Heterogeneous multi-uav task allocation method based on dual-population co-evolution

Abstract

This invention relates to the field of path planning technology, specifically to a heterogeneous multi-UAV task allocation method based on dual-population co-evolution, comprising: constructing an initial population; calculating the fitness of each individual in the initial population based on their position, determining the quality of each individual, and extracting the non-dominated solution set from the initial population to initially construct and activate the restricted external archive, thereby establishing a global elite non-dominated guidance mechanism. The initial population is divided into an exploration subpopulation and a mining subpopulation; the positions of each individual in the exploration and mining subpopulations are updated respectively, the fitness of each individual after the position update is calculated, the quality of each individual after the position update is determined, an archive maintenance strategy based on congestion distance is adopted to maintain the diversity of non-dominated solutions in the restricted external archive, and the archive set is updated until the iteration termination condition is met. The resulting archive set is the optimal solution set of the multi-UAV cooperative task allocation scheme.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, specifically to a heterogeneous multi-UAV task allocation method based on dual-population co-evolution. Background Technology

[0002] Multi-UAV systems can efficiently complete tasks such as disaster reconnaissance, survivor search and rescue, and emergency supplies delivery through parallel and collaborative operations, thereby significantly improving rescue efficiency and regional coverage. Therefore, they are widely used in post-disaster emergency rescue. In this context, multi-UAV task allocation has become a key issue in achieving swarm collaborative operations. Its goal is to generate reasonable task execution plans for heterogeneous UAV swarms while meeting flight performance, mission requirements, and collaborative constraints. Therefore, the multi-UAV task allocation problem is essentially an NP-hard combinatorial optimization problem.

[0003] For the multi-UAV task allocation problem, various solution methods have been proposed in the existing technology. Among them, metaheuristic algorithms have become an important method for solving this type of NP-hard combinatorial optimization problem due to their strong modeling adaptability and global search capability. Metaheuristic methods such as ant colony optimization, pigeon colony optimization, and gray wolf optimization are widely used in multi-UAV cooperative optimization problems.

[0004] However, ant colony optimization, pigeon flock optimization, and gray wolf optimization are all single-swarm heuristic intelligent algorithms. In high-dimensional discrete search spaces, these algorithms are still prone to imbalances between global exploration and local exploitation, which affects the generation of heterogeneous UAV swarm mission plans and leads to reduced rescue efficiency. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide a heterogeneous multi-UAV task allocation method based on dual-population co-evolution, and the specific technical solution adopted is as follows: This invention provides a heterogeneous multi-UAV task allocation method based on dual-population co-evolution, comprising the following steps: (1) Based on the parameters of the task point to be rescued and the parameters of the heterogeneous UAV swarm, construct the mathematical model of the task point to be rescued and the mathematical model of the heterogeneous UAV swarm, and set the constraints and objective functions to obtain the multi-UAV task allocation optimization model; (2) The dual-population co-evolutionary algorithm is used to iteratively optimize the constructed multi-UAV task allocation model to obtain the optimal solution set of the multi-UAV cooperative task allocation scheme; specifically including: (2.1) Construct the initial population; (2.2) Based on the position of each individual in the initial population, calculate the fitness of each individual in the initial population, determine the superiority or inferiority of each individual, and then divide the initial population into an exploration subpopulation and a mining subpopulation; and construct an archive set; (2.3) Update the position of each individual in the exploration subpopulation and the mining subpopulation respectively, calculate the fitness of each individual after the position update, determine the quality of each individual after the position update, and then divide all individuals after the position update into the exploration subpopulation and the mining subpopulation, and update the archive set; (2.4) Iterate through step (2.3) until the iteration termination condition is met. The resulting file set is the optimal solution set for the multi-UAV collaborative task allocation scheme.

[0006] In some possible implementations, the method for updating the position of each individual in the mining subpopulation includes: updating the position of each individual in the mining subpopulation using adaptive differential search and greedy exchange mutation.

[0007] In some possible implementations, methods for updating the position of individuals in the exploratory subpopulation include: To update the position of each individual in the exploratory subpopulation, the method involves randomly performing long-distance jumps based on chaotic sequences or approximating the global optimum along a logarithmic spiral path.

[0008] In some possible implementations, the objective function is a bi-objective function, which includes minimizing the total time to complete all tasks and maximizing the overall rescue benefits.

[0009] In some possible implementations, the fitness is a bi-objective fitness, and the method for determining the superiority or inferiority of individuals includes the Pareto dominance comparison method.

[0010] In some possible implementations, the method of constructing the archive set includes: using the Pareto dominance comparison determination method to identify non-dominant individuals in the initial population, thereby constructing the archive set.

[0011] In some possible implementations, the method for updating the archive set includes: based on the fitness of each individual after the position update and the fitness of each individual in the archive set, using the Pareto dominance comparison method to compare the quality of each individual with that of each individual in the archive set, and for any individual that is not dominated by any individual in the archive set, then adding that individual to the archive set.

