Star swarm autonomous cooperative search and tracking method based on dynamic information gain

Abstract

The application provides a star group autonomous cooperative search and tracking method based on dynamic information gain, which comprises the following steps: acquiring satellite constellation parameters and target initial state information; long-time sequence tasks are decomposed into continuous short-time sequence decisions based on a closed-loop rolling horizon planning framework, the target state distribution is predicted to obtain the uncertainty degree and generate an atomic task candidate set in each rolling period; the tracking or search mode is adaptively selected according to the uncertainty degree and the value is evaluated; the optimal task scheme is generated based on a multi-objective evolutionary algorithm; and the observation is performed and the target state is updated according to the observation result to enter the next period. The application dynamically adapts to the target uncertainty evolution through the closed-loop rolling horizon planning, avoids tracking rupture, uniformly measures the value of search and tracking through adaptive dual-mode cooperative decision, solves the value conflict problem, realizes the space-time optimal configuration of star group resources through the multi-objective evolutionary algorithm, and improves the continuous tracking capability and resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of satellite mission planning and space information system technology, and in particular to a method for autonomous collaborative search and tracking of satellite constellations based on dynamic information gain for maneuvering targets on the sea surface. Background Technology

[0002] Searching for and tracking maneuvering targets at sea has become a core task in constellation mission planning. Traditional static observation methods are insufficient to handle the spatiotemporal dynamics of these targets, which can easily escape the observation area during satellite revisits by maneuvering. To address this challenge, modern remote sensing systems are gradually developing towards multi-mode collaboration, utilizing constellation multi-functional payloads with the ability to switch between wide-field-of-view search and narrow-field-of-view tracking to achieve continuous tracking of large-scale targets through multi-platform synergy.

[0003] However, existing constellation cooperative planning technologies face extremely challenging uncertainties and resource game problems in practical applications. On the one hand, the motion of maneuvering targets is highly random, and their position probability density spreads non-linearly over time in a Gaussian manner, requiring the system to have the cognitive ability to quickly switch modes for re-acquisition the moment the target is lost. On the other hand, the field-of-view switching of constellation multi-functional payloads is strongly coupled with attitude maneuvers and energy consumption, creating hard constraints. Existing methods mostly focus on static coverage or deterministic tracking in a single mode, lacking a unified value measurement system to balance the conflict between robust tracking and loss search. Furthermore, traditional heuristic rules are difficult to effectively schedule multi-functional payloads to cope with the uncertain evolution of targets under strictly limited resource boundaries (such as energy and attitude stabilization time), which can easily lead to the breakdown of the tracking process.

[0004] In summary, the urgent problem to be solved is how to establish a collaborative planning method that can unify and quantify the value of search and tracking, and achieve continuous tracking of large-scale maneuvering targets under the premise of satisfying the strong constraints of satellite energy and attitude. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention proposes a constellation autonomous cooperative search and tracking method based on dynamic information gain, starting from the uncertainty evolution characteristics of moving targets at sea.

[0006] The technical solution of this invention is implemented as follows: The first aspect of this invention provides a method for autonomous cooperative search and tracking of star clusters based on dynamic information gain, comprising the following steps: S1, Obtain satellite constellation parameters and initial state information of maneuvering targets; Based on the closed-loop rolling time-domain planning framework, long-term collaborative tasks are decomposed into continuous short-term decisions, and the following steps are executed in each rolling cycle: S2, based on the current state information of the target, predict the state distribution of the target within the future planning horizon based on the motion model, so as to obtain the uncertainty of the current state of each target, and generate a set of atomic mission candidates based on the satellite constellation parameters; S3. For each target, adaptively select the tracking mode or the search mode according to the degree of uncertainty, construct the corresponding reward function, evaluate the value of each observation opportunity in the atomic task candidate set, and obtain the corresponding reward value. S4, based on a multi-objective evolutionary algorithm, performs joint optimization of the task sequence of the entire system to generate the optimal task solution; S5, based on the optimal task plan, issue the task instruction for the current time step and execute the observation, update the target state based on the observation results; return to step S2 to enter the next rolling cycle, until all planned tasks have been completed or the preset total planning time has been exceeded.

