Unmanned aerial vehicle cluster target search method, device and equipment for complex dynamic environment and storage medium

Abstract

The invention relates to an unmanned aerial vehicle cluster target search method, device and equipment for a complex dynamic environment and a storage medium, and belongs to the technical field of unmanned aerial vehicle cluster target search. The method comprises the following steps: acquiring a collaboration degree index based on an adaptive surface; constructing a cost model based on the ecological heritage, and dynamically predicting the target signal intensity of the unknown region; an unmanned aerial vehicle cluster multi-target search framework is constructed, and unmanned aerial vehicle cluster position distribution is dynamically adjusted through dual-target optimization; determining an optimal path of the unmanned aerial vehicle based on a cost model and an unmanned aerial vehicle cluster multi-target search framework; and designing an unmanned aerial vehicle cluster obstacle avoidance strategy, and performing real-time obstacle avoidance. Through group intelligent emergence and close cooperative interaction, the adaptive ability of the unmanned aerial vehicle cluster and rapid and accurate positioning of the target in a complex environment can be realized, and the search efficiency of the unmanned aerial vehicle cluster in a complex dynamic environment in which multiple dynamic obstacles, multiple dynamic targets and targets automatically evade and search unmanned aerial vehicles is improved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing and information retrieval technology, and in particular to a method, device, equipment and storage medium for searching for unmanned aerial vehicle (UAV) cluster targets in a complex dynamic environment. Background Art

[0002] With the continuous development of drone technology, drones are increasingly recognized for their ability to efficiently and accurately search unknown areas. The rise of the low-altitude economy has further promoted the application of drones in various industries, particularly in search and rescue missions, agricultural monitoring, and urban planning. Due to the ease and flexibility of drone deployment and scheduling, drones have been widely used in these fields.

[0003] Compared to traditional search methods, drones can quickly and accurately search target areas while reducing reliance on manpower and material resources. Drones can search unknown areas in environments such as caves, forests, and oceans. Drones hold significant potential for searching dynamic, unknown areas, and their applications are expanding with the continuous development and innovation of technology. Many researchers are exploring methods to optimize drone search performance in dynamic environments to meet the needs of various fields. However, in dynamic environments, a single drone may not be able to effectively perform search tasks. Therefore, drone swarm target search technology has become an attractive research topic.

[0004] Swarms of drones can work collaboratively to search an area simultaneously, significantly reducing search time. Therefore, research on drone swarm systems and the design of more efficient drone swarm target search methods that can adapt to more complex and dynamic environments are essential. Swarm intelligence-based algorithms enable drones to analyze and interpret data collected during search missions, improving their target detection accuracy. Drones can also learn from previous search missions, helping them adapt to different environments and improve their performance. Therefore, using swarm intelligence technology to design drone swarm control systems has become an important research topic. The concept of swarm intelligence is derived from the cooperative behavior of biological groups, such as bird flocks, ants, and bees. Because the perception range of a single drone is limited and the search area may be dynamically changing, the entire swarm must move in a coordinated manner when searching for targets in dynamic and unknown environments.

[0005] Currently, common swarm intelligence algorithms use movement strategies such as formation advancement and coordinated movement. Phung et al. studied a particle swarm algorithm using motion coding to improve the efficiency of searching for moving targets. Wang et al. combined the bat algorithm to design an optimization method suitable for drone swarms to detect dynamic intrusion targets in oilfield environments. Duan et al. explored a dynamic discrete pigeon swarm optimization method based on the pigeon swarm algorithm, which aims to optimize the planning of drone swarms to perform search and attack missions. In addition, Zheng et al. proposed a human-machine collaborative strategy to improve the efficiency of capture for the fugitive pursuit problem. Although existing research on drone swarm target search algorithms based on swarm intelligence algorithms has enabled drone swarms to conduct collaborative searches in some static environments or simple dynamic environments, there are still some challenges in related research, such as the inability to achieve simultaneous search for multiple targets, the lack of interference from moving obstacles in the experimental environment, and the static or simple movement trajectories of search targets that do not meet the requirements of real-world search tasks.