[0012] In some possible implementations, methods for constructing the initial population include: An encoding and decoding mechanism is constructed to encode the multi-UAV task allocation scheme into individual position vectors. The encoding structure of the individual position vectors is determined, and then the position vectors of each individual in the initial population are generated by using a constraint-based processing mechanism to obtain the initial population.

[0013] In some possible implementations, the methods for calculating the total time to complete all tasks include: For any drone, sum the time it takes to reach the last mission point, the time it spends at the mission point, and the time it takes to fly back to the base from the mission point to get the total mission completion time for that drone. The total completion time of all tasks is obtained by taking the maximum value from the total completion time of all drones.

[0014] In some possible implementations, the methods for calculating the overall relief benefits include: Substitute the completion time of each task into the time decay function; By combining the effective service volume, task completion rate, and time decay function calculation results of each task point, the rescue benefits of each task are obtained. The total rescue benefit is obtained by summing the rescue benefits from all missions.

[0015] The present invention has the following beneficial effects: The present invention utilizes a dual-population co-evolutionary algorithm to iteratively optimize the constructed multi-UAV task allocation model. This algorithm divides the population into an exploration subpopulation and a mining subpopulation. The two subpopulations update individual positions in different ways, which expands the algorithm's search space and can effectively avoid the optimization model from getting stuck in local optima. At the same time, it can accelerate the convergence speed. Furthermore, the two exploration subpopulations share a common archive set, and co-evolution is achieved under the constraint of the common archive set, which can ensure that individuals continuously approach the global optimum, thereby improving the quality of the multi-UAV task allocation scheme. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a rescue scenario according to an embodiment of the present invention; Figure 2 This is a vector synthesis relationship diagram of an embodiment of the present invention; Figure 3 This is a flowchart of the dual-population cooperative evolution algorithm according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the encoding method according to an embodiment of the present invention; Figures 5-8 These are comparison diagrams of task point distribution in four scenarios according to embodiments of the present invention; Figures 9-12 These are box plot comparison diagrams of various algorithms in four scenarios of the present invention; Figures 13-16These are comparison charts showing the distribution of solution sets obtained in four scenarios according to embodiments of the present invention. Detailed Implementation

[0017] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0018] The heterogeneous multi-UAV task allocation method based on dual-population co-evolution of the present invention includes the following steps: Step 1: Based on the parameters of the task points to be rescued and the parameters of the heterogeneous UAV swarm, construct a mathematical model for the task completion time and comprehensive rescue benefits, and set constraints based on the actual rescue conditions to obtain a multi-UAV task allocation optimization model with the goal of minimizing the total time of all task completion and maximizing the comprehensive rescue benefits (it is also called a dual-objective optimization model because it has two optimization objectives).

[0019] (1.1) Based on the parameters of the rescue mission point and the parameters of the heterogeneous UAV swarm, construct a mathematical model of the rescue mission point and the heterogeneous UAV swarm.

[0020] In this embodiment, the rescue scenario diagram is as follows: Figure 1 As shown, there are several distributed among them ( (Number) pending rescue mission points, marked as Receive the collection of rescue mission points The parameters for each rescue mission point are determined, including: rescue mission type, geographical location, basic rescue value, and time window requirements. In this embodiment, the rescue mission types include: disaster reconnaissance, target confirmation, and emergency supplies delivery, which are used to determine the capabilities required to rescue the mission point. Among them, the disaster reconnaissance mission requires the UAV to have the ability to conduct disaster surveys and search for signs of life; the target confirmation mission requires the UAV to have the ability to accurately identify trapped personnel and carry a small amount of first aid supplies; and the emergency supplies delivery mission requires the UAV to have the ability to carry relatively heavy rescue supplies.

[0021] In this embodiment, the heterogeneous drone swarm is based on Taking a single drone as an example, each drone in a heterogeneous drone swarm is labeled as... , obtain drone collection The type of each drone is determined, and then the performance parameters of each drone are determined, including flight speed, endurance, payload, and functions.

[0022] In this embodiment, UAVs are categorized into reconnaissance, operational, and heavy-load types. Performance parameters for each type of UAV are shown in Table 1.

[0023] Table 1. Performance parameters of various types of UAVs To ensure the effectiveness of the rescue area model and the drone swarm model, and to make the models closely resemble actual rescue scenarios, the following assumptions are made in this implementation: (1) During the flight phase, the UAV maintains independent flight path planning; during the operation phase, when the capability of a single UAV is insufficient to meet the task requirements... When needed, multiple heterogeneous UAVs can work together. (2) All UAVs take off from the same rescue command center (fixed base) and must return to the same rescue command center for resupply after completing the assigned task sequence. During the entire mission cycle, mechanical failure or mid-journey crash of UAVs is not considered. (3) The low-altitude environment after the disaster is subject to time-varying wind field interference, and the actual ground speed of the UAV is determined by its airspeed vector and wind speed vector. (4) Considering the evolution characteristics of the disaster, the mission benefit is modeled as the number of intervals that decay over time. That is, the mission not only has an optimal rescue time window, but its basic benefit value also has a fluctuation range to reflect the uncertainty of environmental information.

