Multi-type task multi-radar resource scheduling method for space object detection

Abstract

A multi-task, multi-radar resource scheduling method for space object detection is proposed. This method constructs an allocation relationship model between detection tasks and detection equipment, establishing an objective function that optimizes the overall task priority and overall equipment time utilization. A genetic algorithm is used for combinatorial optimization to find an optimal equipment allocation scheme for all search tasks. Based on the search task planning results based on the genetic algorithm, the system is updated and tasks are transformed. The remaining tracking tasks and tasks transformed from search to tracking are then processed. A joint algorithm combining deep reinforcement learning and multi-objective evolution is constructed to achieve multi-objective optimization and resource scheduling. This invention uses a two-layer planning strategy to allocate tasks of different natures to two independent planning levels, thereby reducing problem complexity and matching the most suitable optimization algorithm for each type of task.

Description

Technical Field

[0001] This invention relates to a technology in the field of space perception, specifically a multi-task, multi-radar resource scheduling method for space object detection. Background Technology

[0002] In the field of space target perception, the design of ground-based radar resource scheduling systems faces increasingly prominent challenges of complexity, the core of which lies in a series of coupled problems caused by the diversity of tasks. The system needs to simultaneously handle multiple heterogeneous tasks such as search, tracking, and mode switching, with significant differences in objectives, constraints, and performance requirements, making it difficult to establish a unified optimization model. Simultaneously, sensor resources also exhibit significant heterogeneity in functional modes, detection capabilities, and temporal availability, further increasing the complexity of matching tasks and resources. Furthermore, multiple tasks often experience resource competition and conflicts in the spatiotemporal dimensions, necessitating the construction of dynamic priority decision-making and conflict resolution mechanisms. These factors collectively lead to the scheduling problem exhibiting high-dimensional combinatorial optimization characteristics, classifying it as an NP-hard problem. Traditional methods struggle to obtain feasible scheduling solutions within a limited timeframe, thus constituting system bottlenecks in terms of real-time performance and accuracy. Summary of the Invention

[0003] This invention addresses the challenge of existing technologies in solving complex resource scheduling problems involving multiple task types, such as search, tracking, and search-to-track. It proposes a multi-task, multi-radar resource scheduling method for space object detection. By employing a two-layer planning strategy, tasks of different natures are assigned to two independent planning levels for processing, thereby reducing the complexity of the problem and constructing suitable scheduling algorithms for various tasks.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a multi-task, multi-radar resource scheduling method for space object detection. It constructs an allocation relationship model between detection tasks and detection equipment, establishing an objective function with the goal of optimizing the overall task priority and overall equipment time utilization. A genetic algorithm (GA) is used for combinatorial optimization to find an optimal equipment allocation scheme for all search tasks. Based on the search task planning results based on the GA, equipment status is updated and tasks are transformed. The remaining tracking tasks and tasks transformed from search to tracking are then processed. A deep reinforcement learning non-dominated Sorting Algorithm (DRLNS) is constructed to achieve multi-objective optimization and resource scheduling.

[0006] This invention relates to a resource scheduling system for implementing the above-mentioned method, comprising: a device-task selection unit, a search task planning unit, a device-task update unit, and a tracking task planning unit, wherein: the device-task selection unit filters and obtains search devices and search-type tasks based on an input task set and a device set; the search task planning unit performs search task planning based on a genetic algorithm based on the search devices and search-type tasks to obtain a search task allocation scheme; the device-task update unit updates the status of existing tasks and devices according to the search task allocation scheme, realizing the transformation from search to tracking tasks to tracking tasks, and obtaining a merged tracking task set; the tracking task planning unit performs tracking task planning based on DRLNS based on the tracking task set to obtain a tracking task allocation scheme.

[0007] Technical effect

[0008] This invention addresses scenarios where multiple radars need to collaborate, with some responsible for wide-area surveillance and search, and others for precise tracking of high-threat targets or immediate tracking after a search. It decouples the complex scheduling problem into two independent sub-problems, which are solved sequentially. Compared to existing technologies, this invention associates the task type and available resources at each layer, reducing the overall problem-solving difficulty. Furthermore, each layer's algorithm can be configured individually, allowing for the construction of suitable algorithms based on different scheduling scenarios. For example, the first layer is a constrained combinatorial allocation problem, suitable for genetic algorithms; the second layer involves more complex dynamic decision-making and multi-objective trade-offs, effectively solved by constructing a joint algorithm combining deep reinforcement learning and multi-objective evolution. Attached Figure Description

[0009] Figure 1 This is a flowchart of the present invention;

[0010] Figure 2 This is a flowchart of an example implementation. Detailed Implementation

[0011] like Figure 1 and Figure 2 As shown in this embodiment, a multi-task, multi-radar resource scheduling method for space object detection is provided. Specifically, it addresses resource scheduling scenarios involving tracking radars capable only of tracking tasks, search radars capable only of search tasks, and integrated radars possessing both search and tracking capabilities.