[0006] Therefore, in order to solve the problem of UAV swarm target search in complex dynamic environments, it is urgent to realize a method that can perform UAV swarm multi-target search in a dynamic experimental environment with multiple dynamic obstacles, multiple dynamic targets, and targets automatically avoiding the search drones. Summary of the Invention

[0007] The purpose of the present invention is to provide a method, device, equipment and storage medium for drone cluster target search in complex dynamic environments. The method can solve the shortcomings of the existing technology. The method is a drone cluster multi-target search algorithm based on the evolution of plant population distribution, which can solve the problem of low efficiency of drone cluster target search in complex dynamic environments in the existing technology.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] In a first aspect of the present invention, a method for searching for a target in a swarm of drones in a complex dynamic environment is disclosed. The method comprises the following steps:

[0010] S1. Obtain a synergy index based on the adaptability surface, and use signal strength and synergy as parallel optimization targets;

[0011] S2. Build a cost model based on ecological heritage to dynamically predict target signal strength in unknown areas;

[0012] S3. Based on the target signal strength prediction value and coordination index, a multi-target search framework for UAV clusters is constructed, and the position distribution of UAV clusters is dynamically adjusted through dual-objective optimization.

[0013] S4. Determine the optimal path for the UAV based on the cost model and the UAV cluster multi-objective search framework;

[0014] S5. Design obstacle avoidance strategies for drone swarms to perform real-time obstacle avoidance.

[0015] Furthermore, in step S1, the construction of a synergy index based on the adaptability surface, taking signal strength and synergy as parallel optimization targets, includes:

[0016] S11. Construct an objective function based on signal strength, and take signal strength as the first optimization objective.

[0017] The signal strength-based objective function includes an individual signal strength function and a group signal strength function;

[0018] The individual signal strength function is:

[0019] f1(x i ) = signal_strength(x i ) (1);

[0020] The group signal strength function is:

[0021]

[0022] Among them, f1(x i ) represents the signal strength of the position of the i-th UAV, x i represents the position of the i-th drone, NP represents the number of drones in the drone cluster, the signal_strength() function is used to obtain the signal strength, f1(x1,x2,…,x NP ) represents the sum of the signal strengths of each individual drone in the drone group.

[0023] S12. Construct an objective function based on the degree of synergy, and use the degree of synergy as the second optimization goal.

[0024] The objective function based on the degree of coordination is:

[0025]

[0026] Among them, x i represents the position of the i-th UAV, x j represents the position of the jth UAV, NP represents the number of UAVs in the UAV cluster, represents the average distance between the positions of the drone cluster, f2(x1,x2,…,x NP ) represents the variance of the distance between drones and is used to indicate the degree of coordination among drones in a group.

[0027] Furthermore, in step S2, the cost model based on ecological heritage is constructed based on the synergy index to dynamically predict the target signal strength in the unknown area, including:

[0028] S21. Obtain search environment information and construct a dynamically updated training set;

[0029] S22. Design an incremental random forest model, use the incremental random forest model to process the search environment information in batches, and dynamically predict the target signal strength in the unknown area.

[0030] The step S21 specifically includes the following steps:

[0031] S211, obtain the search environment information accumulated in the historical iteration to form the original data set D old ;

[0032] S212, using drone sensors to obtain current search environment information to form incremental data D new ;

[0033] S213, using formula (4) to convert the incremental data D new Compared with the original dataset D old Merge into a total data set D, which is used as the training set:

[0034] D=D new ∪D old (4);

[0035] Among them, D represents the entire data set obtained by the drone sensor under the current search environment information, D new Indicates the newly acquired incremental data under the current search environment information obtained by the drone sensor, D old Represents the historical dataset previously acquired by the drone sensor under the current search environment information.

[0036] The step S22 specifically includes the following steps:

[0037] S221. Design an incremental random forest model.

[0038] S222. Perform incremental training of decision trees on the incremental random forest model, and determine a new decision tree through a double random sampling mechanism and error-driven splitting.

[0039] S223, integrate the newly added decision tree into the original random forest, and based on the incremental data D new Training new decision trees to expand the random forest; the original random forest old Contains T old A decision tree is added based on incremental data training to obtain Tnew The newly added decision trees are integrated into the original random forest using formula (8), and the random forest is expanded to obtain the expanded random forest Forest. new :

[0040]

[0041] Among them, Forest new Represents the expanded random forest, Forest old represents the historical random forest, T new Indicates the number of newly added decision trees, Tree k represents the kth new decision tree.

[0042] S224. Based on the expanded random forest, the original decision tree and the newly added decision tree are merged, and the predicted value of the input feature x is updated to the average value of all the merged trees using formula (9) to obtain the integrated signal strength prediction value

[0043]

[0044] in, Represents the signal strength prediction value, T new Indicates the number of newly added decision trees, T old Represents the number of decision trees in the historical random forest, Tree t Represents the tth new decision tree.