[0024] (1.2) Construct a mathematical model for mission completion time and rescue benefits.

[0025] In this implementation, a dual-objective optimization model is constructed with the objective functions of minimizing the total time to complete all tasks and maximizing the rescue benefits. Therefore, this implementation requires the construction of mathematical models for task completion time and rescue benefits to quantify these two optimization objectives and to calculate fitness during subsequent iterative optimization processes.

[0026] (1.2.1) Construct a mathematical model of task completion time, and then construct an objective function that minimizes the total completion time of all tasks.

[0027] In rescue scenarios, mission completion time is primarily affected by the drone's flight time en route, and wind is a significant factor influencing drone flight time. Therefore, the impact of wind must be considered when calculating drone flight time. This necessitates a correction for the environmental wind field effect, specifically including the following steps: Construct a flight time calculation model under the influence of wind fields: In order to accurately calculate the total time for completing all tasks and thus improve the rationality of task allocation, this embodiment improves the flight time of the UAV to take into account the influence of the environmental wind field.

[0028] drones Starting from the starting point, the flight speed in the air is During the flight segment arrive The airspeed vector on is The corresponding ground velocity vector is The environmental wind field vector is To maintain the drone's predetermined flight path, its flight direction needs to be adjusted to counteract the environmental wind vector. At this point, the drone's actual ground speed is no longer equal to its nominal airspeed, but rather determined by the airspeed vector. and environmental wind field vector Together, the vector composition relationship at this point is as follows: Figure 2 As shown, the mathematical relationship is as follows: in, For drones In the flight segment arrive The ground velocity vector.

[0029] Based on task points and mission points The coordinates are used to determine the flight segment. arrive Direction and distance: In the formula, For the segment arrive directional vector, For the segment arrive The distance is calculated using Euclidean distance in this embodiment. For the task point The position vector, For the task point The position vector.

[0030] Based on environmental wind field vector and flight segment arrive The direction vector is used to calculate the components of the environmental wind field vector in the parallel direction and the perpendicular direction of the flight segment: In the formula, Environmental wind field vector In the flight segment arrive The projected length (scalar) in the parallel direction. Environmental wind field vector In the flight segment arrive The component (vector) in the parallel direction. Environmental wind field vector In the flight segment arrive The vertical component (vector).

[0031] After the drone adjusts its flight direction, its flight path arrive The effective velocity in the direction can be expressed as: In the formula, For drones In the flight segment arrive Effective speed in direction.

[0032] To ensure the drone In the flight segment arrive For a flight to proceed normally, the following conditions must be met: The first condition indicates that the UAV has the ability to counteract crosswinds, and the second condition indicates that its effective propulsion speed along the flight path is positive.

[0033] Based on drones In the flight segment arrive Effective speed in direction and flight segment arrive distance Calculation of drones In the flight segment arrive Flight time: In the formula, For drones In the flight segment arrive Flight time. If the above conditions are not met, the drone will be... In the flight segment arrive The flight time is set to infinity to allow the scheme to be quickly eliminated during subsequent optimization.

[0034] Based on the flight time calculation model under the influence of wind fields, an objective function is constructed to minimize the total completion time of all tasks: In this implementation, the latest time when all drones return to base is used as the total time for all tasks to complete. The objective function is as follows: In the formula, Let represent the first objective function, which in this implementation is the total time to complete all tasks. To perform the minimum value operation, To perform the maximum value operation, For drones The time to return to base. This means the maximum time among all drones that return to base, which is the total time to complete all missions.

[0035] The calculation method for the return time of each drone to the base includes: In the formula, It is the set of all nodes that include the base and mission points. ,in Representing the base, For drones The time to return to base The mission point reached by the drone At that moment, For drones at the mission point The duration of the assignment For drones from the mission point The time required to fly back to base. For the coefficients of the decision variable, if the drone The last task point is And the drone after the mission was completed If the drone flies back to base from the mission point, the decision variable takes the value of 1; otherwise, the decision variable takes the value of 0. The formula represents the total mission completion time of the drone by summing the time it takes to reach the last mission point, the time it spends at the mission point, and the time it takes to fly back to base.

[0036] (1.2.2) Construct a task benefit calculation model, and then construct an objective function that maximizes the comprehensive rescue benefit.

[0037] In complex rescue scenarios, the quality of mission completion depends not only on whether drones arrive at the mission site, but also closely on the type of drones involved, their capability matching, and the actual operational time invested. Therefore, in this embodiment, the specific steps for constructing the mission benefit calculation model include: Construct an effective service volume calculation model: Based on drones With mission points The degree of matching and drones At the mission point Based on the actual operation time, an effective service volume calculation model is constructed. Specific methods include: In the formula, For the task point Effective service volume, For drones For the task point The ability matching coefficient For drones At the mission point Actual working time Indicates drone Do you participate in the mission? If you participate in the task, then ,otherwise .

