathematical model for the SSN (Solar / Ground-based Resource Joint Planning) can then be expressed as: .

[0015] The input information mainly includes task information, resource information, and window information, denoted as... .

[0016] Task Information ; ; TarID refers to the unique identifier of the target. reqTime refers to the time when the observation request is submitted, Priority refers to the priority of the space target, minDurTime represents the shortest observation time required for the target, minAdiTime represents the shortest time interval between two consecutive observations of the target, and Freq represents the number of times the target needs to be observed within a scheduling cycle.

[0017] Resource Information SpaSen represents space-based observation resources, GroSen represents ground-based observation resources, TelGS represents telemetry and control resources, DldGS represents data transmission resources, and DPGS represents ground processing center resources.

[0018] ; ; Among them, satID is the unique identifier of the satellite, OBT is the orbital element number of the satellite, TfVelocity is the data transmission rate, TdVelocity is the data downlink rate, minTleDur is the shortest time between telemetry, tracking, and command, minDldDur is the shortest data transmission time, and Sensor refers to the sensor carried on the satellite.

[0019] ; These include sensor identification, observation type, rotational angular velocity of space-based observation resources, and set of observation angles.

[0020] ; GSID refers to the unique identifier of the ground-based observation resource. lat is the latitude of the ground-based resource, lon is the longitude of the ground-based resource, MinObvAngle is the minimum elevation angle during ground-based resource observation, TransTime is the device switching time during sensor observation, and DisDetec is the radar detection range when the ground-based resource is a radar device.

[0021] ; telgsID refers to the unique identifier of the ground-based telemetry and control resource. telat is the latitude of the ground-based telemetry and control resource, and telon is the longitude of the ground-based telemetry and control resource. MinteAngle is the minimum elevation angle of the ground-based telemetry and control resource, and teTransTime is the equipment switching time of the telemetry and control resource.

[0022] ; dldgsID refers to the unique identifier of the ground-based data transmission resource. dlat is the latitude of the ground-based data transmission resource, dlon is the longitude of the ground-based data transmission resource, MindAngle is the minimum pitch angle of the ground-based data transmission resource, and dTransTime is the data transmission resource equipment conversion time.

[0023] Time window information Where SpaSenWds is the set of visible windows for space-based resources to space targets. GroSenWds is the set of visible windows for ground-based resources to space targets. TelWds is the set of telemetry and control windows. DldWds is the set of data transmission windows.

[0024] ; ; Where spaNumWds is the window number. , SstartTime is the start time of the space-based observation window, and SendTime is the end time of the space-based observation window.

[0025] ; groNumWds is the window number. , GstartTime is the start time of the ground observation window, and GendTime is the end time of the ground observation window.

[0026] ; ; Where telNumWds is the window number. , TstartTime is the start time of the monitoring and control window, and TendTime is the end time of the monitoring and control window.

[0027] ; dldNumWds is the window number. , startTime is the window start time, and endTime is the window end time.

[0028] The SSN (Space-Based Memory Service) joint scheduling problem involves four decision variables: space-based resource observation actions, ground-based resource observation actions, telemetry and control actions, and data transmission actions, denoted as: .

[0029] 1) Space-based resource observation operations ; in SAcStartTime represents the actual start time of space-based resource observation, and SAcEndTime represents the actual end time of space-based resource observation. , , .

[0030] 2) Ground-based resource monitoring operations , ,GAcStartTime is the actual start time of ground resource observation, and GAcEndTime is the actual end time of ground resource observation. , .

[0031] 3) Space-based resource tracking and control actions , TAcStartTime is the actual start time of the measurement and control, and TAcEndTime is the actual end time of the measurement and control. , .

[0032] 4) Space-based resource data transmission actions ; ,DAcStartTime is the actual start time of data transmission, and DAcEndTime is the actual end time of data transmission. , .

[0033] Therefore, the decision variables can be represented by four Boolean variables. , This indicates that the i-th space-based resource performs its observation task within its corresponding j-th window, and vice versa. . This indicates that the i-th ground-based resource performs the observation task in its corresponding j-th window, and vice versa. . This indicates that the i-th satellite performs telemetry and control tasks within its corresponding j-th window, and vice versa. . This indicates that the i-th satellite performs the data transmission task within its corresponding j-th window, and vice versa. .

[0034] The objective function chosen in this application is to maximize the total observational gain, since each space objective is... Each attribute has a priority; the reward is obtained when the objective is observed (subject to satisfying different constraints). Therefore, the objective function can be expressed as: .

[0035] Given the observation plans for space-based and ground-based resources, and the data transmission and telemetry control plans for space-based resources, the observation benefits are obtained.

[0036] The SGORJS model involves various constraints, which can be mainly categorized into three types: 1. Observability constraints: Only after observability is satisfied can the RSO be "seen" by space / ground-based resources, allowing for the calculation of the visible time window and resource scheduling. These constraints have been mentioned in previous studies and will not be elaborated upon here. 2. Mission constraints: Different targets require different orbit determination accuracies. For example, high-Earth orbit targets require orbit determination accuracies on the order of hundreds of meters, while low-Earth orbit targets require accuracies on the order of kilometers. This necessitates coordinated observation by different types of resources. 3. Action constraints: Once the decision variables are defined, the behavioral space of the resources is also defined. However, the behavior of different resources is subject to various constraints, such as the observation capacity of space-based resources, the frequency of telemetry, tracking, and command (TT&C) / data transmission, and the rotational angular velocity of space-based resources.

[0037] Task constraints include the following C1: Uniqueness of observation tasks. Each observation task can be executed at most once.

[0038] , , .

[0039] C2: Task Deadline. All tasks must be completed by the deadline.

[0040] ; .

[0041] Where TS and TE are the start and end times of the scheduling period.

[0042] C3: Minimum observation time. To improve orbit determination accuracy, the minimum observation time for the target cannot be too low.

[0043] ; , , ; if , .

[0044] C4: Adjust observation time. To improve orbit determination accuracy, the time interval between two consecutive observations of the same target should not be too short.

