Agile satellite imaging task scheduling method and system based on data-driven

Abstract

The present invention provides a data-driven agile satellite imaging task scheduling method and system. Due to the complex constraints of satellite task scheduling, the satellite task sequences generated by genetic algorithm iterations often violate constraints and are infeasible. A heuristic algorithm is used to ensure that the generated satellite task sequence ultimately does not violate constraints, forming a satellite task scheduling plan. The start and end times of each task within its visible time window are determined. A neural network model is used to improve the fitness of the individuals generated during the genetic iteration process, thereby increasing the fitness of the entire population. To ensure population diversity, individuals improved by the neural network model are locally adjusted to achieve uniform population diversity and convergence. This reduces the algorithm's search time while quickly generating high-quality individuals, or planning solutions.

Description

Technical Field

[0001] The present invention belongs to the field of satellite mission planning, and in particular relates to a data-driven agile satellite imaging mission scheduling method and system. Background Art

[0002] As one of the key means of acquiring space-based information, satellites have played a significant role in national defense, the economy, and people's livelihoods. With the increasing number of satellites in orbit and the continuous improvement of satellite imaging capabilities, users from all walks of life have submitted a large number of complex observation requests to satellite control departments. To meet the diverse and large-scale observation needs, satellite mission scheduling has become a primary issue for satellite control departments. From an economic perspective, taking my country's first commercial agile remote sensing constellation, Gaojing-1, as an example, the Gaojing-1 remote sensing satellite can capture approximately 84 square kilometers of imagery per second, generating approximately ¥10,000 in economic benefits. This shows that reasonable and efficient satellite mission scheduling can bring huge economic benefits.

[0003] The satellite mission scheduling problem refers to the satellite management and control department arranging a reasonable mission execution sequence and mission execution start time for the satellite based on the observation requirements put forward by different users without violating the satellite usage constraints. The visible time window refers to the time range in which the satellite can observe the target. Only when the satellite observes the target within the visible time window can the observation mission be meaningful. In a certain period, the satellite orbits the earth multiple times, so there may be multiple visible time windows for the satellite to observe the target. Therefore, the essence of the satellite mission scheduling problem is to arrange the visible time window and determine the mission start time within the time window. Compared with traditional satellites, agile satellites have two more degrees of freedom, pitch and yaw, and the visible time window between the satellite and the target is longer, which makes the agile satellite mission scheduling problem (hereinafter referred to as satellite mission scheduling) more difficult to solve.