[0038] In the formula, For drones Execution task point The operational efficiency coefficient is a preset empirical value, which is set when the drone is fully capable of fulfilling the mission requirements. When the required operational ability is 1, it has no task points whatsoever. When the required operational capability is not met (e.g., a reconnaissance drone cannot transport supplies), the value is 0, and the drone partially possesses mission capabilities. When the required operational capability is (for example, operational drones can also perform reconnaissance missions), its value is an empirical value between (0,1).

[0039] Build a task completion rate model: The relationship between task completion rate and effective service volume at a task point is not linear. For a specific rescue task, when the effective service volume is low, insufficient task information acquisition or inadequate service leads to a low success rate; when the effective service volume approaches the task requirement threshold, the success rate increases rapidly; and as the effective service volume continues to increase, the task completion quality gradually approaches saturation. To describe this non-linear change, the task completion rate model constructed in this embodiment is as follows: In the formula, Indicates task point Task completion rate For the task point Minimum effective service volume required. This is a coefficient for increasing the task completion rate, used to control the steepness of the task completion rate. This function ensures that the task completion rate remains within the (0,1) interval and reflects the change in the task completion rate from slow growth to rapid increase and then gradual saturation as the effective service volume increases. In this embodiment, =5, The value range is [0.5, 2]. A larger value indicates greater completion. For resource delivery missions, all resources must be delivered; any shortfall results in mission incompleteness. The mission is instantly completed after the last resource is delivered. Larger values, such as 2, conform to this trend: for target confirmation and disaster reconnaissance missions, the longer the operation time, the more accurate the confirmation. When the value is small, such as 0.5, the task completion rate changes relatively smoothly.

[0040] Constructing a rescue benefit model: In disaster relief scenarios, the value of a mission is typically highly time-sensitive. The earlier a mission is completed, the greater the relief benefits; as completion is delayed, the mission's value gradually diminishes. Let... Indicates task point The basic rescue value, Indicates task point The completion time, The time decay function representing the value of a task indicates the task point. The rescue benefit is defined as: In the formula, For the task point The benefits of the rescue efforts.

[0041] In this embodiment, the time decay function of the task value is as follows: In the formula, It is a natural exponential function. The time sensitivity coefficient, , The earliest start time window for services allowed at this task point. This refers to the actual time the task was completed.

[0042] Construct an objective function that maximizes the benefits of comprehensive rescue efforts: In the formula, This is the second objective function, which is also the comprehensive rescue benefit objective function.

[0043] (2.3) Set constraints.

[0044] In this embodiment, the constraints include: Task completion constraints: At least one drone capable of performing the task must travel to each task location to execute the task, and the drone's mission path must pass through the task location. The formula is expressed as: In the formula, Indicates drone From the mission point Fly to the mission point If so ,otherwise .

[0045] Endurance Constraint: For all drones, the time spent performing a mission and the time spent traveling to the mission point must not exceed the drone's endurance. The formula is as follows: In the formula, To indicate drones From the mission point Fly to the mission point If so ,otherwise .

[0046] Time window constraint: The drone arrives at the mission point Time It must be within the effective time window of the task. Inside. For the task point The earliest start time, For the task point The latest start time.

[0047] If the drone started earlier than the earliest start time Upon arrival, a waiting period will be incurred. This waiting time is also included in the total task time.

[0048] Payload Constraint: For UAVs performing material delivery missions, the cumulative weight of materials in a single flight sequence must not exceed their maximum design payload. The formula is expressed as follows: In the formula, For the task point The demand for supplies.

[0049] Step 2: Use the dual-population co-evolutionary algorithm to iteratively optimize the constructed multi-UAV task allocation model to obtain the final multi-UAV collaborative task allocation scheme.

[0050] The flowchart of the dual-population co-evolution algorithm of this invention is as follows: Figure 3 As shown, it includes the following steps: (2.1) Based on the number of UAVs and the number of mission points, construct an encoding and decoding mechanism, determine the encoding structure of individual position vectors, and construct an initial population based on the encoding structure.

[0051] To match the continuous evolution mechanism of LCE with the discrete characteristics of UAV task allocation, continuous variables are interpreted as probabilistic tendencies and relative priorities. This ensures that the neighborhood structure of the continuous space is effectively mapped to the discrete space, enabling the algorithm to utilize gradient-like information to guide the search. This invention establishes an encoding / decoding mechanism. The encoding / decoding mechanism is used to realize the conversion between multi-UAV task allocation schemes and individual position vectors, as follows: The encoding of each individual's position vector in the population consists of three parts: task assignment encoding, task ordering encoding, and task duration encoding, denoted as follows: , and The values ​​of each element in the vector range from [-1, 1]. The encoding methods for each encoding are illustrated in the diagram below. Figure 4 As shown, the details are as follows: The task assignment encoding method utilizes a threshold mapping mechanism for encoding and decoding, specifically including: In the formula, Assign codes to tasks. , Indicates drone For the task To determine the assignment preference, this implementation uses a threshold mapping mechanism to convert continuous values ​​into task participation variables. When the value is greater than or equal to the preset threshold, set A value of 1 indicates a drone. Execute the task Otherwise set A value of 0 indicates a drone. Do not perform task In this embodiment, the preset threshold is 0.5, which is the midpoint of the range [0,1].

