A planning method for agile observation satellites

Abstract

The application discloses a kind of planning methods for agile observation satellite, including the following steps: obtaining satellite dataset, constructs satellite scene model;Satellite task scheduling data generation and labeling are carried out based on satellite scene model;Obtain the dataset with constellation scheduling label;Based on the scheduling model of Transform architecture is constructed, and embedding internal constraint module, satellite attribute matrix and task attribute matrix are used for satellite task allocation;Using dataset, training and testing scheduling model, through reinforcement learning fine-tuning, obtain the final satellite task scheduling model;When the target task planning of agile observation satellite is carried out, the scheduling of corresponding agile observation satellite is carried out through the final satellite task scheduling model.The application relates to the field of satellite communication technology, and satellite task allocation is carried out by constructing benchmark test suite and scheduling model;The goal of optimizing large AEOS constellation scheduling task planning in complex environment is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite communication, and more particularly to a planning method for agile observation satellites. BACKGROUND

[0002] Agile Earth Observation Satellites (AEOS) have become a transformative technology in the field of remote sensing, enabling rapid and flexible monitoring of the Earth's surface. By coordinating planning in a constellation, multiple AEOS can significantly increase revisit frequency and expand coverage, outperforming a single satellite.

[0003] AEOS constellations provide unprecedented flexibility for monitoring the Earth's surface, but their scheduling remains challenging under strict conditions of large-scale scenarios, dynamic environments where tasks can be published and due at any time, and constraints on power and attitude. Existing methods often simplify these complexities, limiting their performance in the real world.

[0004] Existing methods can be broadly divided into optimization-based methods and neural network-based methods.

[0005] I. Optimization-based methods.

[0006] Early research relied on exact solvers to optimize satellite assignments. There are methods that use a constraint programming framework for AEOS scheduling; there are methods that use sequential convex programming to accelerate target acquisition. Although these methods guarantee optimality, their computational cost rises sharply with the size of the problem. Subsequent heuristic algorithms aim to improve adaptability, with methods that balance performance and runtime by switching between perception tasks assignments; there are methods that use a random hill-climbing strategy for on-time assignment optimization; other methods include ant colony optimization, evolutionary algorithms, and genetic algorithms. Although these methods provide faster runtimes, their performance decreases in large-scale or dynamic scenarios.

[0007] II. Neural network-based methods.

[0008] The powerful fitting capabilities of neural networks have driven breakthroughs in constellation scheduling. New methods based on neural networks have emerged, describing the problem as a Markov Decision Process (MDP) and using reinforcement learning for scheduling; pointer networks provide a seq2seq formula for multiple task assignment optimization; there are also methods that use GNN and deep reinforcement learning to solve AEOS planning problems, with very competitive performance; multi-agent RL is combined with a polynomial-time greedy solver to balance assignment quality and speed. Despite good research results, these methods simplify key physical constraints.

[0009] Therefore, it is an urgent problem for those skilled in the art to propose a benchmark and model combined with physical constraints for constellation scheduling of agile observation satellites. SUMMARY

[0010] Therefore, the application provides a planning method for agile observation satellites, which effectively solves the limitations of existing satellite constellation scheduling methods by constructing a SatData benchmark test suite and a Sat-Former scheduling model, and achieves the goal of optimizing large AEOS constellation scheduling task planning in complex environments.

[0011] To achieve the above purpose, the application adopts the following technical solutions:

[0012] The application provides a planning method for agile observation satellites, comprising the following steps:

[0013] S1, obtain a satellite data set, and construct a satellite scene model based on a satellite system model, satellite task definition, satellite action space control and multiple constraint limitations;

[0014] S2, generate and label satellite task scheduling data based on the satellite scene model, and obtain a data set SatData with constellation scheduling labels;

[0015] S3, construct a scheduling model Sat-Former based on a Transformer architecture, and embed an internal constraint module, a satellite attribute matrix and a task attribute matrix for satellite task allocation;

[0016] S4, train and test the scheduling model Sat-Former using the data set SatData with constellation scheduling labels, and obtain a final satellite task scheduling model through reinforcement learning fine-tuning;

[0017] S5, when planning a target task of an agile observation satellite, schedule the corresponding agile observation satellite through the final satellite task scheduling model.

[0018] Further, the generation process of the satellite data set comprises:

[0019] MRP empirical formula calculation is performed for each satellite task, completion rate checking and manual quality checking are performed, and it is decided whether to accept or regenerate the MRP according to the checking result;

[0020] Satellites that pass the check are regarded as qualified satellite assets, and a satellite data set is formed.