[0012] Step 1: Construct a model of the allocation relationship between detection tasks and detection equipment. Specifically, the task set Tasks = {Task[1], ..., Task[i], ..., Task[m]} and the equipment set Devices = {Radar[1], ..., Radar[j], ..., Radar[n]}, where: each task Task[i] in the task set (Tasks) contains task attribute information, task_type[i] is the task type, Task[i] = {starttime[i], endtime[i], priority[i], maxtrackdist[i], maxsearchdist[i], tasktype[i]}, and the radar equipment Radar[j] in the equipment set (Devices) contains equipment attribute information, Radar[j] = {schedule[j], devicetype[j], subdist[j], maxtrackdist[j], searchdist[j], tasktype[j]}.

[0013] The task attribute information includes: starttime[i] is the task start time, endtime[i] is the task end time, priority[i] is the task priority, maxtrackdist[i] is the maximum tracking distance requirement, and maxsearchdist[i] is the maximum search distance requirement.

[0014] The device attribute information includes: schedule[j] is the available time window of the device, devicetype[j] is the device type, subdist[j] is the tracking distance of the sub-beam under multi-target device (only available when devicetype is phased array), maxtrackdist[j] is the maximum tracking distance of single-target device (only available when devicetype is single pulse), searchdist[j] is the maximum search distance of the device, and tasktype[j] is the type of task that the device can perform.

[0015] The task types mentioned include: {1: Tracking task, 2: Search task, 3: Search-to-tracking task}.

[0016] The device types mentioned include: {monopulsive: single-target device, phased array: multi-target device}.

[0017] The types of tasks that the device can perform include: {1: tracking type, 2: search type, 3: search to tracking type}.

[0018] Step 2, Objective Function Modeling: Prioritize the overall priority of the detection tasks and the total time utilization of the detection equipment as optimization objectives. Under the premise of satisfying all constraints, maximize the optimal balance between these two indicators: overall task priority and total equipment time utilization. Specifically: , These are weighting coefficients used to balance the importance of two objectives. . The sum of priorities of all successfully scheduled tasks, where i represents the i-th task. The sum of the proportions of available time spent by all devices performing tasks, where j represents the j-th device.

[0019] Step 3, the first layer of the two-layer planning strategy, namely, the search task planning based on genetic algorithms: Genetic algorithms (GA) are used for combinatorial optimization to find an optimal device allocation scheme for all search tasks, maximizing the overall priority of successfully scheduled tasks. Specifically, this includes:

[0020] 3.1 Initialization: Filtering and preprocessing, specifically including: building a task pool, filtering all search tasks from the initial task list; and building a device pool, filtering all search devices from the initial device information.

[0021] 3.2 Constructing Chromosome Encoding for Genetic Algorithms. A solution (i.e., a complete assignment scheme) is encoded as an integer array, i.e., a chromosome.

[0022] The length of the chromosome is equal to the number of search tasks. The value of each gene position on the chromosome represents the device index assigned to that task.

[0023] For example, chromosome[i] = j means assigning the i-th search task to the j-th searchable device. When a task is not assigned, its corresponding gene value can be set to a special value, such as -1.

[0024] 3.3 Construct the fitness function of the genetic algorithm to evaluate the quality of each chromosome (each solution). Specifically, for each gene chromosome[i] = j in a chromosome, a validity check is performed. Then, for chromosomes that have passed the validity check, the fitness is calculated by traversing the entire chromosome. ,in: For the fitness calculation function, Let be the priority of the i-th task. When a task assignment is valid, its priority is added to the total score. Invalid task assignments contribute a priority of 0.

[0025] The validity checks include:

[0026] a) Equipment capability verification: Check whether the search distance of equipment j is greater than or equal to the search distance requirement of task i. For multi-target equipment, the search distances of all its sub-beams need to be checked; as long as the search distance of any one sub-beam meets the requirement, it is acceptable.