[0007] Specifically, in step S2, the target position is predicted in a gridded manner based on the Gauss-Markov motion model to obtain the existence probability of each grid center point, which constitutes the spatial probability distribution of the target. Based on this distribution, the grid within the satellite field of view is integrated or summed to obtain the hit probability of the target appearing in the field of view. Based on the orbital parameters in the satellite constellation parameters, the satellite position is deduced using an orbital prediction model. Combined with sensor field-of-view parameters and geometric field-of-view constraints, the continuous visible time window of the satellite to the target is calculated and discretized into atomic tasks with a fixed step size.

[0008] Specifically, in step S3, the degree of uncertainty is quantified by a mode indicator function, and the tracking mode or the search mode is adaptively selected by the mode indicator function: when the number of consecutive successful observations of the target is greater than or equal to a set threshold or the hit probability of the target in the current field of view is greater than a preset probability threshold, the target is determined to be in the tracking mode with uncertainty convergence; otherwise, the target is determined to have uncertainty divergence and enters the search mode.

[0009] Specifically, when the mode is determined to be tracking, the revenue function is the tracking mode value function: ; in, Indicates the target in tracking mode Execute the The observational value of each candidate task; This indicates the dynamic priority of the prediction grid where the target is located, and this priority increases non-linearly with the target loss time; Indicate target Appear in the field of view The probability of hitting within; The geometric mass coefficient representing the atomic task; This is the constant amplification factor used for the balance order; When the search mode is determined, the revenue function is the search mode value function: ; in, This indicates the comprehensive value of observations under search mode; Indicates the probability of target hit. The calculated expected information gain quantifies the degree of decrease in information entropy of the target probability distribution before and after observation; This represents the base probability reward weight.

[0010] Specifically, in step S4, the multi-objective evolutionary algorithm applies energy budget hard constraints during the optimization process, and establishes an energy recursion model based on discrete time steps and energy hard constraint conditions. satellite At any moment Available power Derived from the state at the previous moment, within any planned horizon, the satellite The total energy consumption of all assigned tasks must not exceed their current remaining power and the safety threshold. The difference between them.

[0011] Specifically, in step S4, the multi-objective evolutionary algorithm uses a constraint-aware heuristic initialization strategy to generate the initial population, specifically including: Calculate a comprehensive heuristic score for each observation opportunity; After sorting all observation opportunities in descending order of score, a Top-K strategy was used to randomly select high-value tasks, and energy and attitude pre-checks were performed before adding the tasks to the chromosome. Only tasks that pass both energy and attitude pre-checks will have their corresponding gene loci activated and injected into the initial population; other conflicting tasks will be discarded.

[0012] Specifically, in step S4, the multi-objective evolutionary algorithm uses a masked crossover operator based on satellite grouping for evolutionary iteration, specifically including: Generate a length equal to the total number of satellites binary mask vector: ; in, ; The generation rules for offspring individuals are as follows: For the The mission sequence of the satellites, if Then they will inherit completely from their parents. The mission sequence of the satellite; if Then they will inherit completely from their parents. The mission sequence of the satellite.

[0013] Specifically, step S5 also includes performing adaptive repair on the generated optimal task plan, specifically including: Dynamic redundancy control is implemented, with strict deduplication of targets in the tracking state and moderate redundancy allowed for high-risk targets in the search state. A spatiotemporal mutual exclusion mechanism is implemented to calculate the Euclidean distance between the field of view centers of candidate tasks for the same target. If the distance is less than a preset threshold, the task with the lower score is forcibly eliminated. A second energy and attitude hard constraint check was performed to ensure the physical feasibility of the repaired individual.

[0014] A second aspect of the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the computer program to implement the steps of the search and tracking method.

