EM pre-clustering unmanned aerial vehicle cluster task allocation method based on probability inference

Abstract

The invention discloses an EM pre-clustering unmanned aerial vehicle cluster task allocation method based on probability inference. The method comprises the steps that 1, parameters are initialized; step 2, according to all initialized parameters, calculating to obtain a posterior probability by adopting an E step in an EM algorithm; 3, calculating the maximum likelihood probability by adopting M steps according to the posterior probability, and estimating a new model parameter; 4, judging whether the model parameters converge or not, and if yes, executing the step 5; otherwise, returning to the step 2; 5, outputting the model parameter and the posterior probability of the last iteration; step 6, setting the range of epsilon values and the attenuation rate according to the model parameters; the epsilon value in the Markov process is initialized in the range of the epsilon value; for the epsilon value, the position of the unmanned aerial vehicle is initialized, and the current unmanned aerial vehicle path and the data structure of the total moving distance are recorded; and 7, for each unmanned aerial vehicle, searching the path and the total moving distance of the unmanned aerial vehicle. The method provided by the invention can reduce the overhead and improve the task allocation efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle cluster task allocation model, and specifically relates to an EM pre-clustering unmanned aerial vehicle cluster task allocation method based on probability inference. Background Art

[0002] The task allocation of drone swarms refers to the scientific allocation of specific tasks that need to be performed under a certain task environment. The goal is to optimize the task allocation plan and achieve the optimal task execution efficiency under the premise of being able to complete the specific tasks. This problem is essentially a multi-objective optimization problem with multiple constraints. It refers to the process of allocating specific tasks or goals to single or multiple drones in the group under constraints such as the drone's own capabilities and limitations, environmental conditions, and the interdependence between tasks, so as to efficiently and effectively achieve the global task goals. The goal is to optimize task allocation to maximize the collective performance of the drone swarm while ensuring collaboration and avoiding conflicts. Although the current algorithms have made some progress in the study of task allocation, as the number of drones or the complexity of tasks increases, the existing algorithms usually encounter difficulties in scalability and other aspects. Here, the current status of task allocation research is analyzed through two classification methods as follows:

[0003] (1) According to different allocation methods, it is divided into centralized and distributed task allocation.

[0004] The centralized task allocation problem of drone swarms refers to the decision-making method of allocating tasks to individual drones in the swarm by a single central controller or coordinator. The goal is to find a task allocation scheme that maximizes the performance of the entire swarm under constraints such as the maximum load capacity of each drone, the time required to complete the task, and the need for coordination between drones to avoid conflicts. Centralized task allocation methods have been widely studied, including algorithms based on mathematical models and algorithms based on swarm intelligence. The mathematical model-based method has been widely verified by scientific research and industry due to its early research time. In 2018, Hefei University of Technology considered the dynamic constraints of drones, constructed the Durbin path variable speed VRP model using the Durbin algorithm, and studied the task allocation of fixed-wing drones under the influence of wind. In the same year, Wang Xiaoqun set the search object as a Markov motion target, and used the Markov decision process to allocate search tasks to each drone based on the probability of target motion through multi-drone collaborative search, and proposed a variety of different search strategies for task allocation. In 2022, Chen Yuheng and others from Northwestern Polytechnical University used directed graphs in graph theory to represent the collaborative flight constraint relationship between tasks, and based on the improved greedy algorithm, they performed local optimal allocation and optimization of tasks, and proposed a UAV cluster collaborative task allocation algorithm based on graph theory; the UAV cluster task allocation algorithm based on swarm intelligence is derived from the intelligent bionic algorithm of the objective laws of nature and the behavior of biological groups, which can better handle large-scale optimization problems. Including ant colony algorithm, pigeon colony algorithm, wolf colony algorithm, particle swarm algorithm, etc. In 2022, Zhang Ruipeng proposed a strategy to jump out of local convergence in response to the problem that the traditional particle swarm optimization (PSO) algorithm is prone to fall into local convergence, established a startup probability criterion for local search, and achieved a balance between jumping out of local convergence and computational cost. In 2024, in order to improve the optimization ability of the wolf pack algorithm, Zhang Chenyang proposed a UAV cluster task allocation method based on an improved wolf pack algorithm based on adaptive adjustment step size to avoid falling into the local optimal problem.