[0004] At present, the methods for solving satellite mission scheduling problems can be roughly divided into three categories. The first is the typical exact solution algorithm based on branch and bound and dynamic programming. Among them, the branch and bound (B&B) algorithm was proposed by Land et al. in 1960 and formally named by Little et al. in 1963. It is an exact solution algorithm. The B&B algorithm narrows the solution space by means of branching, bounding and pruning, and then searches for the optimal solution through each branch. It is one of the most commonly used algorithms for solving integer linear programming problems at this stage. Since satellite mission scheduling problems are usually simplified and modeled as linear programming models, the branch and bound algorithm is widely used in solving satellite mission scheduling problems. Xiaoyu Chen et al. established a mixed integer programming model for multi-satellite mission scheduling problems under limited observation capabilities, and used the B&B algorithm to solve it. Yiyong Xiao et al. considered the joint planning problem of satellite imaging and data transmission tasks under weather uncertainty, established a mixed integer programming model for the problem and used the B&B algorithm to solve it; for the problem of maritime target recognition, Xiaogeng Chu adopted a dual-satellite constellation observation mode composed of low-resolution satellites (LRS) guiding high-resolution satellites (HRS), established a linear programming model for the problem and proposed an ananytime branch and bound (AB&B) algorithm. Experimental results show that in a scenario with 25 targets and a 500km×2000km sea area, the performance of the dual-satellite constellation is equivalent to the collaborative observation of 30 high-resolution satellites. At the same time, in order to improve the solution efficiency in large-scale task scheduling problems, the B&B algorithm is often used together with the column generation method, cutting plane method and Lagrangian relaxation method to solve satellite task scheduling problems or boundary solutions. The dynamic programming algorithm (DP) is an exact solution algorithm proposed by Bellman et al. in 1965 that searches for the optimal solution through problem decomposition and recursive means. For the remote sensing satellite imaging task scheduling problem, et al. employed a DP algorithm based on a graph theory model to solve the problem. Damiani et al. designed an evaluation vector that includes the current mission, satellite power, and storage capacity, and decomposed the problem into an optimal task combination problem based on the evaluation vector. Furthermore, Peng et al. employed dynamic programming to decode the imaging start time of agile remote sensing satellites within the visible time window during an iterative search, replacing traditional heuristic decoding rules such as earliest start and image quality priority. The problem decomposition approach employed by these DP algorithms is also valuable for solving large-scale satellite mission scheduling problems. Overall, while exact solution algorithms can theoretically obtain optimal solutions, they are limited in their applicability to small and medium-scale scenarios and require simplified problem constraints to establish a linear programming model, which clearly cannot meet practical engineering needs. A second approach is rule-based heuristic algorithms, which typically allocate resources and time to a task sequence based on priority and other rules. These algorithms have the advantages of simple structure, fast computational speed, and are consistent with human subjective experience. They are commonly used in satellite mission scheduling research and in practical scheduling systems. The sorting rules in these priority sorting algorithms are typically based on task priority, but also include some combined priority orders related to attributes such as VTW time, length, and quantity. Common resource allocation rules include the earliest start principle, the latest start principle, the maximum image quality principle, and the minimum possible conflict principle. Heuristic algorithms effectively reduce the decision-making dimension and solution difficulty of the problem, but their application effectiveness depends largely on the rationality of the algorithm design. Most are tightly coupled to the problem scenario, task, and resource characteristics, resulting in insufficient versatility. Metaheuristic algorithms, thirdly, have been applied to TSP, VRP, and OP problems with good results. They are also widely used in satellite mission scheduling, especially for medium- to large-scale missions. Liu et al. proposed an adaptive large neighborhood search (ALNS) algorithm. Its insertion strategy considers temporal dependencies by introducing forward / backward time relaxations, enabling rapid task insertion. He built on Liu's approach by incorporating a tabu search algorithm, fundamentally improving the short-term loop problem that can occur in the ALNS algorithm and extending the problem from single-satellite to multi-satellite coordination. Du proposed an adaptive parallel memetic algorithm, in which a parallel local search strategy overcomes the drawback of traditional local search algorithms, which often struggle to obtain diverse solutions. Different local search algorithms are used to generate different search trajectories and solution sets, thereby increasing solution diversity. Similarly, traditional metaheuristic algorithms, such as genetic algorithms, ant colony algorithms, and simulated annealing, have been applied in the field of satellite mission scheduling, achieving a range of relevant results. However, the main drawbacks of metaheuristic algorithms are their randomness, prone to premature convergence or slow convergence, and low relevance to the problem itself. Furthermore, the algorithm parameters require extensive prior experimentation, making them less generalizable.To address this issue, researchers both domestically and internationally have incorporated knowledge models into traditional metaheuristic algorithms to guide algorithmic search. Some studies have combined knowledge models such as elite individual knowledge, construction knowledge, operator knowledge, and parameter knowledge with genetic algorithms, achieving promising results. However, these methods rely heavily on prior information to mine explicit knowledge about the problem or algorithm, lacking the capacity to capture implicit knowledge. With the rise of machine learning, artificial neural networks have demonstrated superior learning capabilities and are well-suited to capturing implicit knowledge. While some studies have directly used the parent population as neural network input and the offspring population as output, this approach suffers from the drawback of completely replacing the genetic algorithm, losing population diversity and its inherent advantages. For continuous optimization problems, some have used neural networks to improve individuals with lower fitness within a population, but their effectiveness has not been validated for combinatorial optimization problems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to quickly provide an agile satellite scheduling solution for large-scale observation tasks. A data-driven agile satellite imaging task scheduling method and system are proposed.

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

[0007] A data-driven agile satellite imaging task scheduling method includes the following steps:

[0008] Step 1: Get the task set;

[0009] Step 2: Construct an agile satellite task scheduling model. In this model, the objective function maximizes the total observed benefit after task set planning. The decision variables include whether a task is executed within a time window and the start execution time of a task within a time window.

[0010] The constraints include that a task will only be executed once, the actual execution range of the task must be within the visible time window, the task has only one predecessor and one successor, and the task execution is constrained by power, task execution duration, storage, and attitude transition time.