[0045] ; , if , , .

[0046] .

[0047] C5: Observation frequency. To improve orbit determination accuracy, the target needs to be observed multiple times within a scheduling cycle.

[0048] , , .

[0049] C6: Minimum monitoring and control time. To ensure data transmission quality, the monitoring and control time cannot be too short.

[0050] , if , .

[0051] C7: Minimum data transmission time. To ensure data transmission quality, the data transmission time cannot be too short.

[0052] , if , .

[0053] C8: TT&C Task Uniqueness. Each TT&C task can only be executed once.

[0054] .

[0055] C9: Data transfer task uniqueness. Each data transfer task can be executed at most once.

[0056] .

[0057] Motion constraints include the following: C10: Action Time. Space / ground-based resources and data transmission / measurement and control resources need to complete their respective actions within the corresponding time window.

[0058] , , , .

[0059] C11: Observation resource capability. When observing targets using sky-based / ground-based resources, its observation capability cannot exceed its upper limit (precision tracking radar tracks one target at a time, while phased array radar can track multiple targets).

[0060] ; , if , , .

[0061] For a goal: .

[0062] For multiple objectives: .

[0063] C12: The uniqueness of the observed resources. Space-based sensors can only observe GEO targets, while ground-based optical sensors can observe both GEO and LEO targets.

[0064] ; , , .

[0065] C13: Maximum observation time for space-based resources in each cycle. Space-based sensor resources have limited storage capacity; when the capacity reaches saturation, data needs to be downloaded.

[0066] Let the time period of a single orbit of space-based resources be... The maximum total observation time is , , ; If , or , then: .

[0067] C14: Maximum observation time for space-based resources in each cycle. Space-based sensor resources have limited storage capacity and cannot be frequently switched on and off.

[0068] Let the time period of a single orbit of space-based resources be... The maximum number of observations is , , If , Then: .

[0069] C15: Rotation speed of space-based resources. Space-based sensor resources have a relatively small field of view, so rotational angular velocity needs to be considered.

[0070] , ; If , , then, .

[0071] C16: Transfer time of ground resources. The interval time needs to be considered when the same sensor performs two adjacent tasks.

[0072] , ; If , , then, .

[0073] C17: Telemetry, tracking, and command (TT&C) resource capabilities. The same set of TT&C equipment can only provide TT&C services to one satellite at a time.

[0074] ; If , , , .

[0075] C18: Data transmission capacity. A single set of telemetry and control equipment can only provide data transmission services to one satellite at a time.

[0076] If , ; ; .

[0077] C19: Coupling constraint 1. The first telemetry and control action of space-based observation resources precedes the first observation action.

[0078] , , , .

[0079] C20: Coupling constraint 2. The first observation action of space-based observation resources precedes the first data transmission action.

[0080] , , , .

[0081] C21: Coupling constraint 3. If space-based resources are performing observation missions, then there will inevitably be telemetry, tracking, and command (TT&C) actions and data transmission actions.

[0082] , ; If ,then ; .

[0083] As the above analysis shows, the joint planning problem of space target tracking using space / ground-based resources is a complex combinatorial optimization problem involving multiple sources and heterogeneous structures. It requires consideration not only of the action plans for space / ground-based sensors but also of the execution transmission and data downlink plans. With increasing mission scale, its solution space grows exponentially. Traditional evolutionary algorithms struggle to solve this type of multi-source, heterogeneous, and complex constrained combinatorial optimization problem.

[0084] Currently, space-based and ground-based observation centers operate in different locations, and their equipment and workflows also differ significantly. To improve inter-center coordination, reduce disturbances during mission execution, and maximize resource utilization, this application proposes a solution method, namely the DTJS method, as follows: Figure 4As shown, it mainly includes task allocation, central scheduling, and collaborative optimization.

[0085] The DTJS method takes as input a set of RSOs to be observed, along with observation resources, TT&C resources, and data transmission resources. The output is an observation plan for each observation resource and the corresponding TT&C and data transmission plans for space-based resources. The DTJS method consists of two phases: task allocation and central scheduling. In the task allocation phase, the ground-based and space-based resource scheduling centers are treated as nodes in a graph. Feature information and interactions with neighboring nodes are used to allocate the RSOs to the corresponding resource scheduling centers, and the allocation probability is introduced as prior information into the subsequent collaborative optimization algorithm to guide further optimization. In the central scheduling phase, the ground-based and space-based resource scheduling centers perform HCO (resource allocation, conflict resolution, and collaborative scheduling) to obtain a scheduling scheme for each center. Finally, to further optimize the task allocation scheme, the ASAPS optimization strategy is introduced. The ASAPS optimization strategy aims to maximize observation gains by iteratively optimizing and adjusting the task allocation scheme.

[0086] Step 104: Using resource constraint information, orbital parameters of on-orbit targets, and observation requirements as input, a graph convolutional neural network is used to allocate tasks, determine the scheduling center to which each on-orbit target belongs, output the initial task allocation scheme, and use the allocation probability corresponding to the initial task allocation scheme as prior information.

[0087] This application designs a graph neural network for task allocation, that is, to allocate the RSO to be observed to space-based / ground-based resource scheduling centers. The network uses the framework of variational autoencoder (VAE) to simultaneously optimize the graph and allocation models during training, realizing end-to-end optimization of the graph and allocation models.

[0088] The observed gains for each RSO The shortest observation time of RSO The orbital parameters, the number of time windows, and the earliest and latest observation times corresponding to each time window are used as input features for network training. Different features exhibit spatial correlations, collectively determining the RSO assignment result. Since GCN utilizes the connections between nodes to update node representations, it incorporates local structural information from the graph into the node representation, thus improving assignment accuracy. This application uses GCN as its input... and its corresponding adjacency matrix Adjacency matrix This represents the spatial relationships between feature labels. This application first uses a graph convolution module to extract the spatial relationships between features, and then uses a two-dimensional convolution module for dimensionality reduction. The constructed GCN structure graph is shown below. Figure 5 As shown.