[0005] The distributed task allocation problem of drone swarms is a decentralized decision-making process, in which each drone makes local decisions on task allocation independently or in collaboration with a subset of other drones. The goal is to ensure that the tasks assigned to each drone can be effectively executed, while taking into account its individual capabilities and the current environmental context, maintaining the overall mission objectives, and avoiding task conflicts. Distributed algorithms include contract network methods similar to market mechanisms, auction algorithms, and multi-agent algorithms applied in the field of robotics. In 2021, the China Ship Research and Design Center proposed an improved contract network task allocation method to address the problems of unbalanced resources and low real-time performance in the task allocation problem of heterogeneous multi-UAVs, combined with actual combat needs. It effectively improves the global load balancing capability and has good scalability.

[0006] (2) Depending on whether cluster allocation is performed in advance, it is divided into overall allocation and task pre-allocation.

[0007] The overall allocation algorithm directly allocates tasks to drones. This method focuses on how to optimize the overall allocation strategy to achieve overall performance optimization of the drone cluster. Overall allocation algorithms include greedy algorithms, genetic algorithms, etc. In 2017, the University of Science and Technology of China proposed a task allocation algorithm based on greedy strategy and minimum weighted set cover for this NP-hard problem, and proved the superiority of the proposed algorithm. In 2024, Wu Zhangao used genetic algorithms to optimize the objective function and obtained the optimal drone inspection task allocation algorithm, which improved the efficiency of the algorithm.

[0008] Task pre-allocation means that the task planning platform obtains the location and task type of the target point on the battlefield in advance through detection equipment, first clusters or groups the tasks, and then assigns these clustered specific tasks to each drone. This method can reduce the complexity of the problem and improve the allocation efficiency. In 2022, in order to improve the capture efficiency of large-scale task allocation problems, Northwestern Polytechnical University used an improved K-means algorithm to stratify the multi-target capture problem, and used the shortest total time mechanism to establish a one-to-one matching relationship between subtasks and drones, proving the effectiveness of the allocation strategy. In 2023, a certain document used the EM algorithm based on the Gaussian mixture model (GMM) to obtain the initial position of each drone, and applied the algorithm to the drone task allocation based on the initial position of the drone. This method improves the iteration speed.

[0009] To sum up the above methods, although there has been some progress in the research of task allocation, with the increase in the number of UAVs or the complexity of tasks, algorithms are needed that can adapt to the dynamic changes of the environment or task requirements in real time without significant computational overhead. At the same time, there is a lack of research on how to effectively coordinate complex tasks that require multiple UAVs to act synchronously. Summary of the invention

[0010] The purpose of the present invention is to provide an EM pre-clustering UAV cluster task allocation method based on probability inference to meet the problem of real-time adaptation to dynamic changes in the environment or task requirements while not significantly increasing the computational overhead, thereby ensuring task efficiency.

[0011] In order to achieve the above object, the present invention adopts the following technical solution:

[0012] A method for allocating tasks of drone clusters based on EM pre-clustering based on probability inference specifically includes the following steps:

[0013] Step 1, initialize the following parameters: number of drone clusters U, number of targets N, mountain terrain map, model parameters, clustering probability, and distance matrix;

[0014] Step 2: Based on all the parameters initialized in step 1, the E step in the EM algorithm is used to calculate the posterior probability;

[0015] Step 3: According to the posterior probability, the M step in the EM algorithm is used to calculate the maximum likelihood probability, and the new model parameters are estimated according to the current posterior probability;

[0016] Step 4, determine whether the model parameters obtained in step 3 have converged, if so, execute step 5; otherwise, return to step 2 for iteration;

[0017] Step 5: Output the model parameters and posterior probability of the last iteration;

[0018] Step 6, set the range and decay rate of the epsilon value according to the model parameters output in step 5; initialize the epsilon value in the Markov process within the range of the epsilon value; for the epsilon value, initialize the drone position, and record the data structure of the current drone path and total moving distance;

[0019] Step 7: For each drone, set its search starting point to the current drone position, and search for the drone path and total moving distance.

[0020] Furthermore, in step 1, the model parameters include clustering probability, cluster center and covariance matrix.

[0021] Furthermore, step 7 specifically includes the following sub-steps:

[0022] Step 71, for each drone, set its search starting point to the current drone position. When there are still unfound targets, generate a random number. If the random number is less than the current epsilon value, the drone performs exploration behavior and then executes step 72; if the random number is not less than the current epsilon value, no exploration behavior is performed and step 73 is executed; when there are no unfound targets, step 73 is executed;

[0023] Step 72, the drone moves to the selected target position, calculates the moving distance and accumulates the current moving total distance to update it, updates the search path, removes the found target, and updates the current position of the drone, then gradually reduces the epsilon value according to the preset decay rate within the range of the epsilon value as the current epsilon value, and returns to step 71;

[0024] Step 73, record the paths and total moving distances of all drones under the current epsilon value.