[0052] In this implementation, to avoid the problem of a task being missed and no drone performing it, if a task is not selected by any drone with execution capability after initial coding, a compensation allocation strategy is adopted to assign the task to the drone with the highest assignment preference and execution capability. For tasks requiring collaborative completion, the same task point is allowed. ,different The value of is 1.

[0053] The task sorting encoding method uses the minimum position value rule for encoding, specifically including: Using the minimum position value rule, for The N-dimensional vectors in the sequence are sorted in ascending order to obtain the global task priority sequence. The execution path of each UAV is then obtained by scanning the tasks at each task point according to the global task priority sequence. For each UAV... If ordered according to the global task priority sequence ( (in order), judge in sequence Is it 1? If it is 1, then set the task point. Add to drones The execution path of each drone is obtained from the execution path.

[0054] The encoding method for job duration is based on linear scaling transformation: in, For drones At the mission point The job duration is encoded, with a value range of [0,1]. In this embodiment, it is decoded and mapped to the job time constraint interval using a linear mapping method. Specific decoding methods include: In the formula, For drones At the mission point The duration of the assignment For drones At the mission point The job duration code, This represents the lower limit of the task time constraint interval. This represents the upper limit of the time constraint interval for the task.

[0055] Through the above encoding and decoding mechanism, the encoding structure of the individual position vector is determined as follows: In the formula, Represents the position vector of an individual.

[0056] Based on the encoding structure of individual position vectors, the position vectors of each individual are generated using a constraint processing mechanism, thereby obtaining the initial population.

[0057] (2.2) Based on the initial population, the threshold mapping mechanism, minimum position value rule and linear scaling transformation are used for decoding mapping, and the results are substituted into the bi-objective optimization model to obtain the bi-objective fitness of the individual. Then, the initial population is divided into mining subpopulation and exploration subpopulation, and an archive set is constructed.

[0058] To address the shortcomings of single-population algorithms, this invention divides the initial population into an exploratory subpopulation and a mining subpopulation, forming a dual-population co-evolutionary algorithm. The specific method for dividing the exploratory and mining subpopulations is as follows: The top 50% of the initial population with better fitness are designated as the mining subpopulation (also known as the mining group), which performs deep local searches in the discovered high-potential areas; the bottom 50% are designated as the exploration subpopulation (also known as the exploration group), which performs large-scale global searches to find new potential optimal solution areas.

[0059] Since the optimization model constructed in this embodiment is a bi-objective optimization model, the obtained fitness is a bi-objective fitness. The method for judging the quality of individuals based on bi-objective fitness employs Pareto dominance judgment comparison. Specific comparison methods include: For individuals A and B, if neither of individual A's two optimization objectives is inferior to that of individual B, and at least one of individual A's two optimization objectives is strictly superior to that of individual B, then individual A is determined to be superior to individual B, and individual A dominates individual B. In this embodiment, the formula is expressed as follows: For individuals A and B, their bi-objective fitness are respectively ( , )and( , ),like and Established at the same time, and and If at least one of the following conditions is met, then individual A is considered superior to individual B, and individual A dominates individual B. and These represent the total time taken to complete all tasks for individuals A and B, respectively; a smaller value is better. and These represent the comprehensive rescue benefits for individuals A and B, with higher values ​​indicating better outcomes.

[0060] The method for dividing the mining subpopulation and the exploration subpopulation in this embodiment includes: All individuals were ranked from best to worst based on two indicators: non-dominated ranking hierarchy and comprehensive crowding distance. The top 50% were selected as the mining subpopulation, and the bottom 50% were selected as the exploration subpopulation.

[0061] Methods for determining non-dominated ordering levels include: For all individuals in the population, perform pairwise Pareto dominance comparisons. The non-dominated individuals (individuals not dominated by any individual) obtained from the first comparison are designated as level 1. Perform a second pairwise Pareto dominance comparison on the dominant individuals, and the non-dominated individuals (individuals not dominated by any individual) obtained from the second comparison are designated as level 2, and so on, to obtain the non-dominated ranking levels.

[0062] Individuals in level 1 are ranked higher than those in level 2, and within the same level, they are ranked according to their overall crowding distance.

[0063] The method for calculating the overall crowding distance includes: sorting all individuals within a level according to F1 and F2 respectively, resulting in two sequences (an F1 sequence and an F2 sequence). For any sequence, the crowding distance of the individuals at both ends of the sorted sequence is set to infinity (indicating that the individuals are the sparsest and must be retained). For other individuals, the absolute value of the difference between the values ​​of the two adjacent individuals is calculated, and then divided by the difference between the maximum and minimum values ​​of the level to obtain one crowding distance for that individual (if the difference between the maximum and minimum values ​​of the level is 0, then the crowding distance of all individuals in that level is directly set to 0). The same process is performed on the other sequence to obtain another crowding distance for the individual. The two crowding distances of the same individual are summed to obtain the overall crowding distance for that individual.