[0011] Step 3: Solve the agile satellite mission scheduling model using a data-driven satellite mission scheduling problem solving algorithm;

[0012] Step 4: Output the optimal solution as a scheduling plan.

[0013] Furthermore, the data-driven algorithm for solving the satellite mission scheduling problem refers to:

[0014] Step 3.1: Initialize the population using a standard genetic algorithm, and encode the individuals in the population using real numbers;

[0015] Step 3.2: For the current population P t Individuals in P use heuristic algorithms to check constraints and delete tasks that violate constraints in individuals. t represents the t-th generation population;

[0016] Step 3.3: Check the current population P after the constraint t Perform selection, crossover, and mutation to generate a new offspring population P t _c;

[0017] Step 3.4: From the new offspring population P t A certain proportion of individuals are randomly selected from _c and improved using the neural network model. If the fitness of the individual after the output of the neural network model is greater than the fitness of the individual before the input, the individual after the output replaces the individual before the input to obtain a new population P t _new;

[0018] Step 3.5: If the iteration termination condition is not met, return to step 3.2. Otherwise, calculate the objective function value of each individual in the current population and take the individual with the highest objective function value as the optimal solution output.

[0019] Furthermore, in step 3.4, a new population P is obtained. t _new, it is also necessary to calculate the new population P t _new individual similarity, locally adjust the individuals whose similarity is greater than the preset threshold, and obtain the adjusted new population as the current population.

[0020] Furthermore, the sample set used by the neural network model is:

[0021] Step 3.4.1: Select from tN t Generation population By the tth generation population P t All individuals in the generated individual set U X , represents the t-th generation population P t The first N t Generation population;

[0022] Step 3.4.2: According to the individual set U X Generate target solution set U T and dataset U R , the target solution set U T refers to the individual set U X The dataset U is a set of individuals that meet the target number of solutions and whose Euclidean distance between individuals is greater than the preset threshold ζ. R refers to the individual set UX The set of individuals remaining after removing the target solution set;

[0023] Step 3.4.3: Calculate the dataset U separately R Each individual in the target solution set U T The Euclidean distance between different target solutions is used to select the target solution with the smallest distance to match the individual, which is used as a set of input and output individuals of the neural network. All matched individuals are combined to generate the sample set U s .

[0024] Furthermore, during the population crossover and mutation process, every N t A new sample set is formed by generations. After the neural network model in step 3.4 is trained with the new sample set, the new neural network model is evaluated. The new neural network model that passes the evaluation updates the original neural network model for individual improvement.

[0025] Furthermore, the method for calculating individual similarity is:

[0026]

[0027] Among them, S i represents the similarity of the i-th individual, PN represents the number of individuals in the population, GN represents the number of individual gene bits, C ij It represents the number of individuals whose j-th gene position has the same value as the remaining PN-1 individuals at the j-th gene position, and T represents the number of individual genes.

[0028] Furthermore, the local adjustment method is:

[0029] 1): Pre-set several local adjustment strategies and make the initial probabilities of the local adjustment strategies the same;

[0030] 2) For the individuals improved by the neural network, similarity is judged. For the individuals whose similarity is greater than the preset threshold, the local adjustment strategy is used n times and the individual with the largest objective function value is selected. * Replace the original individual;

[0031] 3): Recalculate the probability of using each local adjustment strategy in the next round based on the contribution of the local adjustment strategy.

[0032] Furthermore, the contribution degree is calculated as follows:

[0033] i represents the number of local adjustment strategies

[0034] P num represents the total number of individuals after the neural network is improved, Indicates the proportion of individuals whose similarity exceeds the threshold, Indicates the number of individuals that replace the original individuals with the largest objective function value contributed by the i-th local adjustment strategy, Represents the contribution of the i-th local adjustment strategy.

[0035] Furthermore, when calculating the Euclidean distance between individuals, the individuals are decoded into task execution sequences, and the Euclidean distance of the task sequence is used to represent the distance between individuals.

[0036] The present invention also provides a data-driven agile satellite imaging task scheduling system, which uses the steps in the scheduling method described above.