[0027] b) Time window verification: Check whether the time interval [starttime[i], endtime[i]] of task i falls entirely within the available time window of device j. Since a device may be assigned multiple tasks, it is necessary to check whether task i conflicts with other tasks already assigned to device j in terms of time.

[0028] 3.4 Constructing the selection, crossover, and mutation operations of the genetic algorithm, i.e., genetic operations.

[0029] The selection process employs a tournament selection method. K individuals are randomly selected from the population, with the individual with the highest fitness being chosen for the next generation. This method balances selection pressure and population diversity.

[0030] The crossover described herein is a single-point crossover. Two parent chromosomes are randomly selected, and a crossover point is randomly chosen. The gene sequences after the crossover point are exchanged, generating two new offspring. This facilitates the combination of fragments from different optimal schemes.

[0031] The mutation is achieved by randomly selecting a gene locus i and modifying its value to another random and valid device index (or -1). Mutation is key to escaping local optima and maintaining population diversity.

[0032] 3.5 Implement the genetic algorithm iteration, specifically: calculate the fitness of each individual in the population, repeat the genetic operation to generate a new population, and calculate the corresponding fitness until the chromosome with the highest fitness is found, which is the best allocation scheme for the search task set.

[0033] The genetic operations mentioned above refer to: selection operations to select superior individuals, and crossover and mutation operations to generate new offspring populations.

[0034] Step 4: Update the system and transform the task based on the search task planning results of the genetic algorithm, specifically including:

[0035] 4.1 Update the task list and remove search tasks that have been successfully assigned to the first level from the global task list.

[0036] 4.2 Update device information. For devices performing search tasks, their available time windows need to be segmented and updated based on the time of the assigned tasks. Remove the search device from the device pool for subsequent planning.

[0037] 4.3 Convert search to tracking tasks: Filter out all search-to-tracking tasks and convert them into standard tracking tasks. Specifically, subtract a preset search preparation time ΔT (set to 10 minutes in this invention) from the task's start time. All other parameters (end time, priority, distance requirement) remain unchanged.

[0038] Step 5: The second layer of the two-layer planning strategy, namely, tracking task planning based on DRLNS: This involves processing the remaining tracking tasks and tasks converted from search to tracking, constructing the multi-objective optimization and resource scheduling mechanism of this invention, specifically including:

[0039] 5.1 Merge the original tracking tasks and the transformed search-to-tracking tasks. Combine all tracking tasks in the initial task list with the new tracking tasks transformed in the previous step to form the second-level set of tasks to be planned.

[0040] 5.2 Task scheduling is implemented using the DRLNS method, specifically including:

[0041] 5.2.1) Initialization phase, specifically including:

[0042] a. Constructing and initializing the D3QN network, randomly initializing the weight parameters of the online and target D3QN networks. Clear the experience replay pool. .

[0043] b. NSGA-III Population Initialization: This is achieved by having the D3QN agent execute a process guided by a combination of random and heuristic policies. A complete episode (i.e., generated) (Initial scheduling scheme). The scheduling scheme generated at the end of each episode and its two corresponding target values. Joined as an individual in the initial population of NSGA-III middle.

[0044] c. NSGA-III Reference Point Generation: Based on the number of targets and preset partitioning parameters, generate a set of uniformly distributed reference points in the target space. .

[0045] 5.2.2) The training of D3QN and the population co-evolution cycle of NSGA-III (main training cycle) specifically include:

[0046] a. D3QN decision-making and solution generation, specifically including:

[0047] i) Calculate a weight vector from the Pareto front currently maintained by NSGA-III, based on the elite selection strategy (selecting solutions from a specific region on the front). This weight vector is used to scalarize the multi-objective Q-value function of D3QN to guide the agent's decision preferences in this episode. In this invention, the scalarized reward is... .

[0048] ii) Obtain the initial environment state from the scene generator The D3QN agent begins sequential decision-making (making a series of task assignment decisions in one round), that is, at each time step... Inside, the agent observes the current state. Calculate all possible actions using a D3QN network. To determine the Q-value, and considering multi-objective optimization, this invention uses weights... This guides the calculation of the Q-value and helps obtain the optimal action.