[0064] Using the Pareto dominance comparison method, non-dominant individuals in the initial population are identified, and these identified non-dominant individuals are grouped into an archive set. Figure 3 Also known as the non-dominated solution set, it is used for global comparison with the individuals after position updates.

[0065] Within the same level, the greater the difference in overall congestion, the higher the ranking.

[0066] (2.3) Iteratively update the individual positions and calculate the bi-objective fitness of each individual after the update. Re-divide the exploration subpopulation and mining subpopulation, and compare their performance with that of individuals in the archive set, and then update the archive set.

[0067] The present invention relates to a dual-population co-evolutionary algorithm, comprising an exploratory subpopulation and a mining subpopulation. The exploratory subpopulation employs chaotic mapping and geometric spiral search mechanisms to maintain population diversity; the mining subpopulation uses an adaptive mapping exchange mechanism and a greedy mutation operator to mine potentially high-quality regions, thereby improving convergence accuracy. The two subpopulations achieve co-evolution by sharing a globally non-dominated solution set (archive set).

[0068] The iterative update methods for the individual positions in the two subpopulations are as follows: Exploring iterative update methods for individual positions in subpopulations: The goal of exploring the subpopulation is to maximize the search range. Unlike the original LGC (Logistic-Gauss Circle, LGC optimizer) algorithm, which only relies on Logistic and Gaussian mappings, this implementation proposes a hybrid strategy that combines chaotic perturbation and spiral approximation. This strategy selects the update mode based on the switching probability. Specifically, before updating the individual position in the exploring subpopulation, a random number is generated within the range [0,1]. ,like If so, the individual uses the chaotic sequence to make long-distance jumps to escape local extrema and increase the search range; if At that time, the individual approaches the global optimal solution along the logarithmic spiral path, while simultaneously shrinking the encirclement and searching for the local optimal solution. In this embodiment, as the switching probability, .

[0069] This individual utilizes chaotic sequences to make long-distance jumps: In the formula, To explore the first subpopulation Individuals The position vector of the next iteration. To explore the first subpopulation Individuals The position vector of the next iteration. It is a chaotic vector generated by the Logistic mapping, also with dimension 1. A continuous real vector of dimensionless numbers, where the values ​​of each element range from [0,1]. The upper bound of the search space is a A vector in dimension where all elements are 1. Let be the lower bound of the search space, and let be a D-dimensional vector with all elements equal to 0.

[0070] Methods for individuals to approximate the global optimum along a logarithmic spiral path include: In the formula, This is the position vector of the global optimal solution. To explore the first subpopulation Individuals The position of the next iteration and the global optimal solution The distance between the positions, To define the constants for the spiral shape, A random number between [-1, 1]. In this embodiment, The value is taken as an empirical value of 1. Since the fitness in this implementation is a bi-objective fitness, and the quality of an individual is determined by the Pareto dominance comparison method, the resulting global optimal solution is a non-dominated solution in the archive set, and there may be multiple such solutions. To address this issue, this implementation calculates the comprehensive crowding distance of each individual in the archive set and performs sum normalization to obtain the probability that each individual in the archive set is the global optimal solution. Then, based on this probability, a global optimal solution is randomly selected from the archive set. The larger the crowding distance, the higher the probability of an individual being the global optimal solution. In this way, individuals can be guided to search towards the sparse region of the Pareto front using the guided algorithm.

[0071] The purpose of mining subpopulations is to improve convergence accuracy. In this implementation, EES (Experience Exchange Strategy) is used for iterative updates of individual positions. The original EES includes three stages: experience scarcity, crossover, and sharing, which is a relatively cumbersome process. This implementation transforms the idea of ​​historical differences and global guidance into a lightweight adaptive differential search, introducing a greedy exchange mutation for discrete problems. Specifically, it includes: Adaptive Differential Search: To balance the learning and self-evolution mechanisms of the algorithm, the position update formula is as follows: In the formula, To mine the first subpopulation The first individual The position vector of the nth iteration. To mine the first subpopulation The first individual The position vector of the nth iteration. and For two randomly selected individuals, and These are time-varying adaptive weights, representing exploration weights and mining weights respectively: This mechanism ensures that the algorithm focuses on utilizing the differences between populations in the early stages, while focusing on rapid convergence to the global optimum in the later stages.

[0072] After the data computation of adaptive differential search, the elements in the updated position vector of an individual position may exceed the range of [0,1]. At this time, truncation is required. If an element in the updated position vector of an individual position is greater than 1, it is truncated to 1. If an element in the updated position vector of an individual position is less than 0, it is truncated to 0.

[0073] Greedy swap mutation: Since the UAV mission sequence is a discrete problem with a specific order, updating individual positions using adaptive differential search with continuous numerical calculation formulas would result in invalid mission sorting. To address this issue, this implementation incorporates a greedy exchange mutation mechanism, with mutation probability... Trigger mutation, randomly swap individuals The positions of two task points in the task sequence are used to obtain candidate solutions. If candidate solutions Its fitness is better than the original solution. If the mutation is positive, then the mutation is accepted; otherwise, the mutation is not accepted. The formula is expressed as follows: In the formula, In order to carry out the greedy exchange mutation in the mining subpopulation, the first The first individual The position vector of the iteration. In this embodiment, the fitness is a bi-objective fitness, and the Pareto dominance comparison is used to determine the quality. The specific method has been described in step (2.2) and will not be repeated here.