[0021] Further, the satellite scene model is constructed based on a satellite system model, satellite task definition, satellite action space control and multiple constraint limitations, and specifically comprises:

[0022] constructing a satellite system model based on orbit dynamics, attitude control, power system and sensor payload;

[0023] defining a satellite mission based on release time, expiration time, observation duration and ground target coordinates;

[0024] controlling a satellite action space by using high-level control to schedule low-level control; wherein the high-level control includes mission assignment commands; and the low-level control includes power switch commands and attitude pointing instructions;

[0025] constructing dynamics constraints, energy constraints, field of view constraints, continuity constraints and time window constraints to limit the execution of satellite missions; and obtaining a satellite scenario model including problem definition and scenarios.

[0026] Further, the step S2 specifically includes:

[0027] In the satellite scenario model, a greedy algorithm is used to perform initial assignment to obtain a preliminary assignment result;

[0028] simulate the preliminary assignment result, and retain the observation trajectories that are successfully simulated;

[0029] Through artificial quality review, high-quality scheduling trajectories are marked to obtain a dataset SatData with constellation scheduling annotations.

[0030] Further, in step S3, the satellite attribute matrix includes a static satellite matrix and a dynamic satellite matrix; and is expressed by a formula as:

[0031]

[0032] wherein L represents the satellite attribute matrix; L S represents the static satellite matrix; L d represents the dynamic satellite matrix; N L represents the number of satellites; d L represents the satellite feature dimension.

[0033] The mission attribute matrix includes a static mission matrix and a dynamic mission matrix; and is expressed by a formula as:

[0034]

[0035] wherein R represents the mission attribute matrix, R S represents the static mission matrix; R d represents the dynamic mission matrix; N R represents the number of tasks; d R represents the mission attribute dimension.

[0036] Further, in step S3, the internal constraint module is configured to predict the feasibility of the satellite to perform the task; the execution process comprises:

[0037] S31, for each satellite-task pair , the internal constraint module predicts the feasibility of the satellite i to perform the task j , which is expressed by the formula as follows:

[0038]

[0039]

[0040]

[0041]

[0042] wherein, represents the predicted logarithm of the feasibility of the satellite i to complete the task j; L i represents the matrix of the satellite i, R j represents the matrix of the task j; represents a set of real numbers;

[0043] S32, define an approximate label , process the task that needs to be completed by multiple satellites in cooperation;

[0044] S33, define a feasibility loss function and introduce time supervision, further guide the internalization constraint of the internal constraint module ;

[0045] The feasibility loss function is expressed by the formula as follows:

[0046]

[0047] wherein, represents the feasibility loss.

[0048] Further, in step S4, the data set SatData with constellation scheduling labels is used to train and test the scheduling model Sat-Former, specifically comprising:

[0049] S41, the embedding module is used to respectively convert the and into a vector form and embed a sine time step; the satellite feature vector and the task feature vector are obtained;

[0050] S42, the satellite feature vector and the task feature vector are decoded by using the Transformer decoder to obtain the satellite feature h L and the task feature h R ;

[0051] S43, quantifying the matching degree between satellites and tasks by assigning a score matrix M, adding a feasibility prediction of the embedded constraint module to guide the planning, and using the formula:

[0052]

[0053]

[0054] wherein, represents the predicted logarithm of the feasibility of satellite i completing task j; represents a trainable vector, represents a Hadamard product of element-wise multiplication, and F represents a feasibility matrix obtained after 1 padding on

[0055] S44, defining a model loss function to evaluate the current task allocation result and optimizing the scheduling model Sat-Former; the model loss function is a loss function of the model allocation target:

[0056]

[0057] wherein, represents the model loss, and m represents the task true value;

[0058] S45, in the test phase, exclude unfeasible satellite-task pairs when quantifying the assigned score matrix M;

[0059] using the formula:

[0060]

[0061] wherein, represents a satellite-task pair; represents an indicator function, represents a predefined feasibility threshold; represents the ith row and jth value of the assigned score matrix.

[0062] Further, in step S4, the final satellite task scheduling model is obtained by fine-tuning through reinforcement learning, specifically including:

[0063] In the supervised pre-training phase, the Sat-Former is initialized with random weights, and is trained based on the labeled trajectories in the SatData to obtain a total loss function;

[0064] The total loss function is represented by the formula:

[0065]

[0066] wherein,​ represents a total loss function; and respectively represent weights of and

[0067] In the reinforcement learning fine-tuning stage, comprehensive evaluation is carried out according to the corresponding total loss function; and the trajectory exceeding the feasibility threshold is added to the SatData for reinforcement learning of the Sat-Former until a final satellite task scheduling model is obtained by reaching model convergence.