[0015] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the search and tracking method.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Based on the closed-loop rolling time-domain planning framework, this invention decomposes long-term collaborative tasks into continuous short-term decisions and executes a complete closed loop of "prediction-optimization-execution-feedback" in each rolling cycle. It can dynamically adapt to the high uncertainty evolution of maneuvering targets at sea and adjust the planning scheme in real time according to the latest observation results in each rolling cycle, effectively avoiding the tracking break problem caused by target maneuvering or revisit intervals. At the same time, by unifying the search and tracking values ​​into a multi-objective optimization framework and combining energy and attitude hard constraints for collaborative planning, the optimal allocation of constellation resources in the spatiotemporal dimension is realized, which significantly improves the constellation's continuous tracking capability and resource utilization efficiency for large-scale maneuvering targets. (2) This invention quantifies the uncertainty of the target state through a modal indicator function and adaptively switches between the tracking mode and the search mode. In the tracking mode, the maximum capture probability is used as the reward function, and in the search mode, the expected information gain based on KL divergence is used as the reward function. It unifies the value measurement of search and tracking, solves the technical problem of the difficulty in balancing the conflict between the two values ​​in traditional methods, and enables the system to concentrate resources for high-precision tracking when the target position is clear, and actively search for high-entropy regions when the target is lost or the uncertainty is divergent, thereby greatly improving the success rate of target recapture and avoiding the waste of resources caused by invalid observation. (3) This invention constructs an energy recursive model based on discrete time steps, fully considers the charging and consumption process of the satellite in the alternating environment of the illuminated area and the shadow area, and applies hard constraints on energy budget during the mission planning process to ensure that the total energy consumption of the mission allocated within any planning horizon does not exceed the difference between the satellite's current remaining power and the safety threshold; it ensures the physical feasibility of the planning scheme in the actual operation of the satellite. Even in the shadow area where charging is not possible, the system can still prioritize the execution of high-value missions or trigger hibernation protection, effectively avoiding mission interruption or platform failure due to energy depletion, and significantly improving the engineering robustness of the constellation collaborative planning. (4) In view of the problem that random initialization generates a large number of infeasible solutions in multi-objective evolutionary algorithms, this invention proposes a constraint-aware heuristic initialization strategy. By calculating the comprehensive heuristic score of each observation opportunity and performing Top-K random greedy sampling, energy pre-check and attitude pre-check are performed before injecting the task into the initial population, so that the initial population is naturally in the feasible solution space, which greatly reduces the generation and repair overhead of invalid solutions in the evolution process, and significantly improves the convergence speed and solution efficiency of the algorithm. (5) This invention designs a masked crossover operator based on satellite grouping, using the entire satellite's task sequence as the smallest exchange unit, and controlling the source of the task sequence inherited by the offspring individual from the two parent individuals through a binary mask vector. Compared with the traditional crossover operator, it structurally ensures that the temporal continuity and attitude constraint integrity of the tasks within a single satellite are not destroyed, avoids task conflicts or constraint violations caused by crossover operations, effectively maintains the feasibility of the population individuals while preserving the excellent genes of the parent generation, and improves the search efficiency and solution quality of the multi-objective evolutionary algorithm; (6) In the evolution process, the present invention performs an adaptive repair operator on the offspring individuals, implements strict deduplication of the tracking target through dynamic redundancy control, allows moderate redundancy for high-risk targets in the search state, eliminates redundant observations with too close distance to the center of the field of view through the spatiotemporal mutual exclusion mechanism, and performs hard constraint checks on energy and attitude again. This repair mechanism can effectively eliminate infeasible tasks and resource waste while ensuring sufficient observation information, and ensure that the final Pareto optimal task scheme has higher task value and resource utilization while satisfying all hard constraints. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a flowchart of the autonomous cooperative search and tracking method for star clusters based on dynamic information gain according to the present invention.

[0019] Figure 2 This is a schematic diagram of the closed-loop rolling time-domain planning framework in an embodiment of the present invention.