[0074] Mutation probability In this embodiment, the hyperparameters are preset. .

[0075] After updating the positions of all individuals, based on the bi-objective fitness of each individual after the position update, the non-dominated ranking level and the comprehensive crowding distance are recalculated. Based on the two indicators of non-dominated ranking level and comprehensive crowding distance, all individuals are ranked in order from best to worst. The top 50% are taken as the mining subpopulation and the bottom 50% are taken as the exploration subpopulation, thus realizing the re-division of the mining subpopulation and the exploration subpopulation.

[0076] After updating the positions of all individuals, the updated bi-objective fitness of each individual is compared with the bi-objective fitness of all individuals in the archive. If the updated bi-objective fitness of an individual is not dominated by the bi-objective fitness of any individual in the archive, then that individual is added to the archive. If adding that individual causes the number of individuals in the archive to exceed the preset limit, then the overall crowding distance of all individuals in the original archive is calculated, and the individual with the smallest overall crowding distance is deleted.

[0077] In this implementation, the exploration subpopulation utilizes chaotic perturbations for large-scale blind search, potentially "parachuting" into a completely new and better region at any time. This avoids getting trapped in local optima and expands the algorithm's global search capabilities. The mining subpopulation does not have to undertake the task of "running around aimlessly" but instead focuses on "refining" known good solutions using an adaptive difference mechanism (such as fine-tuning the operation time by a few minutes or swapping the order of two tasks). This makes the algorithm converge faster and the solution more accurate when approaching the final optimal solution. At the same time, the two subpopulations share a common archive set, which consists of globally optimal solutions, enabling the two subpopulations to form a co-evolutionary method of "global exploration and local deep mining".

[0078] (2.4) Iterate through step (2.3) until the set number of iterations is reached, then stop iterating. The resulting set of files is the Pareto optimal solution set of the final multi-UAV collaborative task allocation scheme. Each individual file represents a multi-UAV collaborative task allocation scheme.

[0079] The following comparative analysis is conducted with the dual-population co-evolution method of this invention, using several other methods: A comparative experiment was conducted between the Knee-point based Coevolution Multi-objective Particle Swarm Optimization (KnCMPSO) algorithm for solving the UAV cooperative task allocation problem and three representative algorithms: Coevolutionary Multiswarm Particle Swarm Optimization (CMPSO), Coevolutionary Particle Swarm Optimization (CPSO), and Coevolutionary Multi-objective Local Search based on Decomposition (CoMOLS / D). To ensure fairness, the population size for all algorithms was set to 30, the maximum number of iterations to 500, and the external archive size to 100. The Hypervolume (HV) metric was used to evaluate the quality of each algorithm in obtaining the non-dominated solution set. The HV metric is a comprehensive evaluation metric that can simultaneously assess the algorithm's performance in convergence and diversity. A higher HV value indicates better performance in both convergence and diversity. Considering the physical significance of the target in this study, the reference point is set as... .in, This represents the sum of the maximum flight times of all drones, which is the upper bound for achieving the time target. This represents the lower bound of the task's reward objective. This setting ensures that the reference point can encompass the space of all valid solutions generated by the algorithms, guaranteeing the fairness of the evaluation.

[0080] The experiment constructed a A two-dimensional task area of ​​meters was designed with an environmental wind field set to [3, -2] m / s. Four scenarios of different scales were designed, and the task point distribution of the four scenarios is compared as follows: Figures 5-8 As shown, scenario 1 includes 5 drones and 10 tasks, with the task locations distributed as follows: Figure 5 As shown, Scenario 2 includes 8 drones and 15 tasks, with the task locations distributed as follows: Figure 6 As shown, Scenario 3 includes 11 drones and 20 tasks, with the task locations distributed as follows: Figure 7 As shown, Scenario 4 includes 14 drones and 25 tasks, with the task locations distributed as follows: Figure 8 As shown, the mission objectives in each scenario are randomly set at fixed locations within a two-dimensional rescue area. Each mission objective contains three key attributes, representing the set of capabilities required for that rescue point (determined based on the rescue type), the mission's time window constraint, and the basic rescue value. The heterogeneous drones in each scenario are also initialized at fixed base locations. Each drone includes its flight speed, maximum endurance, and the type of payload it carries.

[0081] The comparison algorithms all adopted the same encoding strategy and initialization method as the MO-LCE (Multi-Objective Two-Population Co-evolutionary Method) of this invention. The HV index data for each scenario are shown in Table 2. Mean represents the average HV index calculated by each algorithm in 30 experiments; its magnitude reflects the convergence and diversity of the non-dominated solution set obtained by the algorithm. Std represents the variance of the HV index calculated by each algorithm in 30 experiments; its magnitude reflects the stability of the algorithm. Best represents the optimal HV value obtained after 30 runs of the algorithm. Worst represents the worst HV value obtained by the algorithm. The optimal results for each test function in Table 2 are marked in bold, and the suboptimal results are marked with an underline.