[0049] iii) According to -The greedy strategy clearly defines the next action. (assign tasks) Give equipment And in time (Execute, or not assign a task). The environment performs actions. Perform constraint checks and return a new state. and rewards (Includes multi-objective reward components). [The following is a list of components:] Stored in the experience replay pool In the middle. After meeting the minimum training sample requirement, periodically (every... (Step 1) Randomly sample a mini-batch of data from the experience replay pool to train the D3QN network (update online network parameters). Periodically (every) Step 1) Copy the parameters from the online network to the target network. .

[0050] iv) When all tasks have been processed, a round ends. Record the complete scheduling scheme generated in that round and its final multi-objective evaluation value. .

[0051] b. Assessment and population renewal, specifically including:

[0052] i) Combine the scheduling scheme generated by the current episode with... Joining the current NSGA-III population as a new individual middle.

[0053] ii) When the population size reaches the preset iteration condition (per generation) After implementing the D3QN scheme, the NSGA-III evolutionary operation is run, merging the parent population and the newly generated offspring individuals. The merged population is then subjected to a non-dominated sorting and divided into... In non-dominant hierarchies, starting from the highest priority level (first level), individuals are selected layer by layer to enter the next generation of the population. until the population size reaches Or it just happens to contain a certain level Part of it. When the last layer If truncation is required, it will be achieved through NSGA-III's unique reference point association and niche preservation mechanism. The remaining individuals are precisely selected to ensure the convergence and diversity of the population.

[0054] iii) NSGA-III is based on the updated Pareto frontier (New non-dominated solution set) Recalculate the weight vector This will provide higher-quality information guidance for the next round of D3QN agent decision-making.

[0055] 5.2.3) The termination condition is: when the preset maximum number of iterations (training epochs) is reached, the process terminates and outputs the Pareto optimal solution set maintained in the NSGA-III population, the optimal weight vector of the multi-optimization objective, and the trained D3QN agent.

[0056] Step 6: Load the D3QN model parameters and multi-target weight coefficients trained in Step 5 into the inference model with the same network structure. Input the space object detection task to be planned and the detection equipment information into the inference model. The inference model will complete the task allocation of all tracking tasks and search-to-tracking tasks, and output the task plan of each detection device according to the time sequence.

[0057] Based on real-world application scenario experiments, the hardware platform consisted of an Intel Core i5-13500H processor (2.6GHz) and 16GB of RAM. The software environment used a 64-bit Windows 11 system built on a Python 3.9 platform, relying on the PyTorch 2.4 deep learning framework. The experimental dataset was dynamically constructed using a custom parameterized generator, and independent test samples were generated using the Monte Carlo method. Specifically, 500 sets of data were generated for training, and 100 sets were used for testing. To simulate the dynamic characteristics of a real detection environment, each set of data maintained differences in dimensions such as device type distribution, task time window overlap, and resource demand conflict intensity.

[0058] The comparative methods used in the experiment mainly fall into three categories: intelligent methods, heuristic optimization methods, and traditional modeling and solution methods. Among them, intelligent methods include task planning models based on D3QN reinforcement learning, heuristic optimization methods (GA) use task planning models based on genetic algorithms, and traditional modeling and solution methods use task planning models based on integer linear programming (ILP). Three evaluation metrics are used: overall priority, task success rate, and method execution time.

[0059] 1) Overall Priority Score: Overall Task Priority , where: Pri i p represents the priority of task i. i This indicates whether the i-th task was successful (value is 0 / 1).

[0060] The total priority score only accumulates the priority of successful tasks; failure of high-priority tasks will have a significant negative impact on the total score. This metric is value-oriented, quantifying the system's ability to guarantee high-value tasks, and sacrificing low-priority tasks to ensure high-priority tasks when resources are scarce.

[0061] 2) Equipment Time Utilization: When the equipment is a single-pulse device, the equipment time utilization is calculated as follows: , where: Ocu j For the j-th task to be executed on device i, N i The total number of tasks assigned to device i, Ava i This represents the entire available time interval for device i.

[0062] If the equipment is a phased array device, the specific calculation of equipment time utilization is as follows: Where: Mer(Ocu1, Ocu2,…, Ocu) Ni ) represents the occupied time after combining all sub-beams of device i, Ava i Let be the total available time interval for device i.

[0063] For example, if the available time for a single-pulse device is 100 minutes, and all assigned tasks occupy a total of 80 minutes, then the utilization rate is 80%. If the available time for a phased array device is 100 minutes, and sub-beam 1 occupies a time slot of 0-50 minutes, while sub-beam 2 occupies a time slot of 40-60 minutes, then after time combining, the occupied time slot is 0-60 minutes, resulting in a utilization rate of 60%.