[0020] Figure 3 This is a flowchart illustrating the multi-objective evolutionary algorithm in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the global collaborative coding strategy in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the masking crossover operator based on satellite grouping in an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figure 1 The first aspect of this invention provides a method for autonomous cooperative search and tracking of star clusters based on dynamic information gain, comprising the following steps: Step S1, Initialization: Obtain satellite constellation parameters and initial state information of maneuvering targets.

[0025] This step is the starting point of the entire method, used to load all the basic data required for subsequent planning. The satellite constellation parameters include the orbital elements of each satellite, sensor field-of-view parameters, and battery capacity. Light-induced charging rate These parameters determine the satellite's observation capabilities and resource constraints. The initial state information of the maneuvering target includes the target's initial latitude and longitude position, speed, heading angle, and prior physical diffusion parameters (such as speed standard deviation, heading diffusion rate, and maximum maneuvering speed boundary), which are used to drive the predictive model for the evolution of target uncertainty. This initialization step lays the data foundation for subsequent closed-loop rolling planning.

[0026] like Figure 2 As shown in (a), this invention employs a closed-loop rolling time-domain planning framework to decompose long-term collaborative tasks into continuous short-term decisions, and executes the following steps S2-S6 within each rolling cycle. Each rolling cycle includes an alternation between "prediction planning cycle" and "execution cycle"; planning horizon length Strictly greater than the execution time step of a single task issuance This ensures that each planning step covers a sufficient window of future information while maintaining a rapid response to dynamic changes. The diagram illustrates three consecutive rolling cycles: in the first cycle, after completing the predictive planning cycle, cycle 1 is executed; subsequently, in the second cycle, based on the feedback update from the previous cycle, predictive planning cycle 2 is performed again and executed, and so on. At the observation level, observations of "ship A" show that after the initial observation (observation 1), if the target is not captured, the system enters an uncertainty accumulation phase (e.g., cycle 2), during which the probability distribution of the target's position spreads over time; after the next predictive planning cycle begins, the system re-plans based on the accumulated uncertainty, triggering a second observation (observation 2), and if it fails again, it continues to accumulate uncertainty (cycle 3). This rolling time-domain mechanism, through alternating planning and execution, achieves a rapid response to dynamic changes.

[0027] like Figure 2 As shown in Figure (b), the spatial mechanism of "prediction-execution-feedback" based on uncertainty evolution is further revealed; in the figure, the blue area represents the prediction stage in which the system extrapolates future time based on the target motion model. The graph shows the possible location area of ​​the target, which expands as the prediction time increases, reflecting the accumulation of uncertainty. The orange dashed box represents the grid target selected in the planning. The graph shows that after the first observation node fails, the accumulated uncertainty increases significantly (the blue area widens), requiring the system to expand the search range in the next planning round. After subsequent planning and execution, once the target is successfully captured, the uncertainty rapidly converges. This graph visually illustrates the temporal coupling relationship between prediction, execution, and feedback, as well as the dynamic evolution of uncertainty during the rolling process.

[0028] Step S2, Target State Prediction and Observation Opportunity Generation: Based on the preset target motion model and combined with the target's current state information, predict the state distribution of the target within the future planning horizon to obtain the uncertainty of the current state of each target, and generate a set of atomic mission candidates for subsequent decision-making based on the satellite constellation parameters.

[0029] First, the target position is predicted in a gridded manner based on the Gauss-Markov motion model, and the unnormalized probability exists at the center point of each grid. The calculation formula is as follows: ; Where d is the actual distance from the grid center to the vertex of the predicted sector. and These represent the mean and standard deviation of radial diffusion over time. The angle of deviation of the grid center relative to the predicted heading. The standard deviation of the yaw angle grows dynamically over time. This formula, based on the Gaussian distribution assumption, describes the probability density distribution of the target's position: the closer the grid is to the predicted heading and the smaller the deviation from the predicted distance, the higher the probability of the target's presence; conversely, the probability decreases exponentially. The unnormalized presence probability of the grid center point is also included. The spatial probability distribution constituting the target can be used to integrate or sum the values ​​of the grid within the satellite's field of view to obtain the hit probability of the target appearing within the field of view. This probability will serve as a core input for subsequent value assessment.