[0045] S225, according to the incremental data D new , using formula (10) to calculate the importance of features and identify the features that contribute most to signal strength prediction, so as to dynamically adjust the model focus:

[0046]

[0047] Among them, Importance(j) represents the importance of the j-th feature, ∑ split on j △MES represents the sum of all MES improvements based on feature j, Total Reduction in MES represents the sum of MES improvements brought by all features, and MES represents the mean square error of the current sample set.

[0048] S226, using formula (11) to calculate the incremental data D new Make predictions and calculate the generalization error for the newly added data. Adjust the number or depth of trees based on the generalization error to modify the incremental random forest model:

[0049]

[0050] Among them, ε new represents the generalization error of the newly added data, y i Indicates the true value of signal strength, represents the predicted value of signal strength, Express Perform mathematical expectation operation.

[0051] The step S222 specifically includes the following steps:

[0052] S2221, through Bootstrap sampling, from the incremental data D new The random sample subset S is formed by sampling.

[0053] S2222, at each split point, randomly select a feature subset from the total feature set F of the random sample subset S The feature subset The size m is determined by formula (5):

[0054]

[0055] Among them, m is the size of the feature subset randomly selected at each split, that is, the number of features; the total feature set F refers to the set of all available features in the training data.

[0056] S2223, at each decision tree node, based on the sample subset S and feature subset F generated in step S2221 rand By calculating the mean square error MES and weighted error Loss, the features and their thresholds that minimize the prediction errors of the left and right child nodes are selected as split points, the left and right subtrees are generated, and the newly added decision tree is determined.

[0057] The mean square error MES is calculated using formula (6):

[0058]

[0059] Among them, MES is used to measure the prediction error of the sample subset S, S represents the prediction error from the incremental data D new Sampling forms a random sample subset, |S| represents the number of samples in the sample set S, y i Indicates the true value of signal strength, Indicates the predicted signal strength value.

[0060] The weighted error Loss is calculated using formula (7):

[0061]

[0062] Among them, formula (7) is used to minimize the weighted error. The split point is selected to minimize the sum of the weighted MSEMES of the left and right child nodes after the split. Loss(S) represents the weighted loss function after the split point is selected. S represents the weighted loss function from the incremental data D new Sampling forms a random sample subset, |S| represents the number of samples in the sample set S, S L is the sample set of child nodes of the left subtree after splitting, S R is the sample set of child nodes of the right subtree after splitting, MES(S L ) represents the mean square error of the left subset, MES(S R ) represents the mean squared error of the right subset.

[0063] Furthermore, the step S3, based on the target signal strength prediction value and the coordination index, constructs a multi-target search framework for the drone cluster and dynamically adjusts the position distribution of the drone cluster through dual-objective optimization, specifically including the following steps:

[0064] S31. Initialize the location set X of the drone cluster t , X t As the initial candidate solution for multi-objective optimization; initialize the UAV cluster position set The location of each drone is a randomly generated point in the search space Ω, Expressed as:

[0065]

[0066] Among them, Ω represents the UAV cluster search space, and NP represents the number of UAVs in the UAV cluster.

[0067] S32, based on a certain UAV cluster target search algorithm, from the current position set X t Generate NP*M f Next-generation positions, forming the next-generation drone cluster position set X t+1 :

[0068]

[0069] Among them, NP represents the number of drones in the drone cluster, and the parameter M f Determines the number of next generation drones to be generated, M f The larger it is, the more locations are generated. Indicates the location of the next generation drone No. i.

[0070] S33, based on the next generation of drone cluster position collection X t+1 , combined generation distribution patterns, forming a distribution pattern set; among them, It represents the number of all possible drone cluster location distribution schemes combined from all generated candidate locations.

[0071] S34. Based on the distribution pattern set and the cost model, the signal strength of each distribution pattern is calculated using formula (14) to obtain a signal strength value set for each distribution pattern:

[0072] f1(y1,y2,…,y NP )=max(CM(y1),CM(y2),…,CM(y NP )) (14);

[0073] Among them, CM is the current cost model; (y1,y2,…,y NP ) represents the location set of the drone cluster in the current distribution pattern, f1(y1,y2,…,y NP ) represents the predicted value of the strongest signal strength in the current distribution pattern.