[0025] Furthermore, in step 71, the drone performs the exploration behavior specifically by calculating the distances between all remaining targets and its current position, giving priority to targets within a strict range, expanding the range if there are no targets within the strict range, and finally randomly selecting a target as the next target. If there are targets within the strict range, the drone selects the target that is closest.

[0026] Compared with the prior art, the present invention has the following technical effects:

[0027] (1) The pre-clustering method is used to cluster the task allocation area first. This processing method improves the accuracy and efficiency of the algorithm and reduces the complexity; thereby effectively reducing the overhead and improving the efficiency of task allocation.

[0028] (2) The mountain search based on the three-dimensional scene of the present invention is not only based on the common two-dimensional plane modeling, but also more intuitive and realistic in applicable scenes;

[0029] Through algorithm analysis and simulation experiments, the present invention is compared with other task allocation methods and has better effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the EM pre-clustering UAV cluster task allocation method (preEM-EG) based on probability inference of the present invention;

[0031] Figure 2 It is a three-dimensional graph of the EM pre-clustering results in preEM-EG;

[0032] Figure 3 It is the path plan after the task allocation in preEM-EG is completed;

[0033] Figure 4 It is a three-dimensional diagram of the path after the task allocation in preEM-EG is completed;

[0034] Figure 5 is a graph of total distance versus time;

[0035] Figure 6 It is a graph of total distance changing with epsilon value;

[0036] Figure 7 It is the task allocation result of the annealing algorithm of K-means pre-clustering;

[0037] Figure 8 It is a graph showing the total distance of the annealing algorithm for K-means pre-clustering changing with the number of iterations;

[0038] Fig. 9 It is the task allocation result of particle swarm algorithm;

[0039] Fig.10 It is the curve of the total distance of the particle swarm algorithm changing with the number of iterations;

[0040] Fig.11 This is a comparison curve of the total distance of the three algorithms changing with the number of iterations.

[0041] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0042] Mountain search mission is a common drone swarm mission. The mountain search mission is established in a three-dimensional scene environment with undulating terrain. Since the target is generally a group action and will move in a small range when fleeing and hiding, the mission goal of the drone swarm is to search for multiple moving targets. For the mountain search mission, the present invention provides an EM pre-clustering drone swarm task allocation method based on probability inference, which specifically includes the following steps:

[0043] Step 1: Initialize the following parameters: number of drone clusters U, number of targets N, mountain terrain map, model parameters, clustering probability, and distance matrix; the model parameters include clustering probability π n , cluster center μ n and the covariance matrix Σ n ;

[0044] Step 2: Based on all the parameters initialized in step 1, the E step in the EM algorithm is used to calculate the posterior probability;

[0045] Step 3: According to the posterior probability, the M step in the EM algorithm is used to calculate the maximum likelihood probability, and the new model parameters are estimated according to the current posterior probability;

[0046] Step 4, determine whether the model parameters obtained in step 3 have converged, if so, execute step 5; otherwise, return to step 2 for iteration;

[0047] Step 5: Output the model parameters and posterior probability of the last iteration.

[0048] Step 6, set the range and decay rate of the epsilon value according to the model parameters output in step 5; initialize the epsilon value in the Markov process within the range of the epsilon value; for the epsilon value, initialize the drone position, and record the data structure of the current drone path and total moving distance;

[0049] Step 7: For each drone, set the search starting point to the current drone position, and search the drone path and total moving distance in the following way. This includes the following sub-steps:

[0050] Step 71, for each drone, set its search starting point to the current drone position. When there are still unfound targets, generate a random number. If the random number is less than the current epsilon value, the drone performs exploration behavior and then executes step 72; if the random number is not less than the current epsilon value, no exploration behavior is performed and step 73 is executed; when there are no unfound targets (that is, all targets have been found), execute step 73;

[0051] The specific exploration behavior of the drone is as follows: calculating the distance between all remaining targets and its current position, giving priority to targets within a strict range (such as 50 meters), if there is no target within the strict range, expanding the range (such as 100 meters), and finally randomly selecting a target as the next target. If there is a target within the strict range, the drone selects the target that is closest.

[0052] In step 72, the UAV moves to the selected target position, calculates the moving distance and accumulates the current total moving distance to update it, updates the search path, removes the found target, and updates the current position of the UAV. Then, the epsilon value is gradually reduced according to the preset decay rate within the range of the epsilon value as the current epsilon value, and returns to step 71; wherein, gradually reducing the epsilon value according to the preset decay rate can reduce exploration behavior and increase the opportunity to use known information.