[0089] In the above figure, the graph convolution module has a kernel size of 3 and a number of 16. The input... and The output is , For the length of the training set, the dimensional convolutional module maps the output to... After passing through a fully connected layer, the output is... This yields the final prediction result.

[0090] The first step in using GCN is graph construction, but currently, graph construction is generally based on pre-defined expert knowledge. Due to the limited experience of experts, generating a graph with complete spatial relationships is quite difficult. Therefore, this application utilizes stochastic gradient descent in deep learning and, with the help of the VAE framework, designs an automatic graph optimization method. First, a weighted adjacency matrix is ​​randomly generated. The adjacency matrix is ​​then optimized to obtain the final GCN. The GCN is used as both the encoder and decoder. This represents a hidden layer. As shown in the following formula: .

[0091] According to the above formula, during the optimization process, the adjacency matrix... The GCN is continuously optimized to obtain the final prediction result. During the optimization process, the optimization objective is to minimize the mean error (MSE) of the training results, as shown in the following formula: .

[0092] If the adjacency matrix The value is greater than If the spatial association is valid, then the spatial association is considered valid. Through the above methods, an end-to-end framework for RSO allocation based on GCN can be obtained.

[0093] Using a trained GCN model, task allocation can be performed on RSO, with the input being... However, at this point, only the final task allocation scheme can be obtained. The DTJS algorithm designed in this application not only generates the task allocation scheme for RSO through GCN, but also uses the allocation probability as prior information to guide subsequent ASAPS for collaborative optimization, thereby optimizing the task allocation scheme. In this application, it transforms the probability into a range of [0,1] through max-min normalization, as shown in the following formula. Then, based on the probability, it is converted into 0 and 1 labels to obtain the initial task allocation scheme.

[0094] .

[0095] By leveraging GCN's ability to learn relationships in the feature space, the scheduling center to which each on-orbit target belongs can be quickly and accurately determined, avoiding the blindness of traditional task allocation. The output initial task allocation scheme and prior information on allocation probabilities provide clear division of labor for subsequent center scheduling and directional guidance for collaborative optimization, significantly reducing the search range of subsequent algorithms and improving overall computational efficiency.

[0096] Step 106: Design the HCO algorithm to plan the task execution sequence of each scheduling center and obtain the central scheduling scheme. The central scheduling scheme includes the task execution priority, resource allocation strategy and time window planning of each scheduling center.

[0097] After the task allocation phase, the space-based / ground-based resource scheduling center obtains the RSOs that need to be observed. Then, the center needs to uniformly schedule these RSOs to maximize observation benefits. Since collaborative optimization places high demands on the stability and speed of the center's scheduling algorithm, this step proposes the HCO algorithm. This method adopts a hierarchical approach, dividing the center scheduling into three stages: task-resource matching, conflict resolution, and multi-resource collaboration, reducing the solution space. At each level of the HCO algorithm, heuristic rules are used to solve resource scheduling, and through multi-level combinatorial optimization, the algorithm approaches its optimum. The main process of the HCO algorithm is as follows: Figure 6 As shown.

[0098] according to Figure 6 As can be seen, the HCO algorithm takes as input the RSOs to be observed by each center and outputs a resource scheduling scheme for each center. It mainly consists of three stages: task-resource matching, conflict resolution, and multi-resource collaborative scheduling. In the task-resource matching stage, space-based / ground-based centers allocate the received RSOs to their respective observation resources (each center contains multiple observation resources). After task-resource matching, conflicting tasks may occur within the same resource. To avoid these conflicts, conflict resolution is required to generate a scheduling plan for each resource. Finally, since resources within the same center can exchange information, a multi-resource collaborative scheduling method is designed to improve resource execution efficiency and observation benefits.

[0099] Through three steps—task-resource matching, conflict resolution, and multi-resource collaborative scheduling—HCO can quickly generate stable central scheduling schemes. These schemes ensure that tasks are rationally allocated to observation resources and fully utilize the center's resources, achieving efficient task execution and collaborative optimization. These three steps are described in detail below.

[0100] (1) Task-Resource Matching: Task-resource matching refers to assigning an RSO to its corresponding observation resource. In SGORJS, the number and length of visible time windows for different observation resources have a significant impact on the matching results. Generally, observation resources with more and longer visible time windows are advantageous for RSO allocation, meaning the RSO is more likely to be observed by that resource. Therefore, this step first calculates the average start time, average end time, average duration, and number of visible time windows for each resource. Then, combining these factors, the matching fitness index is proposed to guide the allocation of RSOs to different observation resources. Assumptions Indicates the first i The first observation resource k A visible time window, Then, the first i The average start time for each resource is: The average completion time is The average duration is The number of visible time windows for each resource is n i Therefore, regarding resources i The formula for calculating its matching fitness is: .

[0101] Because of its high resource fitness, it has more opportunities to observe RSOs. Therefore, the approach to task allocation in this step is as follows: Step 1: Sort the tasks to be observed according to priority; Step 2: Calculate each resource ; Step 3: Divide the sorted tasks to be observed into equal parts according to the amount of resources, and assign high-priority tasks to high-priority tasks. Observational resources; Step 4: Obtain the observation tasks required for each resource.

[0102] (2) Conflict resolution for each resource Based on the matching fitness, the RSOs to be observed under each resource can be obtained. However, when multiple RSOs are assigned to the same observation resource, it can lead to conflicting observation plans, resulting in observation failure. Therefore, conflict resolution is required for the observation tasks under each resource. To obtain a conflict-free task scheduling scheme for each resource in a short time, this step designs the RapidConflictResolution algorithm, which includes Direct and Shifting (RCRDS). The inputs of the RCRDS algorithm are the number of observation resources, observation tasks (RSOs), observation task priorities, and scheduling intervals of the scheduling scheme. The output is a conflict-free scheduling scheme for each resource. First, all observation tasks under each resource are sorted according to priority. Then, under each resource, the tasks to be observed are added to the scheduling scheme in order of priority. If the start time of a later task is greater than the end time of a previous task, i.e., If two tasks do not conflict, the task can be directly added to the scheduling scheme, i.e., an insertion strategy. Conversely, if the start time of the later task is earlier than the end time of the earlier task, i.e., ... At this point, two situations may occur. Case 1, task time Less than the time of adjacent tasks, i.e. At this point, the task can be directly inserted into the original plan, i.e., the Insert strategy is executed. Case 2, Task Time Greater than the time of adjacent tasks, i.e. In this case, the shifting strategy should be selected.