[0037] By adopting the above technical solution, the present invention has the following beneficial effects:

[0038] The present invention provides a data-driven agile satellite imaging task scheduling method and system. Due to the complex constraints of satellite task scheduling, the satellite task sequences generated by genetic algorithm iterations often violate constraints and are infeasible. A heuristic algorithm is used to ensure that the generated satellite task sequence ultimately does not violate the constraints, forming a satellite task scheduling plan. The start and end times of each task within its visible time window are determined. A neural network model is used to improve the fitness of the individuals generated during the genetic iteration process, thereby increasing the fitness of the entire population. To ensure population diversity, individuals improved by the neural network model are locally adjusted to achieve uniform population diversity and convergence. This method reduces the algorithm's search time and improves planning efficiency while quickly generating high-quality individuals, or planning solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the system of the present invention;

[0040] Figure 2 Schematic diagram of the solution algorithm flow;

[0041] Figure 3 Schematic diagram for individual coding;

[0042] Figure 4 Schematic diagram of the principle of generating neural network sample sets;

[0043] Figure 5 Schematic diagram of individual similarity detection;

[0044] Figure 6 Schematic diagram of random partial sequence crossover;

[0045] Figure 7 Schematic diagram of individual local random permutation;

[0046] Figure 8 Schematic diagram of local random frameshifting in the body. DETAILED DESCRIPTION

[0047] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0048] The agile satellite mission scheduling problem is described as follows: given a satellite and a set of ground observation targets, the satellite's orbital parameters are known, and each observation target corresponds to a different observation benefit. The satellite flies along a fixed orbit (with fixed orbital parameters). Since both the satellite orbit and the targets are fixed, the visibility window can be calculated in advance through preprocessing and other methods. The scheduling goal is to arrange the satellite's observation targets in the order and observation start time to maximize the observation benefit. To facilitate the development of the satellite mission scheduling model and to reduce the scale and difficulty of the problem, this paper makes the following assumptions:

[0049] 1) The satellite carries only one imaging payload;

[0050] 2) Once a task starts executing, it cannot be interrupted or preempted by other tasks;

[0051] 4) The task can be executed at most once and cannot be executed repeatedly;

[0052] 5) The mission is considered to be successfully completed if the target can be successfully observed by the satellite, regardless of the quality of the captured image;

[0053] 6) In this paper, satellite observation missions refer to meta-tasks that can be directly executed by satellites. Complex missions can ultimately be decomposed into meta-tasks. Taking regional targets as an example, large areas can be decomposed into meta-tasks that satellites can observe multiple times through.

[0054] Figures 1 to 8 A specific embodiment of the data-driven agile satellite imaging task scheduling method of the present invention is shown. Figure 1 As shown, the following steps are included:

[0055] Step 1: Get the task set;

[0056] Step 2: Construct an agile satellite task scheduling model. In this model, the objective function maximizes the total observed benefit after task set planning. The decision variables include whether a task is executed within a time window and the start execution time of a task within a time window.

[0057] The constraints include that a task can only be executed once, the actual execution range of the task must be within the visible time window, the task has only one predecessor and one successor, and the task execution is constrained by power, storage and attitude transition time;

[0058] The agile satellite mission scheduling model is:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] Among them, formula (1) is the objective function. Even if the total observed benefit is maximized, x ij is a 0-1 decision variable, indicating the i-th task t i In the jth visible time window w ij 1 means execution, 0 means not execution; i is the i-th task t i Observed return value of ;

[0073] w={w ij |1≤i≤N,1≤j≤|w i |},w ij Represents task t i The jth visible window, w i For task t i The set of visible windows, |w i |For task t i The number of visible windows, w ijBy a triple (st ij ,et ij , d ij ),of which St ij 、et ij Represent the time window w ij The start and end time of d ij represents the time window w ij The length of time; N represents the number of tasks;

[0074] Formula (2) indicates that a task can only be executed once, T = (t i |1≤i≤N} represents the set of tasks to be observed, and the i-th task t i A four-tuple (p i , m i , e i , I i ) composition, m i It is task t i The storage capacity occupied per unit time, e i It is task t i The amount of electricity consumed per unit time; i It is task t i The execution time of the task is: Formula (3) indicates that the actual execution range of the task must be within the visible time window, u ij represents the start execution time of the i-th task in the j-th visible time window, which is a decision variable; l i Indicates the execution time of the i-th task;