[0074] S35. Based on the distribution pattern set, the synergy degree of each distribution pattern is calculated using formula (15) to obtain a synergy degree value set for each distribution pattern;

[0075]

[0076] Among them, NP represents the number of drones in the drone cluster, represents the average distance of the drone cluster position, (y1,y2,…,y NP ) represents the location set of the drone cluster in the current distribution pattern, f2(y1,y2,…,y NP ) represents the variance prediction value of the distance between drones in the current distribution pattern, which is used to indicate the degree of coordination of the group of drones.

[0077] S36. Compare the cost model signal strength f1 and the coordination degree f2 of all distribution patterns, use the non-dominated sorting method to screen out the non-dominated solution as the optimal solution, and form a non-dominated solution set.

[0078] S37. In the search for non-dominated solutions, randomly select a distribution pattern as the location set X of the next generation UAV cluster. t+1 , determine the final distribution pattern.

[0079] Furthermore, in step S4, determining the optimal path of the drone based on the cost model and the drone cluster multi-objective search framework specifically includes the following steps:

[0080] S41. Design a cost function based on the movement distance of the drone cluster and the time required to complete the task;

[0081] The cost function is shown in formula (16):

[0082] F d =min{a*γ+(1-a)*Δ} (16);

[0083] in,

[0084]

[0085] Tspan(t)=mat τ i (t),1≤i≤n (20);

[0086] Among them, F d represents the cost function based on the moving distance of the drone cluster and the time required to complete the task, a represents the weight parameter, a∈[0,1], which is used to balance the optimization weight of the two costs of total moving distance and maximum time spent in the drone cluster path planning; γ represents the total moving distance of the drone cluster after completing the entire task; Δ represents the maximum moving time of the drone cluster after completing the task; Z(t) represents the cumulative moving distance of all drones in the tth generation; represents the moving distance of the i-th UAV in the t-th generation; Tspan(t) represents the maximum moving time spent by the t-th generation UAV cluster; τ i (t) represents the specific moving time of the i-th UAV of the t-th generation.

[0087] S42: Initialize the state of the drone cluster, calculate the comprehensive cost of each drone moving to each target location, and determine the optimal path for the drone based on the comprehensive cost.

[0088] S43, record the maximum distance extreme value parameter γ between the UAV and the target position calculated in the current generation max and the maximum time consumption parameter Δ for the UAV to complete the task max , for normalization use when calculating the comprehensive cost matrix in the next generation, and as the benchmark parameter for normalizing distance and time costs when calculating the comprehensive cost matrix in the next generation.

[0089] Furthermore, the step S42 specifically includes the following steps:

[0090] S421, initialize the time vector τ of the tth generation [1×m] (t) = {0,…,0}, where m represents the number of drones and each element corresponds to the cumulative movement time of a drone.

[0091] S422, based on the time vector, normalize the distance cost μ ij and normalized time cost θ ij .

[0092] The normalized distance cost μij Using formula (21), we can calculate:

[0093] μ ij =D(S i (t),X j (t+1)) / γ max (twenty one);

[0094] The normalized time cost θ ij Using formula (22), we can calculate:

[0095] θ ij =(τ i +D(S i (t),X j (t+1)) / v) / Δ max (twenty two);

[0096] Among them, D(S i (t),X j (t+1)) represents the Euclidean distance between the current position and the target position, using Calculated; γ max Indicates the maximum distance between the previous generation of drones and the target location, used for normalization to avoid the distance dimension affecting the cost comparison; τ i represents the cumulative time of the i-th UAV in the previous generation of the current iteration, which is initially 0; v represents the average moving speed of the UAV; Δ max It represents the maximum time consumption of the previous generation of drones to complete the task, which is used to normalize the time cost.

[0097] S423, according to the normalized distance cost μ ij and normalized time cost θ ij , calculate the comprehensive cost F(i,j) required for each drone in the drone cluster to move to the next target point, and form the cost matrix F [m×n] The cost matrix F [m×n] Each element F(i,j) in represents the i-th drone moving to the next target point X j The comprehensive cost required for (t+1).

[0098] The comprehensive cost F(i,j) is calculated using formula (23):

[0099] F(i,j)=a*μ ij +(1-a)*θ ij (twenty three);

[0100] Among them, a represents the weight parameter, a∈[0,1], which is used to balance the optimization weight of the total moving distance and the maximum time spent in the path planning of the UAV cluster; μij represents the normalized distance cost of the i-th UAV moving to the j-th target point; θ ij represents the normalized time cost of the i-th UAV moving to the j-th target point; F(i,j) represents the comprehensive cost of the i-th UAV moving to the j-th target point.