[0103] In theory, an observation plan for each resource can be obtained through task resource matching and conflict resolution. However, a single observation resource leads to low observation benefits, making it difficult to meet the requirements of SSN task scheduling. Since observation resources belonging to the same center can exchange information, this step uses an improved dynamic contract network algorithm to solve for the observation scheme in order to maximize the cooperation between observation resources. The main process is as follows: Step 1: The center (space-based or ground-based) occupies the tendering agency position. The tender document includes... , It is the contract number. It involves observing task information, including task priority, task duration, and task requirements.

[0104] Step 2: Bidding. After obtaining the bidding information, the resources (bidding) will generate a bidding proposal (scheduling plan). .in For contract number, Determine whether the bidding process should continue. For the benefit of bidding, This is the detection plan.

[0105] Step 3: Bid Evaluation. After receiving all tender documents, the best scheduling scheme is selected from the bids. The winning agent will be identified and notified through the evaluation algorithm, and the task will be assigned to the winning agent.

[0106] Step 4: The bidder signs a contract with the successful bidder and transmits the contract signing information to all bidders. Simultaneously, the winning agent merges the tasks into its task pool and executes the planned tasks.

[0107] This step efficiently resolves task conflicts and resource coordination issues within a single scheduling center by hierarchically planning task execution priorities, resource allocation strategies, and time windows. It achieves precise matching of observation resources and tasks, shortens task scheduling response time, fully leverages the resource potential within a single center, improves local resource utilization and scheduling stability, and provides a high-quality local scheduling foundation for global optimization.

[0108] Step 108: Based on the prior information, the initial task allocation scheme and the central scheduling scheme are jointly optimized according to the ASAPS optimization strategy, and the optimized task allocation scheme and central scheduling scheme are output.

[0109] This step proposes an ASAPS algorithm for joint scheduling of SSN space and ground observation resources. The method uses the results of GCN training as the initial allocation scheme and the probability of the allocation scheme as prior information to guide the algorithm's convergence. When the algorithm meets the termination condition, it outputs the optimal allocation scheme at that point. During the iterative optimization process, an adaptive mechanism updates the temperature, and a tabu list strategy is used to avoid repeated searches in the short term. Local perturbation strategies are employed to escape local optima, among other strategies, to assist the algorithm in continuously optimizing towards the optimal goal. Through collaborative optimization, the data-driven framework and the search capabilities of the simulated annealing algorithm can be fully utilized to obtain a more accurate and effective task allocation scheme. This optimization process will iterate until it converges to the optimal task allocation result. Figure 7 As shown.

[0110] The ASAPS algorithm takes as input an initial RSO allocation scheme and prior probabilities generated by a GCN. First, a solution transformation strategy is used to generate a new task allocation scheme based on the original RSO allocation scheme. The HCO algorithm is then used to determine if the new RSO allocation scheme will yield greater observation gains. If the new task allocation scheme is better, it is accepted and used as the new task allocation scheme; otherwise, an improved Metropolis method is used to generate a new scheme. Finally, combining a tabu list strategy and a perturbation strategy, an adaptive update mechanism is used to iteratively optimize the algorithm until the optimal RSO allocation scheme is output. The solution transformation strategy, adaptive neighborhood structure, improved Metropolis, tabu list strategy, and perturbation strategy appearing in ASAPS will be explained in detail below.

[0111] (1) Initial Population Construction Based on GCN: In this application, the initial solution of ASAPS is obtained by GCN. By utilizing the characteristics of GCN, key information can be extracted from a large amount of task and resource data and transformed into the representation of the initial population. Specifically, GCN can model the relationship between tasks and resources, and learn the potential rules of task allocation through the representation of nodes and edges, so as to serve as the initial solution of ASAPS.

[0112] Solution Transformation Strategy Based on Prior Search (SVPS): SV plays a crucial role in SA. A good SV can improve the current solution, avoid getting trapped in local optima, and accelerate algorithm convergence. In this step, we propose SVPS as a neighborhood transformation strategy for the SA algorithm. This method utilizes GCN to learn and extract the complex structure of SSN task scheduling, and uses the probability distribution output by GCN to guide the transformation of the neighborhood structure, thereby guiding the algorithm to converge quickly towards a better solution. As mentioned above, the input of the SVPS algorithm is the initial solution. Prior information Minimum probability and maximum probability The output is a new solution. The main process of the algorithm is as follows: First, based on the task allocation scheme and prior information... Calculate the adaptive information corresponding to each task. Then, based on... and Preserve the best solution, that is: =0&& < || =1&& > In this case, the allocation scheme remains unchanged. Finally, for other cases, an adaptive transformation is performed, i.e. =0&& > || =1&& < The task allocation scheme is changed based on random numbers.

[0113] (2) Adaptive Neighborhood Structure Selection: SVPS incorporates prior information from the GCN output, reflecting the RSO's preference for different centers and guiding the ASAPS algorithm towards optimal optimization. However, since ASAPS is also a probability-based stochastic optimization algorithm, a single neighborhood structure may cause the algorithm to get stuck in local optima. To overcome this problem, a stochastic transformation factor is introduced in this step. With adaptive dynamic neighborhood structure selection mechanism. Random transformation factor. This refers to randomly changing the task allocation scheme; it is also a solution transformation strategy, such as... Figure 8 As shown. Therefore, in the actual algorithm execution, SVPS and the random transformation factor will be used. These are all considered as neighborhood structures within the algorithm. Because the algorithm successfully improves the solution through its structures during iterations, the probability of further improvement in subsequent iterations is also higher. That is, if SVPS performs well, there is no need to execute it. Operation. Based on this, this step proposes an adaptive neighborhood dynamic selection mechanism, which adaptively selects the neighborhood structure during the algorithm's iterative search to improve the algorithm's search efficiency. The specific calculation formula is as follows: ; in, For inertial weights, This refers to the number of times the solution scheme has been successfully improved using the k-th neighborhood structure during the most recent iteration. This refers to the number of times the k-th neighborhood structure is selected during the most recent iteration.