[0064] The average time utilization rate of the equipment is: , where: N dev P represents the total number of devices. utli P represents the time utilization rate of device i. utl This represents the average time utilization rate of the equipment.

[0065] 3) Execution time: The execution time is: ,in: , where: P timei Ts represents the total time allocated to the algorithm from environment initialization to completion of task i. i Te i P is the time when the algorithm completes its planning. time This metric measures the time consumed by all planned tasks. It measures the real-time performance of the algorithm; in dynamic task scenarios, the algorithm needs to respond within seconds.

[0066] Table 1 Comparison of performance methods

[0067] To evaluate the performance of this invention in radar space object mission planning, five mission scheduling methods were evaluated on 100 sets of test data based on the above three evaluation indicators. According to the results in Table 1, among intelligent algorithms, the DRLNS method significantly outperforms DQN and D3QN in terms of total priority (19098) and average time utilization (35.374%), despite a slightly longer execution time (0.305 seconds), resulting in a significant performance improvement. Compared to the heuristic method GA, the DRLNS method significantly leads in all indicators, especially in execution time (0.305 seconds vs. 19.789 seconds). Compared to the traditional method ILP, although ILP has higher total priority (25768) and equipment time utilization (37.189%), and can be approximated as a standard solution, its execution speed is extremely slow (0.305 seconds vs. 1112.029 seconds). This demonstrates that the DRLNS method successfully integrates the advantages of deep reinforcement learning and evolutionary algorithms, exhibiting superior performance across multiple key performance dimensions, reflecting the method's efficiency and overall performance improvement.

[0068] Taking the 100th group of tracking task scheduling test samples as an example, the task planning of this invention is illustrated. Known information is shown in Tables 2 and 3.

[0069] Table 2 Test Group 100 - Radar Information

[0070] Table 3. Task Information for Test Group 100

[0071] This planning task involves three sets of detection equipment and six detection tasks. Equipment 1 has three sub-beams, equipment 2 has a single main beam, and equipment 3 has two sub-beams. After algorithmic selection, five tasks—1, 2, 4, 5, and 6—are executed. Based on the scheduling output, equipment 1 uses sub-beam 1 to execute task 1, sub-beams 2 and 3 work together to execute task 2, equipment 2 executes task 4 alone, and equipment 3 uses sub-beams 1 and 2 to execute tasks 5 and 6 respectively. Task 3 was excluded after comprehensive consideration because it conflicted with tasks 5 and 6 in terms of time and had a lower priority than the sum of the priorities of tasks 5 and 6. This demonstrates that the tracking task planning model can reasonably avoid conflicting tasks, optimize resource allocation, and ensure that tasks remain within the coverage area of ​​the equipment.

[0072] Compared with existing technologies, this invention employs a two-layer planning strategy in a resource scheduling architecture for multi-task applications, decoupling various mixed tasks and reducing problem complexity. Simultaneously, it enables flexible configuration of advantageous models for decoupled single-type tasks. Secondly, for tracking tasks, the proposed radar task planning method based on DRLNS establishes an online co-evolutionary mechanism of deep reinforcement learning and multi-objective optimization algorithms. It uses the D3QN model to make online decisions on task schemes and periodically adjusts the planning strategy. By exploring the Pareto front in sequential decision-making using NSGA-III, it ensures global optimization of the solution. The combination of these two approaches achieves an effective trade-off between objectives such as overall task priority and time utilization.

[0073] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A multi-task, multi-radar resource scheduling method for space object detection, characterized in that, By constructing an allocation relationship model between detection tasks and detection equipment, an objective function is established with the goal of optimizing the overall priority of tasks and the overall utilization rate of equipment time. A genetic algorithm is used for combinatorial optimization to find an optimal equipment allocation scheme for all search tasks. The system is then updated and tasks are transformed based on the search task planning results based on the genetic algorithm. The remaining tracking tasks and tasks transformed from search to tracking are then processed. A joint algorithm of deep reinforcement learning and multi-objective evolution is constructed to achieve multi-objective optimization and resource scheduling.