[0068] Further, in step S4, the reinforcement learning specifically includes:

[0069] Initializing a value function network to perform value evaluation on the current state; and taking the Sat-Former as a policy function in reinforcement learning, training the Sat-Former step by step, and performing probability sampling through the assignment score matrix output by the Sat-Former;

[0070] Guiding the reinforcement learning through a reward function and performing reward calculation according to the current state; calculating a time difference error according to the calculated reward and the next state; and updating the Sat-Former and the value function network according to the time difference error.

[0071] According to the above technical solution, compared with the prior art, the planning method for agile observation satellites provided by the present disclosure has the following beneficial effects:

[0072] The present disclosure constructs a large-scale and high-precision benchmark test suite SatData, accurately simulates satellite orbit dynamics, attitude control and other physical characteristics, introduces real satellite data testing, ensures that the constraint conditions and evaluation indexes are close to reality, and provides expert-labeled true value scheduling labels, solves the problem of lack of general evaluation standards in existing benchmarks, supports fair comparison of models, and has comprehensiveness and openness.

[0073] The present disclosure is based on the Transformer architecture, embeds an internal constraint module, explicitly models the physical and operational limits of the satellite, and constructs a Sat-Former model. The module predicts the feasibility probability to guide the scheduling decision and improve the feasibility and fidelity of the generated scheme.

[0074] In the Sat-Former model processing, the action space is divided into high-level task allocation and low-level control instruction, the scheduling model focuses on task selection and timing, the platform automatically converts into low-level instructions, and the model complexity is reduced. In dynamic data processing, the static properties and dynamic states of the satellite and the task are fused, the sine time embedding is introduced, and the model's ability to capture time sequence features is improved.

[0075] The internal constraint module is supervised by binary cross-entropy loss (BCE) to optimize the feasibility prediction. BRIEF DESCRIPTION OF DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0077] Figure 1 A planning method flowchart for agile observation satellites is provided for the embodiments of the present application.

[0078] Figure 2 A satellite data set generation schematic diagram is provided for the embodiments of the present application.

[0079] Figure 3 A data set SatData generation schematic diagram is provided for the embodiments of the present application.

[0080] Figure 4 A structure diagram of a scheduling model Sat-Former is provided for the embodiments of the present application.

[0081] Figure 5 A reinforcement learning fine-tuning flowchart is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0083] The embodiments of the present application disclose a planning method for agile observation satellites, as shown in Figure 1 The method comprises the following steps:

[0084] S1, acquiring a satellite data set, constructing a satellite scene model based on a satellite system model, satellite task definition, satellite action space control and multiple constraint limits;

[0085] S2, generating and labeling satellite task scheduling data based on the satellite scene model; obtaining a data set SatData with constellation scheduling labels;

[0086] S3, constructing a scheduling model Sat-Former based on a Transformer architecture, and embedding an internal constraint module, a satellite attribute matrix and a task attribute matrix, for satellite task allocation;

[0087] S4, training and testing the scheduling model Sat-Former by using the data set SatData with constellation scheduling annotations, and obtaining a final satellite task scheduling model by reinforcing learning and fine-tuning the model;

[0088] S5, when planning a target task of an agile observation satellite, scheduling the corresponding agile observation satellite by using the final satellite task scheduling model.

[0089] The embodiment first constructs a large-scale, high-precision and fully open benchmark test suite SatData, which contains 3907 satellite data sets and 16410 scenes, each scene contains 1-50 satellites and 50-300 imaging tasks, covering 3600 time steps. The scene is generated by the Basilisk engine high-fidelity simulation platform, accurately simulates the physical characteristics of satellite orbit dynamics, attitude control, etc., introduces real satellite data testing, ensures that the constraint conditions and evaluation indexes are close to reality. The evaluation indexes cover six dimensions of task completion rate, turnaround time, power consumption, etc., provide expert annotated true value scheduling annotations, data is publicly accessible, solves the problem of lack of general evaluation standard in existing benchmarks, supports fair comparison of models.

[0090] Secondly, the Sat-Former scheduling model is constructed. Based on the Transformer architecture, the internal constraint module is embedded, the physical and operational limits of the satellite are modeled, the scheduling decision is guided, and the feasibility and fidelity of the generated scheme are improved. Combined with two-stage learning of supervised pre-training and simulation exploration, the model adapts to the dynamic environment, discovers better scheduling strategies, and enhances the generalization ability. The action space is divided into high-level task allocation and low-level control instruction, which reduces the complexity of the model and improves the ability of the model to capture time sequence features. Binary cross-entropy loss and mean square error loss are used for double supervision, realizing the deep fusion of constraint and feature matching, and improving the allocation efficiency.