[0114] (3) Improved Metropolis Criterion: The Metropolis criterion is used to determine whether to accept a new solution. In traditional SA, each iteration generates a new solution and calculates its energy difference with the current solution. According to the Metropolis criterion, if the new solution is better than the current solution, it is accepted. Otherwise, the new solution may be accepted based on a certain probability. However, in the SSN task scheduling of this application, HCO is proposed to solve specific scheduling schemes. Since the HCO algorithm is an approximate optimal algorithm, it cannot obtain a globally optimal solution, which may lead to the situation of "the same allocation result, different scheduling benefits". Therefore, this step improves the Metropolis criterion, that is, if the new solution is better than the current solution, the new solution is accepted. Or, if the new solution is not less than 10% of the original solution's benefit, the new solution is accepted with a high probability, and the original solution is kept in the tabu list at the same time. Otherwise, it is accepted based on a certain probability. The calculation formula is as follows: ;

[0115] in, These are control parameters. Initially... It is relatively large, and after a sufficient number of transfers, it slowly decreases. The algorithm repeats this process until a certain criterion is met, at which point it terminates.

[0116] (4) Tabootable strategy: Tabu search lacks memory of solution transformations during the search process, which may lead to the algorithm getting trapped in local optima. By constructing a Tabu table, the lack of memory during the search process can be effectively combined with Tabu search. Neighborhood solutions are generated based on the neighborhood structure. Under the improved Metropolis criterion, once a new solution is received, the scheduling scheme of the original solution is retained in the Tabu table, with L representing the length of the Tabu table. The purpose of the Tabu table is to remember recently executed search operations and to tabu and detab certain operations. Tasks in the Tabu table are not allowed to be scheduled or arranged unless they are released. If the length of the tasks in the Tabu table reaches L, the tasks added to the Tabu table first are released in sequence.

[0117] (5) Perturbation Strategy: To avoid the algorithm getting trapped in local optima during iterative optimization, the improved Metropolis criterion accepts inferior solutions with a certain probability, but it is still difficult to escape local optima. Therefore, this step designs a perturbation strategy, continuing the search based on the perturbation, which may increase the quality of the solutions obtained. When the algorithm's gains are continuous... Cum If no change is observed, the perturbation strategy is triggered. The perturbation process is as follows: randomly select two individuals of the same length from a set of task allocation schemes and swap them, similar to the crossover operation in genetic algorithms. Unlike individual transformation, the perturbation strategy transforms a region, thus being more likely to expand the algorithm's search range and continuously optimize towards the optimal solution.

[0118] (6) Temperature Adaptive Update: The optimization behavior of the algorithm is mainly affected by temperature, and the Metropolis criterion has a direct relationship with the acceptance probability of inferior solutions. Therefore, temperature control is of great significance in the optimization and convergence of the algorithm. If the temperature is monotonically decreasing, the acceptance probability will also decrease monotonically, which can easily lead the algorithm into local optima. To this end, this step proposes an adaptive temperature update mechanism, the calculation formula of which is as follows: ; in This represents the maximum temperature. These are the weighting coefficients. This represents an algebraic expression where the objective function value of the current solution has not improved. This is the temperature adjustment factor.

[0119] (7) Termination Conditions: To improve the efficiency of the algorithm, two termination condition criteria were designed: Criterion 1: Current iteration number Itr The maximum number of iterations has been reached. ; Criterion 2: Continuity The suboptimal solution has not been improved.

[0120] If any of the above criteria are met during the algorithm execution, the algorithm terminates and outputs the global optimal solution.

[0121] ASAPS time complexity analysis: The complexity of ASAPS is mainly determined by two parts: the outer loop stage and the outer loop stage. This stage mainly involves iterative searching by adjusting the temperature. Assuming there are N temperature stages, the time complexity is O(N). Then comes the inner loop phase. During the iteration of the ASAPS algorithm, new solutions need to be generated based on the adaptive neighborhood structure, and the improved Metropolis algorithm is used to determine whether to accept the new solution. Assume that the average time complexity of the inner loop in each temperature phase is O(n log n). ,in This represents the average number of iterations in the inner loop. Generally speaking, That is, the time complexity of the ASAPS algorithm is O(n log n). It can be executed in polynomial time.

[0122] By incorporating prior information, the initial task allocation and central scheduling scheme are globally adjusted. Through mechanisms such as adaptive neighborhood selection, tabu lists, and perturbation strategies, the algorithm avoids getting trapped in local optima and accelerates convergence. This solves the problems of low search efficiency and susceptibility to local optima in traditional optimization algorithms, further optimizes the synergy between task allocation and resource scheduling, and improves the optimization effect of the global scheduling scheme.

[0123] Step 110: Integrate the optimized task allocation scheme with the central scheduling scheme to form the final scheduling scheme, which is used to guide the on-orbit target observation and scheduling execution of the space surveillance network.

[0124] By integrating the optimized task allocation scheme with the central scheduling scheme, a globally unified scheduling scheme is formed, breaking down the barriers between independent scheduling of space-based and ground-based resources. This ensures the integrity and synergy of the scheduling scheme, fully leveraging the complementary advantages of space-based and ground-based resources. The final output scheme can directly guide actual observation and scheduling execution, improving the feasibility of scheduling and the ability to perceive the space situation.