2. The multi-task, multi-radar resource scheduling method for space object detection according to claim 1, characterized in that, include: Step 1: Construct a model of the allocation relationship between detection tasks and detection equipment, specifically: Task set Tasks={Task[1],….Task[i],….,Task[m]}, Equipment set Devices={Radar[1],….,Radar[j],….,Radar[n]}, where: Each task Task[i] in the task set (Tasks) contains task attribute information, task_type[i] is the task type, Task[i]={starttime[i],endtime[i],priority[i],maxtrackdist[i],maxsearchdist[i],tasktype[i]}, and the radar equipment Radar[j] in the equipment set (Devices) contains equipment attribute information, Radar[j]={schedule[j],devicetype[j],subdist[j],maxtrackdist[j],searchdist[j],tasktype[j]}; Step 2, Objective Function Modeling: Prioritize the overall priority of the detection tasks and the total time utilization of the detection equipment as optimization objectives. Under the premise of satisfying all constraints, maximize the optimal balance between these two indicators: overall task priority and total equipment time utilization. Specifically: These are weighting coefficients used to balance the importance of two objectives. The sum of priorities for all successfully scheduled tasks, where i represents the i-th task. The sum of the proportions of available time spent by all devices performing tasks, where j represents the j-th device; Step 3: The first layer of the two-layer planning strategy, namely the search task planning based on genetic algorithm: Use genetic algorithm (GA) for combinatorial optimization to find an optimal device allocation scheme for all search tasks, so that the total priority of successfully scheduled tasks is the highest; Step 4: Update the system and transform the task based on the search task planning results of the genetic algorithm, specifically including: 4.1 Update the task list and remove search tasks that have been successfully assigned to the first level from the global task list; 4.2 Update device information. For devices performing search tasks, their available time windows need to be segmented and updated according to the time of the assigned tasks, and the search devices should be removed from the device pool for subsequent planning. 4.3 Convert search to tracking tasks: Filter out all search to tracking tasks and convert them into standard tracking tasks; Step 5: The second layer of the two-layer planning strategy, namely the tracking task planning based on DRLNS: handle the remaining tracking tasks and the tasks transformed from search to tracking, and construct the multi-objective optimization and resource scheduling of this invention. Step 6: Load the D3QN model parameters and multi-target weight coefficients trained in Step 5 into the inference model with the same network structure. Input the space object detection task to be planned and the detection equipment information into the inference model. The inference model will complete the task allocation of all tracking tasks and search-to-tracking tasks, and output the task plan of each detection device according to the time sequence.

3. The multi-task, multi-radar resource scheduling method for space object detection according to claim 2, characterized in that, Step 3 specifically includes: 3.1 Initialization: Filtering and preprocessing, specifically including: building a task pool, filtering all search tasks from the initial task list; and building a device pool, filtering all search devices from the initial device information. 3.2 Constructing the chromosome encoding for the genetic algorithm: Encode a complete allocation scheme into an integer array, i.e., chromosomes; The length of the chromosome is equal to the number of search tasks, and the value of each gene position on the chromosome represents the device index assigned to that task; 3.3 Construct the fitness function of the genetic algorithm to evaluate the quality of each chromosome. Specifically, for each gene chromosome[i]=j in a chromosome, a validity check is performed. Then, for the chromosomes that have passed the validity check, the fitness is calculated by traversing the entire chromosome. ,in: For the fitness calculation function, Let be the priority of the i-th task. When a task is assigned effectively, its priority is added to the total score. Tasks with invalid assignments contribute 0 priority. 3.4 Constructing the selection, crossover, and mutation operations of a genetic algorithm, i.e., genetic operations; The selection method described uses tournament selection, randomly selecting k individuals from the population, where the individual with the highest fitness is selected to enter the next generation. This method balances selection pressure and population diversity. The aforementioned crossover uses a single-point crossover method, randomly selecting two parent chromosomes, randomly selecting a crossover point, exchanging the gene sequences after the crossover point, and generating two new offspring. This helps to combine fragments of different excellent schemes. The aforementioned mutation, which involves randomly selecting a gene locus i and modifying its value to another random and valid device index, is key to escaping local optima and maintaining population diversity. 3.5 Implement the genetic algorithm iteration, specifically: calculate the fitness of each individual in the population, repeat the genetic operation to generate a new population, and calculate the corresponding fitness until the chromosome with the highest fitness is found, which is the best allocation scheme for the search task set. The genetic operations mentioned above refer to: selection operations to select superior individuals, and crossover and mutation operations to generate new offspring populations.