[0091] The embodiment of the present application is significantly better than the existing techniques such as REDA, EOSSP-RCS in terms of task completion rate, power consumption, comprehensive score and other indicators. It promotes the construction of standardization and tool chain. SatData of the present application will become the first large-scale benchmark for real AEOS constellation scheduling. The open source code and data of Sat-Former promote technology reproduction and method innovation.

[0092] The specific steps of the embodiment will be described in detail below.

[0093] Firstly, a benchmark test suite SatData is constructed, which includes:

[0094] S1, obtaining a satellite dataset, constructing a satellite scenario model based on a satellite system model, a satellite mission definition, a satellite action space control and multiple constraint limits;

[0095] The embodiment with reference to Figure 2 is shown, which shows a process of a satellite mission from initial evaluation to final confirmation, including calculating MRP using an empirical formula, performing completion rate checking and manual quality checking, and deciding whether to accept or regenerate MRP according to the checking results. Finally, the satellite that passes all checks will be regarded as a qualified satellite asset. Thus, the satellite dataset in the embodiment is generated, which includes MRP formula calculation and multiple checks to ensure stable attitude control of each dataset.

[0096] The MRP empirical formula of the embodiment is:

[0097]

[0098]

[0099]

[0100]

[0101] wherein, represents a proportional constant of the MRP formula; represents an integral constant of the MRP formula; represents a differential constant of the MRP formula; represents an upper limit of integration; represents a maximum value of inertia tensor projection; represents a maximum angular momentum size that can be provided by all momentum wheels; α, β, γ, δ ~ U(0, 1) represent random variables of 0-1 uniform distribution, respectively.

[0102] In the embodiment, a specific satellite and mission are taken as an example,

[0103] The satellite information is:

[0104] Mass: 210.05 kg

[0105] Moment of inertia: 107.87 kg·m²

[0106] Angular momentum: 50.00 kg·m²·s - ¹

[0107] The mission point is:

[0108] Longitude: 0.000

[0109] Latitude: 5.239

[0110] Completion progress: 15 / 67

[0111] This embodiment conducts manual quality inspection after completion rate inspection;

[0112] This embodiment uses six indicators to evaluate the scheduler, including task completion rate, timeliness, and energy efficiency. The completion rate (CR) measures the proportion of completed tasks to the total number of tasks. The partial completion rate (PCR) evaluates the ratio of maximum progress to the total required duration. The weighted completion rate (WCR) is a weighted version of CR, taking into account task duration. The turnaround time (TAT) calculates the average time taken to complete a task, reflecting scheduling efficiency. The power consumption (PC) quantifies the total energy consumed by satellite sensors during imaging. Finally, the comprehensive score (CS) aggregates these indicators into a single performance indicator:

[0113]

[0114] In this embodiment, greater than 70% indicates passing; for example, CR: 5.00%; PCR: 16.23%; indicates passing. For example, CR: 90.00%; PCR: 93.28%; indicates passing. For suspicious cases, the MRP is regenerated, and the final passing satellite will be classified as a satellite asset as the satellite dataset of this embodiment.

[0115] This embodiment constructs the benchmark suite SatData, starting from defining the problem setting of AEOS constellation scheduling, and describes the generation of the true value scheduling annotation of the satellite dataset.

[0116] Regarding the process of problem setting, i.e., the construction process of the satellite scene model.

[0117] This embodiment constructs a satellite scene model based on a satellite system model, satellite task definition, satellite action space control, and multiple constraint limitations.

[0118] Among them, the satellite system model is composed of four core subsystems: orbital dynamics, attitude control, power system, and sensor payload. It is used to capture the basic physical characteristics that ensure the feasibility of the task. Specifically, it includes parameters such as satellite occupying low Earth orbit, orbital elements, mass characteristics, and inertia moments, as well as control gains in attitude control and actuator limitations for each satellite. Low control gain leads to slow attitude adjustment, while overloaded actuators may damage the stability of the satellite, increasing the risk of task failure.

[0119] This embodiment collects satellite features into a static satellite matrix where N L represents the number of satellites, This indicates the dimension of the satellite's static features.

[0120] Imaging missions arrive dynamically, and each mission is defined by its release time, expiration time, required observation duration, and ground target coordinates. In this embodiment, these mission descriptors form a static mission matrix. ,have One task and Task attributes.