[0125] The aforementioned data-driven daily scheduling method for large-scale space target surveillance first obtains information on the resource constraints of the scheduling center, the orbital parameters of on-orbit targets, and observation requirements, providing complete data support for modeling. The constructed SGORJS mathematical model simultaneously covers the objective functions and constraints of task allocation and center scheduling, accurately depicting the coupling relationship of multiple resources and avoiding the optimization limitations caused by information gaps or one-sided modeling in traditional models. Secondly, a graph convolutional neural network is used for task allocation. Leveraging its ability to learn relationships in the feature space, it quickly determines the scheduling center to which the target belongs. The output initial scheme and prior information on allocation probabilities lay an accurate foundation for subsequent optimization, reducing blind searches. The HCO algorithm is designed to efficiently solve task conflicts and resource coordination problems within a single scheduling center by hierarchically planning task execution priorities, resource allocation strategies, and time windows, improving local scheduling efficiency. Simultaneously, the ASAPS optimization strategy, combined with prior information, collaboratively optimizes the initial task allocation and center scheduling scheme. Through mechanisms such as adaptive neighborhood selection and tabu lists, it avoids local optima and accelerates convergence, further optimizing the overall scheme. Finally, by integrating and optimizing the two types of schemes, the solution forms the globally optimal final scheduling scheme. It exerts efforts throughout the entire process from data foundation, modeling accuracy, algorithm efficiency to global optimization, effectively breaking through the barriers of traditional independent scheduling and significantly improving the optimization effect and computational efficiency of the scheduling scheme.

[0126] In one embodiment, the resource constraint information includes the number of observation devices, observation range, working time and energy consumption limits of each scheduling center; the orbital parameters of the on-orbit target include orbital inclination, perigee altitude, apogee altitude and orbital period; the objective function includes maximizing observation coverage, minimizing task execution delay and maximizing resource utilization; and the constraints include device capacity constraints, time window constraints and orbital visibility constraints.

[0127] In one embodiment, using resource constraint information, orbital parameters of on-orbit targets, and observation requirements as input, a graph convolutional neural network is used for task allocation to determine the scheduling center to which each on-orbit target belongs. An initial task allocation scheme is output, and the allocation probability corresponding to the initial task allocation scheme is used as prior information, including: Feature data of on-orbit targets are extracted from resource constraint information, orbital parameters of on-orbit targets and observation requirements information. Feature data includes the observation benefits of on-orbit targets, shortest observation time, orbital parameters, number of time windows and the earliest and latest observation times corresponding to the time windows. The feature data is divided into a training set and a test set. The training set includes training feature data and training labels, while the test set consists of test feature data. A GCN network based on a variational autoencoder framework is constructed. The GCN network includes a graph convolution module, a two-dimensional convolution module, and a fully connected layer. The graph convolution module extracts spatial relationships from the input data and outputs the first feature data. The first feature data is input into the two-dimensional convolution module for dimensionality reduction and outputs the second feature data. The second feature data is input into the fully connected layer and outputs the prediction result corresponding to the input data. The GCN network is trained using the training set. The training process includes randomly generating an adjacency matrix, using the training feature data, training labels, and the convolutional module of the adjacency matrix input graph as an encoder to output a hidden layer, performing a transpose adjacency matrix transformation on the hidden layer, and then using the transformed hidden layer and the convolutional module of the adjacency matrix input graph as a decoder to output predicted labels, and optimizing the parameters of the adjacency matrix and the GCN network using stochastic gradient descent based on the training labels and predicted labels, with the minimum average error as the optimization objective, until the average error of the training result meets the preset requirements, thus completing the training of the GCN network. The test feature data is input into the trained GCN network, which outputs the initial task allocation scheme for the on-orbit target to be assigned to the space-based or ground-based scheduling center, and the allocation probability corresponding to the initial task allocation scheme is also output as prior information.

[0128] In one embodiment, an HCO algorithm is designed to plan the task execution sequence of each scheduling center to obtain a central scheduling scheme, including: The space-based dispatch center and the ground-based dispatch center each obtain the RSO assigned to themselves in the initial task allocation scheme as the RSO to be observed; For each RSO to be observed in a scheduling center, perform task-resource matching and calculate the matching fitness of each observed resource in the scheduling center. The matching fitness is determined based on the average start time, average end time, average duration and number of time windows of the observed resource. The RSOs to be observed are sorted by priority, and then assigned to the corresponding observation resources according to the matching fitness of the observation resources. For each observation resource, conflict resolution is performed on the RSO to be observed. A fast conflict resolution algorithm is used to add RSOs to the scheduling scheme of the observation resource in order of RSO priority. If there is no time conflict between the next RSO and the previous RSO, it is directly inserted. If there is a time conflict, insertion or offset adjustment is selected according to the relationship between the duration of the RSO and the time of the adjacent task to obtain a conflict-free scheduling scheme for each observation resource. For conflict-free scheduling schemes of various observation resources within the same scheduling center, multi-resource collaborative scheduling is performed. An improved dynamic contract network algorithm is adopted. The scheduling center acts as a bidding agent to issue bidding information, and each observation resource generates a bidding scheme as a bidder. The scheduling center evaluates the bidding schemes and determines the winning observation resource. The task is assigned to the winning observation resource and the scheduling scheme is integrated to obtain the central scheduling schemes of the space-based scheduling center and the ground-based scheduling center.

[0129] In one embodiment, the initial task allocation scheme and the central scheduling scheme are collaboratively optimized based on prior information according to the ASAPS optimization strategy, and the optimized task allocation scheme and central scheduling scheme are output, including: The initial task allocation scheme is used as the initial solution for the ASAPS optimization strategy, and the initial solution is introduced into the ASAPS optimization strategy. The solution transformation strategy and random transformation factor are used as the neighborhood structure of the algorithm; an adaptive neighborhood dynamic selection mechanism is adopted to calculate the selection probability of each neighborhood structure; based on the calculated selection probability, the neighborhood structure is selected to perform neighborhood transformation on the initial solution, and a new task allocation scheme is generated by combining prior information. Input the new task allocation scheme into the HCO algorithm, calculate the observation benefits corresponding to the new scheme, and compare them with the observation benefits of the current task allocation scheme. The decision to accept a new task allocation scheme is based on the improved Metropolis criterion. If the new scheme has higher observed benefits, it is accepted directly. If the observed benefits of the new scheme are not less than 90% of the current scheme, it is likely to be accepted. Otherwise, it is accepted according to the preset probability. A tabu list strategy is used to record accepted task allocation schemes to avoid repeated searches in the short term. If the observed gains do not change in a preset number of iterations, a perturbation strategy is triggered to randomly select two individuals of the same length in the task allocation scheme and swap them to expand the search range. An adaptive temperature update mechanism is used to adjust the temperature of the ASAPS algorithm until the maximum number of iterations is reached or no optimal solution is found after a preset number of iterations, and then the optimal task allocation scheme is output.