[0082] Table 2. Comparison of performance indicators of different algorithms for solving the UAV task allocation problem As shown in Table 2, the optimal solution sets of the UAV cooperative task allocation schemes obtained by the multi-objective dual-population cooperative evolution method of the present invention have good convergence and diversity in all four scenarios.

[0083] Box plot comparison of each algorithm in 4 scenarios is shown below. Figures 9-12 As shown, where Figure 9 , Figure 10 , Figure 11 , Figure 12 Box plots are shown for scenarios 1, 2, 3, and 4, respectively. To visually demonstrate the distribution of the algorithm's results in different scenarios and to facilitate Pareto front comparison, the negative value of the comprehensive rescue benefit is used. A comparison of the distribution of the solution sets obtained by the algorithm in the four scenarios is shown below. Figures 13-16 As shown, where Figure 13 , Figure 14 , Figure 15 , Figure 16 The figures show the distribution of solution sets for scenarios 1, 2, 3, and 4, respectively. As can be seen from the distribution of solution sets for each scenario, compared to other comparative algorithms, the solution sets obtained by the algorithm of this invention are superior in both diversity and convergence, which is consistent with the evaluation results of the HV index mentioned above.

Claims

1. A heterogeneous multi-UAV task allocation method based on dual-population co-evolution, characterized in that, Includes the following steps: (1) Based on the parameters of the task point to be rescued and the parameters of the heterogeneous UAV swarm, construct the mathematical model of the task point to be rescued and the mathematical model of the heterogeneous UAV swarm, and set the constraints and objective functions to obtain the multi-UAV task allocation optimization model; (2) The dual-population co-evolutionary algorithm is used to iteratively optimize the constructed multi-UAV task allocation model to obtain the optimal solution set of the multi-UAV cooperative task allocation scheme; specifically including: (2.1) Construct the initial population; (2.2) Based on the position of each individual in the initial population, calculate the fitness of each individual in the initial population, determine the superiority or inferiority of each individual, and then divide the initial population into an exploration subpopulation and a mining subpopulation; and construct an archive set; (2.3) Update the position of each individual in the exploration subpopulation and the mining subpopulation respectively, calculate the fitness of each individual after the position update, determine the quality of each individual after the position update, and then divide all individuals after the position update into the exploration subpopulation and the mining subpopulation, and update the archive set; (2.4) Iterate through step (2.3) until the iteration termination condition is met. The resulting file set is the optimal solution set for the multi-UAV collaborative task allocation scheme.

2. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution as described in claim 1, characterized in that, Methods for updating the position of each individual in the mining subpopulation include: using adaptive differential search and greedy exchange mutation to update the position of each individual in the mining subpopulation.

3. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 1, characterized in that, Methods for updating the position of each individual in the exploration subpopulation include: To update the position of each individual in the exploratory subpopulation, the method involves randomly performing long-distance jumps based on chaotic sequences or approximating the global optimum along a logarithmic spiral path.

4. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 1, characterized in that, The objective function is a bi-objective function, which includes minimizing the total time for completing all tasks and maximizing the overall rescue benefits.

5. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 4, characterized in that, The fitness is a bi-objective fitness, and the method for determining the superiority or inferiority of individuals includes the Pareto dominance comparison method.

6. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 5, characterized in that, Methods for constructing archive sets include: using the Pareto dominance comparison determination method to identify non-dominant individuals in the initial population, thereby constructing the archive set.

7. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 5, characterized in that, The methods for updating the archive set include: based on the fitness of each individual after the position update and the fitness of each individual in the archive set, using the Pareto dominance comparison method to compare the quality of each individual with that of each individual in the archive set; for any individual, if it is not dominated by any individual in the archive set, then the individual is added to the archive set.

8. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 1, characterized in that, Methods for constructing an initial population include: An encoding and decoding mechanism is constructed to encode the multi-UAV task allocation scheme into individual position vectors. The encoding structure of the individual position vectors is determined, and then the position vectors of each individual in the initial population are generated by using a constraint-based processing mechanism to obtain the initial population.

9. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 4, characterized in that, The methods for calculating the total time to complete all tasks include: For any drone, sum the time it takes to reach the last mission point, the time it spends at the mission point, and the time it takes to fly back to the base from the mission point to get the total mission completion time for that drone. The total completion time of all tasks is obtained by taking the maximum value from the total completion time of all drones.

10. The heterogeneous multi-UAV task allocation method based on dual-population co-evolution according to claim 4, characterized in that, The methods for calculating comprehensive relief benefits include: Substitute the completion time of each task into the time decay function; By combining the effective service volume, task completion rate, and time decay function calculation results of each task point, the rescue benefits of each task are obtained. The total rescue benefit is obtained by summing the rescue benefits from all missions.

Images