[0075] Formula (4) represents the transition time constraint between tasks. The start execution time of the next task needs to consider the posture transition time after the previous task is completed. is a binary intermediate variable, Represents task t i It's a task The previous task, otherwise 0, Indicates a task The jth * time window, Indicates a task The set of visible windows, Indicates the i * tasks In the jth * The start execution time of the visible time window, Indicates a task Switch to task t i conversion time;

[0076] Formulas (6), (7), and (8) indicate that a task has only one predecessor and one successor, that is, the execution of tasks is sequential, where w e and w s are virtual mission nodes, representing the last node and the first node of the satellite respectively. For the task The number of visible windows,

[0077] Formula (9) represents the storage constraint, m i It is task t i The storage capacity occupied per unit time; M represents the total satellite storage;

[0078] Formula (10) represents the power constraint, e i It is task t i The amount of electricity consumed per unit time; p s , p a They represent the power consumed by the satellite for each attitude conversion and the power consumed per degree when performing attitude conversion; E represents the total power of the satellite;

[0079] Equation (11) represents the satellite switching time, which is a piecewise function, Δ g Indicates the attitude conversion angle between two observations, v1 v2 v3 v4 represents the conversion angle of four different attitudes;

[0080] Formula (12) shows They represent the satellite's position at time u ij Roll, pitch, and yaw angles;

[0081] Step 3: Solve the agile satellite mission scheduling model using a data-driven satellite mission scheduling problem solving algorithm.

[0082] In this embodiment, the satellite mission scheduling problem has been proven to be an NP-Hard problem, so there is no exact algorithm that can solve it in polynomial time. Genetic algorithm, as one of the most classic meta-heuristic algorithms, has been widely used to solve the satellite mission scheduling problem, but it also has common problems with meta-heuristic algorithms. In order to solve the problems of premature or late convergence of genetic algorithms and random search, this paper proposes a data-driven algorithm for solving the satellite mission scheduling problem based on the standard genetic algorithm framework. The basic framework of the algorithm is as follows: Figure 2 The data-driven algorithm for solving the satellite mission scheduling problem is as follows:

[0083] Step 3.1: Use the standard genetic algorithm to initialize the population. The individuals in the population are coded with real numbers. Individual coding: use the following code: Figure 3The real number encoding method shown is then decoded to determine the execution order of the satellite observation tasks. For the satellite observation task set T = {t1, t2, t3...t n}, randomly generate n numbers from 0 to 1 corresponding to each task in the task set, and determine the execution order of the tasks in ascending order of random numbers. Figure 3 Indicates that the task execution order is t1, t3, t2, t5, t6, t4.

[0084] Step 3.2: For the current population P t Individuals in P use heuristic algorithms to check constraints and delete tasks that violate constraints in individuals. t represents the t-th generation population;

[0085] In this embodiment, the satellite task scheduling problem has complex constraints, and the satellite task sequences generated by the algorithm iterations often violate the constraints and are infeasible. A construction heuristic algorithm generates a satellite task scheduling solution based on the satellite task sequence that ultimately does not violate the constraints, and determines the start and end times of each task within its visible time window. The core of the construction heuristic algorithm is to sequentially schedule tasks at the earliest possible location for execution, i.e., front-loading. This approach has the advantage of reserving sufficient time for the scheduling of subsequent tasks. Simultaneously, when scheduling tasks, it is necessary to consider constraints such as power consumption, storage capacity, task execution duration, and task transition time to ensure the feasibility of the resulting satellite task scheduling solution.

[0086] In this embodiment, the specific steps of the heuristic algorithm are:

[0087] Step 1: Determine whether the duration of each task in the satellite task sequence is less than the time window length. If it is less than the time window length, delete the task.

[0088] Step 2: For the satellite task sequence, the start time of each task is determined in sequence by taking a tight scheduling approach. The constraints of power, storage, and conversion time are checked. If there is no conflict, the task is added to the satellite task scheduling plan. Otherwise, the task is deleted.

[0089] Step 3.3: Check the current population P after the constraint t Perform selection, crossover, and mutation to generate a new offspring population P t In this example, the crossover method uses Partial-Mapped Crossover (PMX), randomly selecting a gene segment from two individuals at the same position and swapping them. Constraint conflicts are then checked and fixed for the new individuals. In satellite mission scheduling, a task is executed at most once, so it is necessary to check whether individuals have the same gene. The mutation method uses random exchange, randomly selecting two genes from an individual and swapping them to generate a new individual.