4. The multi-task, multi-radar resource scheduling method for space object detection according to claim 2, characterized in that, Step 5 specifically includes: 5.1 Merge the original tracking tasks and the converted search-to-tracking tasks, and summarize all the tracking tasks in the initial task list with the new tracking tasks converted from the previous step to form the second-level set of tasks to be planned; 5.2 Task scheduling is implemented using the DRLNS method, specifically including: 5.2.1) Initialization phase; 5.2.2) Training of D3QN and the population co-evolution cycle of NSGA-III.

5. The multi-task, multi-radar resource scheduling method for space object detection according to claim 4, characterized in that, The initialization phase specifically includes: a. D3QN Network Construction and Initialization: Randomly initializing the weight parameters of the online and target D3QN networks. Clear the experience replay pool. ; b. NSGA-III Population Initialization: This is achieved by having the D3QN agent execute a population initialization process guided by a combination of random and heuristic policies. A complete episode, i.e., generating An initial scheduling scheme is generated at the end of each episode, along with the corresponding two target values. Joined as an individual in the initial population of NSGA-III middle; c. NSGA-III Reference Point Generation: Based on the number of targets and preset partitioning parameters, generate a set of uniformly distributed reference points in the target space. .

6. The multi-task, multi-radar resource scheduling method for space object detection according to claim 4, characterized in that, The training of D3QN and the population co-evolution cycle of NSGA-III specifically include: a. D3QN decision-making and solution generation, specifically including: i) From the Pareto front currently maintained by NSGA-III, calculate a weight vector according to the elite selection strategy, i.e., select a solution in a specific region on the front. This weight vector is used to scalarize the multi-objective Q-value function of D3QN to guide the agent's decision preferences in this episode. The scalarized reward is... ; ii) Obtain the initial environment state from the scene generator The D3QN agent begins to make sequential decisions, that is, at each time step... Inside, the agent observes the current state. Calculate all possible actions using a D3QN network. The Q value, through weights This guides the calculation of the Q-value and helps determine the optimal action. iii) According to -The greedy strategy clearly defines the next action. That is, assigning tasks Give equipment And in time The environment performs actions, either by assigning tasks or not. Perform constraint checks and return a new state. and rewards ,Will Stored in the experience replay pool In this process, after meeting the minimum training sample requirement, a mini-batch of data is periodically sampled randomly from the empirical replay pool to train the network parameters of the D3QN network. The parameters of the online network are periodically copied to the target network. ; iv) When all tasks have been processed, a round ends. Record the complete scheduling scheme generated in that round and its final multi-objective evaluation value. ; b. Assessment and population renewal, specifically including: i) Combine the scheduling scheme generated by the current episode with... Joining the current NSGA-III population as a new individual middle; ii) When the population size reaches the preset iteration condition, run the NSGA-III evolutionary operation to merge the parent population and the newly generated offspring individuals. Perform a non-dominated sort on the merged population and divide it into... In non-dominant hierarchies, individuals are selected layer by layer from the highest priority to enter the next generation of the population. until the population size reaches Or it just happens to contain a certain level Part of, when the last layer If truncation is required, it will be achieved through NSGA-III's unique reference point association and niche preservation mechanism. The remaining individuals are precisely selected to ensure the convergence and diversity of the population; iii) NSGA-III is based on the updated Pareto frontier Recalculate the weight vector This provides higher-quality information guidance for the next round of D3QN agent decision-making; 5.2.3) The termination condition is: when the preset maximum number of iterations is reached, the process terminates and outputs the Pareto optimal solution set maintained in the NSGA-III population, the optimal weight vector of the multi-optimization objective, and the trained D3QN agent.

7. A multi-task, multi-radar resource scheduling system for space object detection that implements the method of any one of claims 1-6, characterized in that, include: The system comprises a device-task selection unit, a search task planning unit, a device-task update unit, and a tracking task planning unit. Specifically: the device-task selection unit filters and obtains search devices and search-type tasks based on the input task set and device set; the search task planning unit plans search tasks based on the search devices and search-type tasks using a genetic algorithm to obtain a search task allocation scheme; the device-task update unit updates the status of existing tasks and devices according to the search task allocation scheme, transforming search tasks into tracking tasks and obtaining a merged tracking task set; and the tracking task planning unit plans tracking tasks based on DRLNS according to the tracking task set to obtain a tracking task allocation scheme.

Images