[0121] Regarding satellite operations space control, this embodiment divides actions into two abstract levels, separating high-level scheduling from low-level control. The low-level operations space includes power on / off commands and attitude pointing instructions, which are directly assigned to the Basilisk engine to simulate battery cycling, sensor activation, and MRP-based attitude maneuvers. While this provides maximum control flexibility, it introduces excessive complexity to the scheduling model. Conversely, the high-level operations space in this embodiment consists of task assignment commands. The scheduler outputs an assignment vector. Each of them . This indicates that the satellite's i-sensor is powered off, while any This instructs satellite i to activate its sensors and redirect to the service mission. The platform automatically translates these high-level assignments into low-level commands, allowing the scheduling model to focus purely on task selection and timing.

[0122] Real-world AEOS constellation scheduling is subject to multiple constraints. Regarding these constraints, this embodiment enforces five constraints: dynamics, energy, FOV, continuity, and time window. Any high-level assignments that violate these constraints are rejected by the simulator, and only successful observations are recorded for later-stage benchmarking.

[0123] S2, based on the satellite scene model built in S1, generates and labels satellite mission scheduling data; resulting in a dataset SatData with constellation scheduling labels;

[0124] This embodiment refers to Figure 3 As shown, different mission scenarios are generated using satellite assets, involving mission allocation across multiple satellites at different times. A greedy algorithm is used to allocate initial missions and generate mission plans. Figure 3 The task allocation is displayed in a three-dimensional time-space grid: different rows represent different satellites, different columns represent different time points, and each square represents the task status of a satellite at a specific time point. The initial allocation is iteratively filtered according to the completion rate to generate the final satellite trajectories. The final SatData benchmark library with constellation scheduling annotations generated in this embodiment contains 16,000 predefined satellite trajectories for evaluating and optimizing mission planning.

[0125] After the benchmarking suite SatData is constructed, the scheduling model Sat-Former needs to be constructed, which specifically includes:

[0126] S3, constructing the scheduling model Sat-Former based on the Transformer architecture, and embedding the internal constraint module, satellite attribute matrix and task attribute matrix, for satellite task allocation;

[0127] The Transformer architecture in the embodiment only includes a decoder, as shown in FIG. 3, first, the static and dynamic data of the satellite and the task are spliced and embedded. The decoder focuses on the satellite embedding and the task embedding under cross attention. Then, the internal constraint module predicts the feasibility probability to guide the action selection. Figure 4

[0128] Figure 4 The satellite static data of the satellite in the embodiment includes satellite mass, satellite center of mass position, satellite orbit eccentricity, semi-major axis, inclination, ascending node right ascension, perigee amplitude angle, satellite solar panel attribute, satellite imaging device attribute and satellite momentum wheel attribute. Among them, the satellite solar panel attribute includes the direction of the solar panel relative to the satellite, the area of the solar panel, and the photoelectric conversion efficiency of the solar panel; the satellite imaging device attribute includes the load type (may have multiple types such as optical or infrared load), and the half field of view angle; the satellite momentum wheel attribute includes the momentum wheel static attribute (maximum momentum, motor efficiency, orientation).

[0129] Satellite dynamic data includes satellite true anomaly angle, battery attribute, load power, current momentum wheel speed and current attitude of the satellite. Among them, the battery attribute includes battery capacity and current battery percentage; the current attitude of the satellite is represented in the form of modified Rodrigues parameters (MRP).

[0130] Task static data of the task includes task release time, task deadline, continuous observation time, target latitude and longitude, observation task required imaging device load type, task completion status and task mask. As for the task completion status, if the task has been completed, it is 1, otherwise it is 0; and the task mask is used to mask illegal tasks. Task dynamic data includes time step and task progress.

[0131] Among them, as for the dynamic data, it is known that each scene in SatData is composed of a static satellite matrix L S and a static task matrix R S ​By definition, they capture time-invariant properties. Dynamic properties, such as task progress and satellite attitude, are not included in these static matrices. Enabling the scheduling model to infer dynamic states from static properties and past decisions would greatly increase complexity without a clear benefit. Instead, the present embodiment queries the simulator at each time step to retrieve the current dynamic satellite and task properties. The complete input matrix is formed by concatenating the static and dynamic components:

[0132]

[0133]

[0134]

[0135]

[0136] where, is the dynamic satellite matrix, is the dynamic task matrix; denotes the satellite matrix dimension, denotes the satellite matrix static parameter dimension, denotes the satellite matrix dynamic parameter dimension; denotes the task matrix dimension, denotes the task static parameter matrix dimension, denotes the task dynamic parameter matrix dimension.