[0030] Secondly, based on the orbital parameters in the satellite constellation parameters, the SGP4 orbit prediction model is used to extrapolate the satellite position, and combined with the sensor field of view parameters and geometric field of view constraints, the continuous visible time window of the satellite to the target is calculated.

[0031] Specifically, the line-of-sight vector and observation elevation angle of the target relative to the satellite are calculated through coordinate transformation. When the elevation angle is greater than the set minimum observation elevation angle limit, it is determined that the physical visibility condition is met. The continuous visible time period is recorded as a time window and discretized into a set of mutually exclusive atomic task candidates with a fixed step size. At the same time, in order to quantify the geometric quality at different observation times, when generating the task candidate set, the geometric quality coefficient of each atomic task is calculated based on the degree to which the task execution center time deviates from the center of the visible time window. : ; in, This is the central moment of the atomic task execution. The center moment of a continuously visible time window. The total duration of the continuously visible time window. This is the geometric decay weighting factor; this coefficient reflects the degree to which the observation time deviates from the center of the time window. The smaller the deviation, the higher the geometric quality (the closer the value is to 1.0), and the larger the deviation, the more significant the quality decay. Through this step, the system obtains all candidate observation tasks and their corresponding hit probabilities and geometric quality coefficients, providing a complete decision space for subsequent value assessment.

[0032] Step S3, Adaptive Dual-Mode Collaborative Decision Making and Value Assessment: For each objective, the tracking mode or the search mode is adaptively selected according to the degree of uncertainty, and the corresponding reward function is constructed with the goal of maximizing the capture probability or maximizing the expected information gain. The value of each observation opportunity in the atomic task candidate set is assessed to obtain the corresponding reward value.

[0033] Specifically, the degree of uncertainty is indicated by a modal indication function. Quantization is performed using the modal indicator function. Adaptively select tracking mode or search mode: based on the number of consecutive successful observations of the target. ≥ Set threshold (t) or the probability of the target hitting within the current field of view When the probability exceeds a preset probability threshold, the target is determined to be in an uncertain convergence tracking mode. Otherwise, the target is determined to have experienced uncertainty divergence, and the search mode is entered. This judgment mechanism ensures that the system can focus on precise tracking when the target status is clear, and actively conduct a wide-area search when the target is lost.

[0034] When the mode is determined to be tracking, the reward function is the tracking mode value function: ; in, Indicates the target in tracking mode Execute the The observational value of each candidate task; This indicates the dynamic priority of the prediction grid where the target is located, and this priority increases non-linearly with the target loss time; Indicate target Appear in the field of view The probability of hitting within; The geometric mass coefficient representing the atomic task; This is a constant amplification factor used to balance the magnitude; the value function comprehensively considers the target importance, hit probability and observation geometry quality, so that the system prioritizes the observation opportunity with the highest capture probability in tracking mode.

[0035] When the search mode is identified, the payoff function is the search mode value function: ; in, This indicates the comprehensive value of observations under search mode; Indicates the probability of target hit. The calculated expected information gain quantifies the degree of decrease in information entropy of the target probability distribution before and after observation; Indicates the base probability reward weight; Indicates the dynamic priority of the grid; The geometric quality coefficient represents the atomic task; this value function unifies the value measurement of search and tracking, solving the technical problem of balancing the conflicting values ​​of the two in traditional methods.

[0036] Step S4, Cluster Coordination Task Planning: (e.g.) Figure 3As shown, based on a multi-objective evolutionary algorithm, with the atomic task candidate set as the decision space and the maximization of the reward value as the optimization objective, the task sequence of the entire system is jointly optimized to generate a Pareto optimal task scheme that satisfies energy and attitude constraints.