[0053] Step 73, record the paths and total moving distances of all drones under the current epsilon value.

[0054] After the method of the present invention is finished, the total running time of the whole process is displayed. The whole process ensures that the drone can both effectively update and utilize known information during the search process, optimize the search path and reduce the total distance by dynamically adjusting the epsilon value.

[0055] In order to prove the feasibility and effectiveness of the present invention, the following analysis and simulation experiments are carried out.

[0056] 1. Algorithm complexity analysis

[0057] (1) Time complexity

[0058] The time complexity of the preEM-EG algorithm is divided into two parts: EM algorithm + epsilon-greedy algorithm. The algorithm is divided into two parts to calculate their time and space complexities respectively, and then they are linearly added.

[0059] The time complexity of the EM algorithm is determined by the E step and the M step. The time complexity of the E step is In this step, the algorithm needs to calculate the posterior probability value of each target belonging to each cluster. The time complexity of step M is The algorithm updates the mean, covariance matrix and other parameters of each cluster. The time complexity of a single iteration is Since it takes multiple iterations to converge, the total time complexity of the EM algorithm is i is the total number of iterations required.

[0060] The Epsilon greedy strategy algorithm is used to calculate the order in which the drone searches for targets within the defined cluster range. Since the action selection is based on the current state and known strategy conditions, the time complexity of each action selection is Similarly, the time complexity of updating the reward function and the number of actions is Then for each goal, the time complexity of the algorithm is The time complexity of traversing all targets is

[0061] The total time complexity is the sum of the time complexity of the EM algorithm and the epsilon greedy strategy.

[0062]

[0063] (2) Space complexity

[0064] The space complexity of the EM algorithm depends on the parameters required to store the clusters and the information required to assign the target tasks. To store the cluster assignment for each data point and the parameters for each cluster, the space complexity of the EM algorithm is

[0065] The space complexity of the Epsilon greedy strategy algorithm depends on the storage value function and action count. Each state and action has a reward function and action count, resulting in a space complexity of S is the number of states.

[0066] The total space complexity will be the sum of the space complexity of the EM algorithm and the epsilon greedy strategy algorithm.

[0067] 2. Simulation experiment

[0068] (1) Parameter setting

[0069] System parameters: Windows 11 system, hardware parameters: Intel(R) Core(TM) i9-14900HX2.20GHz, 16G RAM, NVIDIA GeForce RTX 4060GPU. Software environment parameters: MATLAB R2023b software environment. The experiments of the present invention are all based on the above experimental parameters, which will not be repeated in the following text.

[0070] In the experimental parameter setting, we first need to assign the task environment for the mountain target search task, and then give the remaining parameters. First, set the mountain environment parameters, including the length, width, and height of the area, as well as the radius and position of the added mountain. Set the drone cluster to have a total of 7 aircraft, set the number of targets to 100, and set the search step size, epsilon maximum and minimum parameter values, epsilon decay rate, etc. The main parameter settings of the experiment are shown in Table 3:

[0071] Table 3 Experimental parameter settings of preEM-EG algorithm

[0072]

[0073] (2) Experimental results

[0074] For the mountain target search task, the experimental results of running the preEM-EG algorithm on MATLAB are as follows: Figure 2 shown.

[0075] Figure 3 The topography of the plan view is represented by the color shades from green to yellow. The greener the color, the lower the altitude, and the yellower the color, the higher the altitude. Figure 4 Display the situation after task allocation from a three-dimensional perspective. Figure 3 and Figure 4 As can be seen, the preEM-EG algorithm divides the 100 target points into 7 clusters (represented by seven colors) and gives the allocation order of the UAV search tasks through arrows.

[0076] In order to better analyze the experimental results, the curve graph of the total distance changing with time is shown as follows Figure 5 As shown in the figure, the total flight distance of the drone cluster increases almost linearly and converges at about the 20th second, with a total path of 10,287 meters. In other words, the drone cluster has completed the search and visit of all targets at the 20th second.

[0077] Next, we analyze how the total distance varies with the value of epsilon. Figure 6 The total distance is plotted against the epsilon value. Figure 6 As shown:

[0078] At the same time, the tables of total distance values ​​when the epsilon value is 1 to 9 are shown in Table 4 and Table 5 below.