[0137] In the present embodiment, the constellation properties at the current time step are as shown in Table 1:

[0138] Table 1. Constellation properties showcase

[0139]

[0140] The task properties to be completed are as shown in Table 2:

[0141] Table 2. Task properties showcase

[0142]

[0143] Regarding the internal constraints module, which is used to predict the feasibility of a satellite to perform a task; the present embodiment, for each satellite-task pair (i,j), the internal constraints module predicts the feasibility of satellite i to perform task j is expressed in formula as:

[0144]

[0145]

[0146]

[0147]

[0148] where, denotes the predicted logit of the feasibility of satellite i to complete task j; L i denotes the matrix of satellite i, R j denotes the matrix of task j; denotes the set of real numbers;

[0149] The embodiment also defines an approximate label , processes the task that needs to be completed by multiple satellites in cooperation; defines a feasibility loss function and introduces time supervision, further guiding the internalization constraints of the internal constraint module .

[0150] The feasibility loss function is expressed by the formula:

[0151]

[0152] This supervision strategy enables the learning of the feasibility of satellite task allocation, effectively capturing the constraints existing in the AEOS-Bench scenario.

[0153] After the model is constructed, according to S4, the scheduling model Sat-Former is trained and tested, and through reinforcement learning fine-tuning, the final satellite task scheduling model is obtained.

[0154] The scheduling model Sat-Former of the embodiment adopts a decoder architecture to match satellites and tasks, which jointly processes satellite and task embeddings guided by the internal constraint module. Its processing process is as follows.

[0155] First, S41, the embodiment projects S and T into the embedding space and appends a sinusoidal time step embedding , which is expressed by the formula:

[0156]

[0157]

[0158] where, denotes the satellite feature vector, denotes the task feature vector; and denote the embedding module, which looks up the classification data (for example, sensor mode) in the embedding matrix, while continuous classification data (for example, quality, progress) uses linear projection.

[0159] S42, decode the satellite feature vector and the task feature vector using the Transformer decoder to obtain satellite features h L and task features h R ;

[0160] S43, quantify the matching degree between the satellite and the task by assigning the score matrix M, and add the feasibility prediction of the embedded constraint module to guide the planning, which is expressed by the formula:

[0161]

[0162]

[0163] wherein, represents a trainable vector, represents a Hadamard product of element-wise multiplication, F represents a feasibility matrix obtained after 1 padding on , when , , otherwise is equal to ;

[0164] S44, define a model loss function to evaluate the current task allocation result and optimize the scheduling model Sat-Former; the model loss function is the loss function of the model allocation target:

[0165]

[0166] wherein, m is the true value of the task;

[0167] S45, in the test phase, exclude the satellite-task pairs that are not feasible when quantifying the assigned score matrix;

[0168] which is expressed by the formula:

[0169]

[0170] wherein, represents a satellite-task pair; represents an indicator function, represents a pre-defined feasibility threshold; represents the i-th row and j-th value of the assigned score matrix. The embodiment closely integrates the learned constraints and feature matching to achieve efficient satellite task allocation.

[0171] The embodiment constructs a reinforcement learning process, and the overall learning process is shown in Figure 5 In the supervised pre-training stage, Sat-Former is initialized with random weights, and is trained based on the labeled trajectories in SatData to obtain a total loss function​ , which is expressed by the formula:

[0172]

[0173] wherein, represents a total loss function; and respectively represent the weights of and .

[0174] In the reinforcement learning fine-tuning stage, comprehensive evaluation is performed according to the corresponding total loss function; and trajectories exceeding the feasibility threshold are added to the SatData for reinforcement learning of the Sat-Former until a final satellite task scheduling model is obtained by reaching model convergence, and the reinforcement learning specifically includes:

[0175] S46, before the reinforcement learning training starts, initializing a value function network wherein, is a learnable vector for value evaluation of the current state, guiding the Sat-Former to generate a better scheduling scheme.

[0176] S47, taking the Sat-Former as a policy function in reinforcement learning , representing the probability of selecting action a under state L, R, wherein, represents all learnable vectors in the Sat-Former, the Sat-Former is trained step by step, and the assignment score matrix output by the Sat-Former is used for probability sampling, which is expressed by the formula:

[0177]

[0178]

[0179] wherein, represents a satellite-task pair sampled at the current time step, i.e., an action in reinforcement learning, P represents a sampling probability, represents the i-th row and j-th value of the assignment score matrix.

[0180] S48, guiding the reinforcement learning through a reward function r, and performing reward calculation according to the current state, which is expressed by the formula:

[0181]

[0182] wherein, represents the number of satellites that have not been assigned tasks, denotes the number of tasks currently being observed, denotes the number of tasks that have completed observation.