[0037] Specifically, an improved non-dominated sorting genetic algorithm (NSGA-II) is used to jointly optimize the task sequence of the entire system. The algorithm employs global cooperative coding, such as... Figure 4 As shown, chromosomes are defined as integer vectors with a length equal to the total number of observation opportunities N, and gene values ​​represent decision variables. (0 represents no observation, and non-zero integers represent selecting the corresponding time slice).

[0038] During the initialization phase, the algorithm employs a constraint-aware heuristic initialization strategy to generate the initial population, addressing the problem of numerous infeasible solutions arising from random initialization. First, it calculates the comprehensive heuristic score for each observation opportunity: ; in, Indicates candidate observation opportunities Comprehensive heuristic scoring; This indicates the dynamic priority of the target corresponding to the opportunity; This represents the expected information gain corresponding to the opportunity; and These represent the preset priority weight coefficient and information entropy weight coefficient, respectively, used to balance the tendency of tracking high-value targets and searching for high-uncertainty targets; After sorting all observation opportunities in descending order of score, a Top-K strategy was used to randomly select high-value tasks, and energy and attitude pre-checks were performed before adding the tasks to the chromosome. Energy Pre-Check: ; in, This indicates that in the process of constructing a single initial individual, the satellite Local real-time remaining power after deducting the energy consumption of allocated historical tasks; To execute the current candidate observation opportunity Estimate the required energy consumption; Attitude pre-check: ; in, Indicates the candidate task currently being validated. The planning start time; Indicates allocation to the same satellite And the preceding tasks that are immediately adjacent in time series. The execution end time; This indicates the time required for the satellite to complete a valid pointing switch, including fixed attitude maneuvers and platform stabilization. Only tasks that pass both energy and attitude pre-checks will have their corresponding gene loci activated and injected into the initial population; other conflicting tasks will be discarded. This initialization strategy ensures that the initial population is naturally within the feasible solution space, significantly reducing the generation and repair costs of invalid solutions during the evolution process.

[0039] During the optimization process, the algorithm operates in a four-dimensional target space. The algorithm performs non-dominated sorting and crowding distance calculation, with four optimization objectives: maximizing weighted task value, maximizing system information gain, minimizing system energy consumption, and minimizing soft constraint violations. Non-dominated sorting is used to divide the population into different tiers based on Pareto dominance. Tournament selection is a commonly used individual selection strategy. Its basic idea is to randomly select several individuals (usually two) from the population each time, compare their fitness or non-dominated levels, and select the best individual to enter the next generation. This process is repeated until the desired population size is reached. This strategy can effectively preserve superior genes while maintaining population diversity.

[0040] Meanwhile, the algorithm imposes hard constraints on energy budget during the optimization process, establishing an energy recursion model based on discrete time steps: ; in, Indicates satellite exist The available remaining battery power at any given time; This indicates the maximum physical capacity of the satellite's battery; Indicates satellite In the previous moment The remaining battery power; The discrete time step represents the state recursion. The illumination factor, with a value between 0 and 1, is used to characterize the proportion of time the satellite is in the illuminated area within that time step. This indicates the solar charging rate of the satellite under pure sunlight conditions.

[0041] Within any planned horizon, satellite The total energy consumption of all assigned tasks must not exceed their current remaining power and the safety threshold. The difference between them: ; in, Indicates allocation to satellites The set of all candidate observation tasks; As a binary decision variable, when the satellite Assigned to execute time slice tasks With the observation target The value is 1 if the condition is met, and 0 otherwise. Indicates execution of the first Total energy consumption required for each candidate mission (set according to whether the satellite payload is in wide field of view search mode or narrow field of view tracking mode). This represents the low-power protection threshold set to ensure the safe operation of the satellite's basic platform. This constraint ensures that the satellite can complete high-value missions or trigger hibernation protection even in the shadow area where it cannot be charged, thus guaranteeing the physical feasibility of the planning scheme during the actual operation of the satellite.