[0090] Step 3.4: From the new offspring population P t A certain proportion of individuals are randomly selected from _c and improved using the neural network model. If the fitness of the individual after the output of the neural network model is greater than the fitness of the individual before the input, the individual after the output replaces the individual before the input to obtain a new population P t In this embodiment, the individual fitness is the objective function value.

[0091] In this embodiment, the neural network model adopts a fully connected multi-layer neural network structure, including an input layer, an output layer, and two hidden layers. The nodes of the two hidden layers are set to three times the scale of the satellite observation task. The neural network optimizer selects the adaptive moment estimation algorithm, the loss function is set to the mean square error, and the number of training rounds is set to 50. The number of nodes in the input and output layers is the same as the task scale. The individuals output by the neural network are then decoded to obtain the task execution sequence. Whether the neural network works significantly depends to a large extent on the quality of the sample set. How to generate a high-quality data set for neural network training is the main challenge faced by this aspect. The genetic algorithm continuously maintains a population through operations such as crossover and mutation during each iteration. As the number of iterations increases, the population continues to evolve, which provides a natural condition for generating a sample set for neural network training.

[0092] In this embodiment, the sample set used by the neural network model is:

[0093] Step 3.4.1: Select from tN t Generation population By the tth generation population P t All individuals in the generated individual set U X , represents the t-th generation population P t The first N t Generation population;

[0094] Step 3.4.2: According to the individual set U X Generate target solution set U T and dataset U R , the target solution set U T refers to the individual set U X The dataset U is a set of individuals that meet the target number of solutions and whose Euclidean distance between individuals is greater than the preset threshold ζ. R refers to the individual set U XThe set of individuals remaining after removing the target solution set. In this embodiment, the preset threshold ζ is mainly used to ensure the diversity of individuals predicted by the neural network and avoid convergence to similar individuals. The target solution set is used as the label of the sample set. In addition, since individuals are encoded as real numbers from 0 to 1, the Euclidean distance between individuals is small and the differences are small. Therefore, when calculating the Euclidean distance between individuals, the individuals are decoded into task execution sequences, and the Euclidean distance of the task sequence is used to represent the distance between individuals.

[0095] Step 3.4.3: Calculate the dataset U separately R Each individual in the target solution set U T The Euclidean distance between different target solutions is used to select the target solution with the smallest distance to match the individual, which is used as a set of input and output individuals of the neural network. All matched individuals are combined to generate the sample set U s .

[0096] like Figure 4 As shown in the figure, the solid circle represents the target solution set, the hollow circle represents the data set, and the final sample set is a one-to-one correspondence between the data set and the target solution set.

[0097] In this embodiment, during the population crossover mutation process, every N t The new sample set is formed by generations. After the neural network model in step 3.4 is trained with the new sample set, the new neural network model is evaluated. The new neural network model that passes the evaluation updates the original neural network model to improve the individual. As the iterative process progresses, the quality of individuals in the population continues to improve. The neural network model trained with the original sample set can no longer make greater improvements to the individual. Therefore, the new neural network model is evaluated every N generations. t The new sample set is formed by generations to train the neural network model, so that the neural network model can adapt to the current individual and be improved.

[0098] Although the fitness of individuals improved by the neural network, i.e., the objective function value, is significantly improved, decoding reveals that the improved satellite mission sequences are relatively small, thus compromising population diversity. To further enhance population diversity, it is necessary to generate new forces based on the current satellite mission sequence. In addition to increasing the number of target solutions (labels) and increasing the distance threshold ζ, a competition-based adaptive local adjustment strategy is proposed. While executing the local adjustment algorithm on all individuals improved by the neural network would likely yield more distinct satellite mission sequences, it would be time-consuming. Often, only local adjustments to individuals with relatively small differences can achieve the same effect. Therefore, it is necessary to identify individuals with relatively small differences for local adjustments.