[0183] S49, using a simulation platform, a state transition function is constructed which can receive the current state and the sampled satellite-task pair, and return the next new state and the reward, which is expressed by the formula:

[0184]

[0185]

[0186] wherein L t and R t respectively represent the satellite attribute matrix and the task attribute matrix at the t time step, denotes the task corresponding to the i-th satellite sampled at the t time step, denotes all satellite-task pairs sampled at the t time step.

[0187] S410, according to the calculated reward and the time difference error of the next state which is expressed by the formula:

[0188]

[0189] wherein, is a discount factor.

[0190] S411, according to the time difference error, the Sat-Former and the value function network are updated, which is expressed by the formula:

[0191]

[0192]

[0193] wherein, and denote the learning rate hyperparameter.

[0194] The pre-trained Sat-Former is used for inference to generate trajectories. Each generated trajectory is evaluated by the score defined in the total loss formula. Then, the embodiment collects the performance exceeding the predetermined threshold These high-quality trajectories are re-added to the SatData training set, and Sat-Former is then fine-tuned using the new dataset. This cycle is repeated until convergence. In this way, Sat-Former continually refines its strategy, discovers new strategies beyond the original annotations, and adapts to simulation-driven exploration, adapting to increasingly diverse scenarios.

[0195] The present application performs satellite scheduling through a unified framework that integrates a standardized benchmark suite and a new scheduling model. The benchmark suite SatData contains 3907 fine-tuned satellite datasets and 16410 scenarios. Each scenario has 1 to 50 satellites and 50 to 300 imaging tasks. These scenarios are generated through a high-fidelity simulation platform to ensure realistic satellite behavior, such as orbital dynamics and resource constraints. True value scheduling annotations are provided for each scenario. Based on this benchmark, the present application introduces Sat-Former, a Transformer-based scheduling model that includes an attention mechanism for perceiving constraints. The internal constraint module explicitly simulates the physical and operational limits of each satellite. Through simulation-based reinforcement learning, Sat-Former can adapt to different scenarios, providing a robust solution for AEOS constellation scheduling.

[0196] The benchmark suite SatData of the present application has four key features: 1) Large-scale. SatData includes 16,410 scenarios, each with 1 to 50 satellites, 50 to 300 imaging tasks, and 3,600 time steps. 2) Realistic. All scenarios are generated and evaluated on our simulation platform, ensuring accurate physical behavior of satellites. The test set uses real satellite data from public sources, allowing for evaluation on real data. 3) Comprehensive. SatData evaluates six indicators, including task completion rate, turnaround time, and power consumption. 4) Public data. Each scenario is annotated with true value tasks through a rigorous process. All benchmark data and annotations are publicly accessible.

[0197] Sat-Former of the present application is a Transformer-based scheduler designed specifically for the AEOS constellation. Its core is a dedicated internal constraint module that explicitly simulates the physical and operational limits of each satellite, including sensor field of view, battery state, and attitude control time. This module guides scheduling by predicting feasibility probabilities.