[0042] Specifically, such as Figure 5 As shown, the multi-objective evolutionary algorithm uses a masked crossover operator based on satellite grouping for evolutionary iteration to avoid the traditional crossover operator from destroying the mission time continuity and attitude constraints within a single satellite; Figure 5 In this context, "sat1", "sat2", and "sat3" represent different satellites. Specifically, generate a length equal to the total number of satellites. binary mask vector: ; in, ; The generation rules for offspring individuals are as follows: For the The mission sequence of the satellites, if Then they will inherit completely from their parents. The mission sequence of the satellite; if Then they will inherit completely from their parents. The mission sequence of this satellite. The design uses the entire satellite as the smallest unit of exchange, structurally ensuring that the integrity of the intra-satellite constraints is not compromised.

[0043] Step S5, Adaptive Repair and Constraint Handling of the Solution: Adaptive repair is performed on the generated Pareto optimal task scheme to eliminate infeasible tasks and adjust redundant observations to ensure the physical feasibility of the scheme.

[0044] First, dynamic redundancy control is implemented, and strict deduplication is carried out for targets in the tracking state. The same target is retained at most once in the same time step to avoid waste of resources. For high-risk targets in the search state, moderate redundancy is allowed, with a maximum of 3 observations retained to meet the needs of multi-angle exploration in high uncertainty areas. Secondly, a spatiotemporal mutual exclusion mechanism is executed to calculate the Euclidean distance between the field-of-view centers of the same target candidate task. If the distance is less than a preset threshold (e.g., ...), the distance is considered equal to the distance between the field-of-view centers of the target candidate task. This will force the removal of tasks with lower scores to avoid resource overlap caused by multiple satellites observing almost the same location at the same time; A second hard constraint check of energy and attitude is performed to ensure the physical feasibility of the repaired individual. This repair mechanism can effectively eliminate infeasible tasks and resource waste while ensuring sufficient observational information.

[0045] Step S6, Rolling Execution and Feedback Update: Based on the repaired task plan, issue the task instructions for the current time step and execute the observation. Update the target status based on the observation results. If all planned tasks have been completed or the preset total planning time has been exceeded, the process ends; otherwise, return to step S2 to enter the next rolling cycle.

[0046] A second aspect of the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the computer program to implement the steps of the search and tracking method.

[0047] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the search and tracking method.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for autonomous cooperative search and tracking of star clusters based on dynamic information gain, characterized in that, Includes the following steps: S1, Obtain satellite constellation parameters and initial state information of maneuvering targets; Based on the closed-loop rolling time-domain planning framework, long-term collaborative tasks are decomposed into continuous short-term decisions, and the following steps are executed in each rolling cycle: S2, based on the current state information of the target, predict the state distribution of the target within the future planning horizon based on the motion model, so as to obtain the uncertainty of the current state of each target, and generate a set of atomic mission candidates based on the satellite constellation parameters; S3. For each target, adaptively select the tracking mode or the search mode according to the degree of uncertainty, construct the corresponding reward function, evaluate the value of each observation opportunity in the atomic task candidate set, and obtain the corresponding reward value. S4, based on a multi-objective evolutionary algorithm, performs joint optimization of the task sequence of the entire system to generate the optimal task solution; S5, based on the optimal task plan, issue the task instructions for the current time step and execute the observation, and update the target state based on the observation results; Return to step S2 to enter the next rolling cycle until all planning tasks have been completed or the preset total planning time has been exceeded.

2. The method for autonomous cooperative search and tracking of star clusters based on dynamic information gain as described in claim 1, characterized in that, In step S2, the target position is predicted in a gridded manner based on the Gauss-Markov motion model to obtain the existence probability of each grid center point, which constitutes the spatial probability distribution of the target. Based on this distribution, the grid within the satellite field of view is integrated or summed to obtain the hit probability of the target appearing in the field of view. Based on the orbital parameters in the satellite constellation parameters, the satellite position is deduced using an orbital prediction model. Combined with sensor field-of-view parameters and geometric field-of-view constraints, the continuous visible time window of the satellite to the target is calculated and discretized into atomic tasks with a fixed step size.