[0099] In this embodiment, after obtaining the new population P tAfter _new, it is also necessary to calculate the individual similarity of the new population P t _new, and perform local adjustment on individuals with similarity greater than the preset threshold to obtain the adjusted new population. Individual similarity is used to describe whether the differences between individuals in the population are obvious and is an index reflecting population diversity. In this embodiment, individual similarity is calculated to identify individuals with small differences. The method of individual similarity is as follows:

[0100]

[0101] where S i represents the similarity of the i-th individual. As Figure 5 shown, P N represents the number of individuals in the population, C ij represents the number of individuals with the same value as the remaining P N -1 individuals at the j-th gene position of the i-th individual at the j-th gene position, and T represents the number of individual genes.

[0102] The method of local adjustment is as follows:

[0103] 1): Preset several local adjustment strategies and make the initial probabilities of the several local adjustment strategies the same.

[0104] The local adjustment algorithm tries to keep part of the sequence of the improved individual while making the fitness value of the output individual as high as possible. In this embodiment, three local adjustment strategies are adopted to change the diversity of the population.

[0105] Strategy 1: Random partial sequence crossover. As Figure 6 shown, for individuals with small differences after the neural network improvement, that is, individuals with large individual similarity, perform partial sequence crossover with the originally input individuals, and select the individual with a larger fitness to replace the individual after the neural network output.

[0106] Strategy 2: Individual local random permutation. As Figure 7 shown, for individuals with small differences after the neural network improvement, assuming the individual length is L, randomly select two positions a, b (a < b) of the individual, and select to exchange the [1, a] and [b, L] segments of the individual to generate a new individual to replace the current individual.

[0107] Strategy 3: Individual local random frame shift. For individuals with small differences after the neural network improvement, randomly select two positions a, b (a < b) of the individual, and generate a new individual to replace the current individual in the manner of Figure 8 .

[0108] 2): Perform similarity judgment on the individuals improved by the neural network, and use the local adjustment strategy n times for individuals with similarity greater than the preset threshold, and then select the individual X with the largest objective function value.* Replace the original individual;

[0109] 3): Recalculate the probability of using each local adjustment strategy in the next round based on the contribution of the local adjustment strategy. The contribution calculation method is:

[0110]

[0111] P num represents the total number of individuals after the neural network is improved, Indicates the proportion of individuals whose similarity exceeds the threshold, Indicates the number of individuals that replace the original individuals with the largest objective function value contributed by the i-th local adjustment strategy, Represents the contribution of the i-th local adjustment strategy.

[0112] Step 3.5: If the iteration termination condition is not met, return to step 3.2. Otherwise, calculate the objective function value of each individual in the current population and take the individual with the highest objective function value as the optimal solution output.

[0113] Step 4: Output the optimal solution as a scheduling solution. In this embodiment, the optimal solution output in step 3.5 needs to be decoded to form a task sequence and then output as a scheduling solution.

[0114] The present invention also provides a data-driven agile satellite imaging task scheduling system, which uses the steps in the scheduling method described above.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven agile satellite imaging task scheduling method, characterized in that: The following steps are involved: Step 1: Get the task set; Step 2: Construct an agile satellite task scheduling model. In this model, the objective function maximizes the total observed benefit after task set planning. The decision variables include whether a task is executed within a time window and the start execution time of a task within a time window. The agile satellite mission scheduling model is: Among them, formula (1) is the objective function, even if the total observed benefit is maximized, is a 0-1 decision variable, representing the i-th task In the jth visible time window 1 means execution, 0 means not execution; is the i-th task Observed return value of ; , Indicates a task The jth visible window of For the task t i The set of visible windows, For the task The number of visible windows, By a triple Composition, of which Represents time windows The start and end time of Represents a time window The length of time; N represents the number of tasks; Formula (2) indicates that a task can only be executed once. Represents the set of tasks to be observed, the i-th task By a four-tuple composition, It's a task The storage capacity occupied per unit time, It's a task The amount of electricity consumed per unit time; Formula (3) indicates that the actual execution range of the task must be within the visible time window. It represents the start execution time of the i-th task in the j-th visible time window and is a decision variable; It's a task Execution time; Formula (4) represents the time constraint for switching between tasks. The start execution time of the next task needs to consider the posture conversion time after the previous task is completed. is a binary intermediate variable, Indicates a task It's a task The previous task, otherwise 0, Indicates a task No. time window, Indicates a task The set of visible windows, Indicates the tasks In the The start execution time of the visible time window, Indicates a task Switch to Task conversion time; Formulas (6), (7), and (8) indicate that a task has only one predecessor and one successor, that is, the execution of tasks is sequential, where and are virtual mission nodes, representing the last node and the first node of the satellite respectively. For the task The number of visible windows; The constraints include that a task will only be executed once, the actual execution range of the task must be within the visible time window, the task has only one predecessor and one successor, and the task execution is constrained by power, task execution duration, storage, and attitude transition time; Step 3: Solve the agile satellite mission scheduling model using a data-driven satellite mission scheduling problem solving algorithm; Step 4: Output the optimal solution as a scheduling plan.