[0130] In one embodiment, the selection probability of each neighborhood structure is calculated as follows: ; in, For inertial weights, This refers to the number of times the solution was successfully improved using the k-th neighborhood structure during the most recent iteration. This refers to the number of times the k-th neighborhood structure is selected during the most recent iteration.

[0131] In one embodiment, calculating the matching fitness of each observation resource within the scheduling center includes: The fitness of the matching of each observation resource within the scheduling center is calculated as follows: ; in, , , , Each represents a different weight. Indicates the first i Average start time for each resource Indicates the first i The first observation resource k The start time of a visible time window. Indicates the first i Average end time for each resource Indicates the first i The first observation resource k The end time of a visible time window. Indicates the first i Average duration under each resource This indicates the number of visible time windows for each resource.

[0132] In one embodiment, the training feature data, training labels, and adjacency matrix input graph convolutional module of the training set are used as an encoder to output a hidden layer; the hidden layer is then subjected to a transposed adjacency matrix transformation, and the transformed hidden layer and the adjacency matrix input graph convolutional module are used as a decoder to output predicted labels, including: The training feature data, training labels, and adjacency matrix input graph convolutional module of the training set are used as the encoder to output the hidden layer. The hidden layer is then transformed by the transposed adjacency matrix, and the transformed hidden layer and the adjacency matrix input graph convolutional module are used as the decoder to output the predicted labels. ; in, Indicates a hidden layer. Represents training feature data, Indicates training labels, Represents the adjacency matrix. This represents the transformed hidden layer. Indicates the predicted label, This represents a graph convolution encoder.

[0133] In one embodiment, the probability of accepting a suboptimal solution in the improved Metropolis criterion is calculated as follows: ; in, The probability of accepting an inferior solution, For control parameters, The observational benefits of the current task allocation scheme. The observational benefits of the new task allocation scheme. It is a natural constant.

[0134] In one embodiment, an adaptive temperature update mechanism is used to adjust the temperature of the ASAPS algorithm, including: The temperature of the ASAPS algorithm is adjusted using an adaptive temperature update mechanism. ; in, This represents the maximum temperature. These are the weighting coefficients. This represents an algebraic expression where the objective function value of the current solution has not improved. This is the temperature adjustment factor.

[0135] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this application, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0136] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0137] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A data-driven daily scheduling method for large-scale space target surveillance, characterized in that, The method includes: To acquire resource constraint information of each scheduling center in the space surveillance network, orbital parameters and observation requirements of on-orbit targets, and to establish a mathematical model of the SSNSGORJS problem, the mathematical model includes the objective function and constraints of task allocation and center scheduling. Using the resource constraint information, the orbital parameters of the on-orbit target, and the observation requirement information as input, a graph convolutional neural network is used to allocate tasks, determine the scheduling center to which each on-orbit target belongs, output an initial task allocation scheme, and use the allocation probability corresponding to the initial task allocation scheme as prior information. The HCO algorithm is designed to plan the task execution sequence of each scheduling center to obtain a central scheduling scheme. The central scheduling scheme includes the task execution priority, resource allocation strategy and time window planning of each scheduling center. Based on the ASAPS optimization strategy and prior information, the initial task allocation scheme and the central scheduling scheme are jointly optimized, and the optimized task allocation scheme and central scheduling scheme are output. The optimized task allocation scheme is integrated with the central scheduling scheme to form the final scheduling scheme, which is used to guide the on-orbit target observation and scheduling execution of the space surveillance network.

2. The method according to claim 1, characterized in that, The resource constraint information includes the number of observation devices, observation range, working time and energy consumption limits of each scheduling center; the orbital parameters of the on-orbit target include orbital inclination, perigee altitude, apogee altitude and orbital period; the objective function includes maximizing observation coverage, minimizing task execution delay and maximizing resource utilization, and the constraints include equipment capacity constraints, time window constraints and orbital visibility constraints.

3. The method according to claim 1, characterized in that, Using the resource constraint information, the orbital parameters of the on-orbit targets, and the observation requirements information as input, a graph convolutional neural network is used for task allocation to determine the scheduling center to which each on-orbit target belongs. An initial task allocation scheme is output, and the allocation probability corresponding to the initial task allocation scheme is used as prior information, including: Feature data of on-orbit targets are extracted from resource constraint information, orbital parameters of on-orbit targets and observation requirements information. The feature data includes the observation benefits, shortest observation time, orbital parameters, number of time windows and the earliest and latest observation times corresponding to the time windows of the on-orbit targets. The feature data is divided into a training set and a test set. The training set includes training feature data and training labels, and the test set is test feature data. A GCN network based on a variational autoencoder framework is constructed. The GCN network includes a graph convolution module, a two-dimensional convolution module, and a fully connected layer. The graph convolution module extracts spatial relationships from the input data and outputs first feature data. The first feature data is input into the two-dimensional convolution module for dimensionality reduction and outputs second feature data. The second feature data is input into the fully connected layer and outputs the prediction result corresponding to the input data. The GCN network is trained using the training set. The training process includes randomly generating an adjacency matrix, using the training feature data and training labels of the training set, along with the convolutional module of the adjacency matrix input graph, as an encoder to output a hidden layer; performing a transpose adjacency matrix transformation on the hidden layer, and then using the transformed hidden layer and the convolutional module of the adjacency matrix input graph as a decoder to output predicted labels; based on the training labels and predicted labels, optimizing the parameters of the adjacency matrix and the GCN network using stochastic gradient descent with the minimum average error as the optimization objective, until the average error of the training result meets the preset requirements, thus completing the training of the GCN network. The test feature data is input into the trained GCN network, which outputs the initial task allocation scheme for the on-orbit target to be assigned to the space-based or ground-based scheduling center, and the allocation probability corresponding to the initial task allocation scheme is also output as prior information.