[0079] Table 4 Total distance values ​​when epsilon values ​​are 1, 2, 3, 4, and 5

[0080]

[0081]

[0082] Table 5 Total distance values ​​when epsilon values ​​are 6, 7, 8, and 9

[0083]

[0084] from Figure 6 It can be seen from Tables 4 and 5 that when e is 1, the total distance is the shortest, which is 10287 meters. At the same time, it can be seen that the overall trend of the total distance increases with the increase of the value of e. This shows that in this terrain environment, the drone cluster has the best effect when the e value is 1, and the larger the e value, the worse the effect.

[0085] (3) Comparative test and analysis

[0086] In order to compare the experimental results, two classic methods of task allocation were selected, and the k-means clustering annealing algorithm and particle swarm algorithm were used for comparison. Among them, the preEM-EG algorithm selected the result when e equals 1. The experimental results are shown in Figure 2. Figure 7 As shown. Figure 7 It can be seen that the task allocation of the K-means pre-clustering annealing algorithm is more reasonable, but most drones need to climb.

[0087] from Figure 8 It can be seen that the global optimal distance of the annealing algorithm reaches 12297.8 meters when the number of iterations is 100 (converged), which is longer than that of the preEM-EG algorithm.

[0088] from Fig. 9 It can be seen that the clustering effect of the particle swarm algorithm is not as good as the first two, and there are many long-distance flying phenomena. Fig.10 It is concluded that the global optimal distance of the particle swarm algorithm reaches 11461.6 meters when the number of iterations is 100 (converged), which is longer than that of the preEM-EG algorithm.

[0089] From the curve graph of the total distance of the three algorithms changing with the number of iterations of each algorithm ( Fig.11 ) It can be seen that when e is equal to 1, the total flight distance required for the drone cluster task allocation of the preEM-EG algorithm is shorter than that of the particle swarm algorithm and the K-means pre-clustering annealing algorithm. The K-means pre-clustering method converges slowly and has the worst effect. Although the particle swarm converges quickly, it is easy to fall into the local optimum and cannot achieve the global optimal strategy. The effect is not as good as the preEM-EG algorithm. The comparison of the three algorithms proves the superiority of the preEM-EG algorithm, that is, the superiority of the EM pre-clustering drone cluster task allocation invention based on probability inference.

Claims

1. A method for allocating tasks of drone clusters based on EM pre-clustering based on probability inference, characterized in that: The specific steps include: Step 1, initialize parameters: number of drone clusters U, number of targets N, mountain terrain map, model parameters, clustering probability, and distance matrix; Step 2: Based on all the parameters initialized in step 1, the E step in the EM algorithm is used to calculate the posterior probability; Step 3: According to the posterior probability, the M step in the EM algorithm is used to calculate the maximum likelihood probability, and the new model parameters are estimated according to the current posterior probability; Step 4, determine whether the model parameters obtained in step 3 have converged, if so, execute step 5; otherwise, return to step 2 for iteration; Step 5: Output the model parameters and posterior probability of the last iteration; Step 6, set the range and decay rate of the epsilon value according to the model parameters output in step 5; initialize the epsilon value in the Markov process within the range of the epsilon value; for the epsilon value, initialize the drone position, and record the data structure of the current drone path and total moving distance; Step 7: For each drone, set its search starting point to the current drone position, and search for the drone path and total moving distance.

2. The EM pre-clustering UAV cluster task allocation method based on probability inference as claimed in claim 1 is characterized in that: In step 1, the model parameters include clustering probability, cluster center and covariance matrix.

3. The EM pre-clustering UAV cluster task allocation method based on probability inference as claimed in claim 1 is characterized in that: Step 7 specifically includes the following sub-steps: Step 71, for each drone, set its search starting point to the current drone position. When there is still an unfound target, generate a random number. If the random number is less than the current epsilon value, the drone performs exploration behavior and then executes step 72; If the random number is not less than the current epsilon value, no exploration behavior is performed and step 73 is executed; when there is no unfound target, step 73 is executed; Step 72, the drone moves to the selected target position, calculates the moving distance and accumulates the current moving total distance to update it, updates the search path, removes the found target, and updates the current position of the drone, then gradually reduces the epsilon value according to the preset decay rate within the range of the epsilon value as the current epsilon value, and returns to step 71; Step 73, record the paths and total moving distances of all drones under the current epsilon value.

4. The EM pre-clustering UAV cluster task allocation method based on probability inference as claimed in claim 1 is characterized in that: In step 71, the drone performs the exploration behavior specifically by calculating the distances between all remaining targets and its current position, giving priority to targets within a strict range, expanding the range if there are no targets within the strict range, and finally randomly selecting a target as the next target. If there are targets within the strict range, the drone selects the target that is closest.

Images