2. The scheduling method according to claim 1, characterized in that: The data-driven algorithm for solving satellite mission scheduling problems refers to: Step 3.1: Initialize the population using a standard genetic algorithm, and encode the individuals in the population using real numbers; Step 3.2: For the current population The individuals in use heuristic algorithms to check constraints and remove tasks that violate the constraints in the individuals. Indicates the t generation population; Step 3.3: Check the current population after constraints Perform selection, crossover, and mutation to generate a new offspring population ; Step 3.4: From the new offspring population A certain proportion of individuals are randomly selected from the population and improved using the neural network model. If the fitness of the individual after the output of the neural network model is greater than the fitness of the individual before the input, the individual after the output replaces the individual before the input to obtain a new population. ; Step 3.5: If the iteration termination condition is not met, return to step 3.

2. Otherwise, calculate the objective function value of each individual in the current population and take the individual with the highest objective function value as the optimal solution output.

3. The scheduling method according to claim 2, characterized in that: The new population is obtained in step 3.4 After that, we need to calculate the new population The individual similarity of the individuals is calculated, and the individuals whose similarity is greater than the preset threshold are locally adjusted to obtain the adjusted new population as the current population.

4. The scheduling method according to claim 3, characterized in that: The sample set used by the neural network model is: Step 3.4.1: Select Generation population To Generation population All individuals in, generate individual set , Indicates the t Generation population Before Generation population; Step 3.4.2: Based on individual collection Generate target solution set and datasets , the target solution set refers to a collection of individuals The dataset is a set of individuals that meet the target number of solutions and whose Euclidean distance between individuals is greater than the preset threshold ζ. refers to a collection of individuals The set of individuals remaining after removing the target solution set; Step 3.4.3: Calculate the data sets separately Each individual in the target solution set The Euclidean distance between different target solutions is used to select the target solution with the smallest distance to match the individual, which is used as a set of input and output individuals of the neural network. All matched individuals are combined to generate a sample set. .

5. The scheduling method according to claim 4, characterized in that: During the population crossover and mutation process, every A new sample set is formed by generations. After the neural network model in step 3.4 is trained with the new sample set, the new neural network model is evaluated. The new neural network model that passes the evaluation updates the original neural network model for individual improvement.

6. The scheduling method according to claim 3, characterized in that: The method for calculating individual similarity is: in, Indicates the i The similarity of individuals, PN represents the number of individuals in the population, GN represents the number of gene positions of an individual, Indicates the Individuals in The remaining PN-1 Individuals in the The number of genes with the same value at each gene position, and T represents the number of individual genes.

7. The scheduling method according to claim 6, characterized in that: The local adjustment method is: 1): Pre-set several local adjustment strategies and make the initial probabilities of the local adjustment strategies the same; 2): For individuals improved by the neural network, similarity is judged, and a local adjustment strategy is used for individuals whose similarity is greater than the preset threshold. After that, the individual with the largest objective function value is selected Replace the original individual; 3): Recalculate the probability of using each local adjustment strategy in the next round based on the contribution of the local adjustment strategy.

8. The scheduling method according to claim 7, characterized in that: The contribution is calculated as follows: , i Indicates the number of local adjustment strategies represents the total number of individuals after the neural network is improved, Indicates the proportion of individuals whose similarity exceeds the threshold, Indicates the The number of individuals that replace the original individuals with the largest objective function value contributed by the local adjustment strategy, Indicates the The contribution of a local adjustment strategy.

9. The scheduling method according to claim 4, characterized in that: When calculating the Euclidean distance between individuals, the individuals are decoded into task execution sequences, and the Euclidean distance of the task sequence is used to represent the distance between individuals.

10. A data-driven agile satellite imaging task scheduling system, characterized in that: Use the steps in the scheduling method according to any one of claims 1 to 9.

Images