3. The method for autonomous cooperative search and tracking of star clusters based on dynamic information gain as described in claim 1, characterized in that, In step S3, the degree of uncertainty is quantified by a mode indicator function, which adaptively selects either the tracking mode or the search mode: when the number of consecutive successful observations of the target is greater than or equal to a set threshold or the hit probability of the target in the current field of view is greater than a preset probability threshold, the target is determined to be in the tracking mode where uncertainty converges; otherwise, the target is determined to have experienced uncertainty divergence and enters the search mode.

4. The method for autonomous cooperative search and tracking of star clusters based on dynamic information gain as described in claim 3, characterized in that, When the mode is determined to be tracking, the revenue function is the tracking mode value function: ； in, Indicates the target in tracking mode Execute the The observational value of each candidate task; This indicates the dynamic priority of the prediction grid where the target is located, and this priority increases non-linearly with the target loss time; Indicate target Appear in the field of view The probability of hitting within; The geometric mass coefficient representing the atomic task; This is the constant amplification factor used for the balance order; When the search mode is determined, the revenue function is the search mode value function: ； in, This indicates the comprehensive value of observations under search mode; Indicates the probability of target hit. The calculated expected information gain quantifies the degree of decrease in information entropy of the target probability distribution before and after observation; This represents the base probability reward weight.

5. The method for autonomous cooperative search and tracking of star clusters based on dynamic information gain as described in claim 1, characterized in that, In step S4, the multi-objective evolutionary algorithm applies hard constraints on energy budget during the optimization process, and establishes an energy recursion model based on discrete time steps and hard energy constraint conditions. satellite At any moment Available power Derived from the state at the previous moment, within any planned horizon, the satellite The total energy consumption of all assigned tasks must not exceed their current remaining power and the safety threshold. The difference between them.

6. The method for autonomous cooperative search and tracking of star clusters based on dynamic information gain as described in claim 1, characterized in that, In step S4, the multi-objective evolutionary algorithm uses a constraint-aware heuristic initialization strategy to generate the initial population, specifically including: Calculate a comprehensive heuristic score for each observation opportunity; After sorting all observation opportunities in descending order of score, a Top-K strategy was used to randomly select high-value tasks, and energy and attitude pre-checks were performed before adding the tasks to the chromosome. Only tasks that pass both energy and attitude pre-checks will have their corresponding gene loci activated and injected into the initial population; other conflicting tasks will be discarded.

7. The method for autonomous cooperative search and tracking of star clusters based on dynamic information gain as described in claim 6, characterized in that, In step S4, the multi-objective evolutionary algorithm uses a masked crossover operator based on satellite grouping for evolutionary iteration, specifically including: Generate a length equal to the total number of satellites binary mask vector: ； in, ; The generation rules for offspring individuals are as follows: For the The mission sequence of the satellites, if Then they will inherit completely from their parents. The mission sequence of the satellite; if Then they will inherit completely from their parents. The mission sequence of the satellite.

8. The method for autonomous cooperative search and tracking of star clusters based on dynamic information gain as described in claim 7, characterized in that, Step S5 also includes performing adaptive repair on the generated optimal task plan, specifically including: Dynamic redundancy control is implemented, with strict deduplication of targets in the tracking state and moderate redundancy allowed for high-risk targets in the search state. A spatiotemporal mutual exclusion mechanism is implemented to calculate the Euclidean distance between the field of view centers of candidate tasks for the same target. If the distance is less than a preset threshold, the task with the lower score is forcibly eliminated. A second energy and attitude hard constraint check was performed to ensure the physical feasibility of the repaired individual.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the search and tracking method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the search and tracking method as described in any one of claims 1 to 8.

Images