4. The method according to claim 1, characterized in that, The HCO algorithm is designed to plan the task execution sequence of each scheduling center, resulting in a central scheduling scheme, including: The space-based dispatch center and the ground-based dispatch center each obtain the RSO assigned to them in the initial task allocation scheme as the RSO to be observed; For each RSO to be observed in a scheduling center, a task-resource matching is performed, and the matching fitness of each observed resource in the scheduling center is calculated. The matching fitness is determined based on the average start time, average end time, average duration and number of time windows of the observed resource. The RSOs to be observed are sorted by priority, and then assigned to the corresponding observation resources according to the matching fitness of the observation resources. For each observation resource, conflict resolution is performed on the RSO to be observed. A fast conflict resolution algorithm is used to add RSOs to the scheduling scheme of the observation resource in order of RSO priority. If there is no time conflict between the next RSO and the previous RSO, it is directly inserted. If there is a time conflict, insertion or offset adjustment is selected according to the relationship between the duration of the RSO and the time of the adjacent task to obtain a conflict-free scheduling scheme for each observation resource. For conflict-free scheduling schemes of various observation resources within the same scheduling center, multi-resource collaborative scheduling is performed. An improved dynamic contract network algorithm is adopted. The scheduling center acts as a bidding agent to issue bidding information, and each observation resource generates a bidding scheme as a bidder. The scheduling center evaluates the bidding schemes and determines the winning observation resource. The task is assigned to the winning observation resource and the scheduling scheme is integrated to obtain the central scheduling schemes of the space-based scheduling center and the ground-based scheduling center.

5. The method according to claim 1, characterized in that, Based on the ASAPS optimization strategy and prior information, the initial task allocation scheme and the central scheduling scheme are collaboratively optimized, and the optimized task allocation scheme and central scheduling scheme are output, including: The initial task allocation scheme is used as the initial solution of the ASAPS optimization strategy, and the initial solution is introduced into the ASAPS optimization strategy. The solution transformation strategy and random transformation factor are used as the neighborhood structure of the algorithm; an adaptive neighborhood dynamic selection mechanism is adopted to calculate the selection probability of each neighborhood structure; based on the calculated selection probability, the neighborhood structure is selected to perform neighborhood transformation on the initial solution, and a new task allocation scheme is generated by combining prior information. Input the new task allocation scheme into the HCO algorithm, calculate the observation benefits corresponding to the new scheme, and compare them with the observation benefits of the current task allocation scheme. The decision to accept a new task allocation scheme is based on the improved Metropolis criterion. If the new scheme has higher observed benefits, it is accepted directly. If the observed benefits of the new scheme are not less than 90% of the current scheme, it is likely to be accepted. Otherwise, it is accepted according to the preset probability. A tabu list strategy is used to record accepted task allocation schemes to avoid repeated searches in the short term. If the observed gains do not change in a preset number of iterations, a perturbation strategy is triggered to randomly select two individuals of the same length in the task allocation scheme and swap them to expand the search range. An adaptive temperature update mechanism is used to adjust the temperature of the ASAPS algorithm until the maximum number of iterations is reached or no optimal solution is found after a preset number of iterations, and then the optimal task allocation scheme is output.

6. The method according to claim 5, characterized in that, The selection probability of each neighborhood structure is calculated as follows: in, For inertial weights, This refers to the number of times the solution was successfully improved using the k-th neighborhood structure during the most recent iteration. This refers to the number of times the k-th neighborhood structure is selected during the most recent iteration.

7. The method according to claim 4, characterized in that, Calculate the matching fitness of each observation resource within the scheduling center, including: The fitness of each observation resource within the scheduling center is calculated as follows: in, , , , Each represents a different weight. Indicates the first i Average start time for each resource Indicates the first i The first observation resource k The start time of a visible time window. Indicates the first i Average end time for each resource Indicates the first i The first observation resource k The end time of a visible time window. Indicates the first i Average duration under each resource This indicates the number of visible time windows for each resource.

8. The method according to claim 3, characterized in that, The training feature data and training labels of the training set are used as an encoder along with the convolutional module of the adjacency matrix input graph to output a hidden layer. The hidden layer is then subjected to a transposed adjacency matrix transformation. Finally, the transformed hidden layer and the convolutional module of the adjacency matrix input graph are used as a decoder to output predicted labels, including: The training feature data and training labels of the training set are used as an encoder along with the convolutional module of the adjacency matrix input graph to output a hidden layer. The hidden layer is then subjected to a transposed adjacency matrix transformation. Finally, the transformed hidden layer and the convolutional module of the adjacency matrix input graph are used as a decoder to output predicted labels. in, Indicates a hidden layer. Represents training feature data, Indicates training labels, Represents the adjacency matrix. This represents the transformed hidden layer. Indicates the predicted label, This represents a graph convolution encoder.

9. The method according to claim 5, characterized in that, In the improved Metropolis criterion, the probability of accepting a suboptimal solution is calculated using the following formula: in, The probability of accepting an inferior solution, For control parameters, The observational benefits of the current task allocation scheme. The observational benefits of the new task allocation scheme, It is a natural constant.

10. The method according to claim 5, characterized in that, The ASAPS algorithm employs an adaptive temperature update mechanism to adjust its temperature, including: The temperature of the ASAPS algorithm is adjusted using an adaptive temperature update mechanism as follows: in, This represents the maximum temperature. These are the weighting coefficients. This represents an algebraic expression where the objective function value of the current solution has not improved. This is the temperature adjustment factor.

Images