[0198] The various embodiments described in this specification are implemented in a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the apparatus disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0199] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those of ordinary skill in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A planning method for agile observation satellites, characterized in that, The method comprises the following steps: S1, acquiring a satellite dataset, constructing a satellite scene model based on a satellite system model, satellite task definition, satellite action space control and multiple constraint limits; S2, generating and labeling satellite task scheduling data based on the satellite scene model; Obtaining a data set SatData with constellation scheduling labels; S3, constructing a scheduling model Sat-Former based on a Transformer architecture, and embedding an internal constraint module, a satellite attribute matrix and a task attribute matrix, for satellite task allocation; S4, training and testing the scheduling model Sat-Former using the data set SatData with constellation scheduling labels, and obtaining a final satellite task scheduling model through reinforcement learning fine-tuning; S5, when planning a target task of an agile observation satellite, scheduling the corresponding agile observation satellite through the final satellite task scheduling model; In step S3, the satellite attribute matrix includes a static satellite matrix and a dynamic satellite matrix; and is expressed by a formula as follows: ； where L represents a satellite attribute matrix; L S represents a static satellite matrix; L d represents a dynamic satellite matrix; N L represents the number of satellites; d L represents a satellite feature dimension; The task attribute matrix includes a static task matrix and a dynamic task matrix; and is expressed by a formula as follows: ； wherein R represents a task attribute matrix, R S represents a static task matrix; R d represents a dynamic task matrix; N R represents the number of tasks; d R represents a task attribute dimension; In step S3, the internal constraint module is used to predict the feasibility of a satellite executing a task; and the execution process comprises: S31, for each satellite-mission pair , the internal constraints module predicting the feasibility of satellite i to perform mission j , is expressed by the formula: ； ； ； ； wherein, represents a predicted log of the feasibility of satellite i to complete task j; L i represents a matrix of satellite i, R j represents a matrix of task j; represents a set of real numbers; S32, defining an approximate label , the case where many satellites need to cooperate to complete many tasks; S33, define the feasibility loss function and introduce time supervision, further guide the internal constraint module internalization constraints; The feasibility loss function is expressed by a formula as follows: ； wherein, represents the feasibility loss; In step S4, the training and testing of the scheduling model Sat-Former using the data set SatData with constellation scheduling labels specifically comprises: S41, converting the L and R into vector forms and embedding a sinusoidal time step through an embedding module respectively; obtaining a satellite feature vector and a task feature vector; S42, decode the satellite feature vector and the task feature vector using a Transformer decoder to obtain a satellite feature h L and a task feature h R ; S43, quantifying the matching degree between satellites and tasks through an assignment score matrix M, and adding the feasibility prediction of the embedded constraint module to guide the planning; and is expressed by a formula as follows: ； ； wherein, represents a predicted log of the feasibility of satellite i to complete task j; represents a trainable vector, represents a Hadamard product of element-wise multiplication, F represents a feasibility matrix after padding with 1s; S44, defining a model loss function to evaluate the current task allocation result and optimize the scheduling model Sat-Former; the model loss function is a loss function of a model allocation target: ； wherein, represents the model loss, m represents the task true value; S45, in the test phase, when quantifying the assignment score matrix M, the unfeasible satellite-task pairs are excluded; is expressed by a formula as follows: ； wherein, denotes a satellite-mission pair; denotes an indicator function, denotes a predefined feasibility threshold; denotes the ith number of the jth value of the assignment score matrix.

2. The planning method for agile observation satellites according to claim 1, characterized in that, In step S1, the generation process of the satellite dataset comprises: Performing MRP empirical formula calculation on each satellite task, completing rate checking and manual quality checking, and deciding to accept or regenerate the MRP according to the checking result; Satellites that pass the checking are regarded as qualified satellite assets, and a satellite dataset is formed.

3. The method of claim 1, wherein, In step S1, the construction of the satellite scene model based on the satellite system model, the satellite task definition, the satellite action space control and the multiple constraint limits specifically comprises: Constructing a satellite system model based on orbit dynamics, attitude control, power system and sensor payload; Defining a satellite task based on release time, expiration time, observation duration and ground target coordinates; Controlling the satellite action space by high-level control scheduling low-level control; wherein the high-level control includes task allocation commands; and the low-level control includes power switch commands and attitude pointing instructions; The satellite scene model is constructed by constructing a kinetic restriction, an energy restriction, a field of view restriction, a continuity restriction and a time window restriction to restrict execution of the satellite task, and obtaining a satellite scene model including problem definition and scene.

4. The planning method for agile observation satellites according to claim 3, characterized in that, The step S2 specifically comprises: In the satellite scene model, an initial allocation is performed by using a greedy algorithm to obtain a preliminary allocation result. The preliminary allocation result is simulated, and an observation track with successful simulation is reserved. Through artificial quality review, a high-quality scheduling track is marked to obtain a data set SatData with constellation scheduling marking.

5. The method of claim 1, wherein, In step S4, the final satellite task scheduling model is obtained by fine-tuning through reinforcement learning, specifically comprising: In the supervised pre-training stage, the Sat-Former is initialized by using random weights, and the Sat-Former is trained based on the marked track in the SatData to obtain a total loss function. The total loss function is expressed by a formula as follows: ； wherein, represents the total loss function; and respectively represent the weights of and In the reinforcement learning fine-tuning stage, comprehensive evaluation is performed according to the corresponding total loss function; and trajectories exceeding the feasibility threshold are added to the SatData for reinforcement learning of the Sat-Former, until a final satellite task scheduling model is obtained by reaching model convergence.

6. The planning method for agile observation satellites according to claim 5, characterized in that, In step S4, the reinforcement learning specifically comprises: A value function network is initialized to evaluate the value of the current state, and the Sat-Former is taken as a policy function in the reinforcement learning to train the Sat-Former step by step, and the value score matrix output by the Sat-Former is used for probability sampling; The reinforcement learning is guided by a reward function, and a reward is calculated according to the current state; a time difference error is calculated according to the calculated reward and the next state; and the Sat-Former and the value function network are updated according to the time difference error.

Images