[0101] Furthermore, in step S5, the design of the drone cluster obstacle avoidance strategy to perform real-time obstacle avoidance specifically includes the following steps:

[0102] S51. Define safety zone: Define a safety zone with a radius of d around each drone to monitor potential collision and obstacle threats in real time.

[0103] S52: Establish a virtual force field. Based on the virtual force field, the UAV performs real-time obstacle avoidance. When other UAVs or obstacles enter the safe area, a repulsive force is generated on UAV i. The repulsive force function is as follows:

[0104]

[0105] Among them, d ij represents the distance between drone i and obstacle / drone j; Represents the repulsive force scale factor, which controls the repulsive force strength. Unit: N·m 2 ;U repij =((i, j) / (i)) / (i) indicates that drone i is repelled by obstacle / drone j. The repulsive force is directed from the obstacle / drone towards drone i, forcing drones away from the threat source. The real-time obstacle avoidance results influence the cluster distribution pattern, indirectly affecting the multi-objective optimization in step S3.

[0106] S53. Use formula (25) to update the position of the drone after being affected by the repulsive force:

[0107] x_new=x+U repi (25);

[0108] Among them, x represents the position of the UAV before being affected by the repulsive force, x_new represents the position of the UAV after being affected by the repulsive force, and U repi is the repulsive force on drone i. The adjusted position is fed back to step S4, which may trigger path replanning.

[0109] In a second aspect of the present invention, a drone cluster target search device for complex dynamic environments is disclosed. The device includes a coordination index acquisition module, a target signal strength prediction module, a drone cluster position distribution adjustment module, a drone optimal path determination module, and an obstacle avoidance module.

[0110] The synergy index acquisition module is used to obtain the synergy index based on the adaptability surface, and takes signal strength and synergy as parallel optimization targets; the target signal strength prediction module is used to construct a cost model based on ecological heritage and dynamically predict the target signal strength in unknown areas; the drone cluster position distribution adjustment module is used to construct a drone cluster multi-target search framework based on the target signal strength prediction value and the synergy index, and dynamically adjust the drone cluster position distribution through dual-target optimization; the drone optimal path determination module is used to determine the drone optimal path based on the cost model and the drone cluster multi-target search framework; the obstacle avoidance module is used to design a drone cluster obstacle avoidance strategy and perform real-time obstacle avoidance.

[0111] In a third aspect of the present invention, an electronic device is disclosed, comprising: at least one processor; and a memory, wherein the memory stores instructions, which, when executed by the at least one processor, enable the at least one processor to execute the above-mentioned drone cluster target search method for complex dynamic environments.

[0112] In a fourth aspect of the present invention, a machine-readable storage medium is disclosed, which stores executable instructions, which, when executed, enable the machine to execute the above-mentioned drone cluster target search method for complex dynamic environments.

[0113] Compared with the prior art, the advantages of the present invention are:

[0114] (1) The present invention solves the limitation of low efficiency of drone cluster search in dynamic environments in existing technologies through the emergence of group intelligence and close collaborative interaction, and can realize the adaptive ability of drone clusters and rapid and accurate positioning of targets in complex environments. The present invention realizes multi-target search of drone clusters based on the adaptive surface of plant population distribution evolution and ecological heritage. This method solves the target search task in unknown complex dynamic environments by constructing a coordination index based on the adaptive surface and a cost model based on the ecological heritage, combined with a drone cluster target search algorithm based on the law of plant population distribution evolution, thereby improving the performance and efficiency of drone cluster search targets and enhancing the benefits of drone cluster collaborative search.

[0115] (2) The present invention constructs an experimental scenario of an unknown complex dynamic environment containing moving targets and obstacles, and evaluates and analyzes the target search method of the drone cluster. In the simulation scenario of the unknown complex dynamic environment, the method proposed by the present invention has a fast search speed, high stability and good robustness. The construction of the coordination index based on the adaptability surface involved in the present invention takes into account the search efficiency of the drone cluster and the collaborative search capability of the drone cluster. The construction of the cost model based on ecological heritage increases the exploration range while effectively reducing the search time and energy consumption of the drone cluster. The combination with the drone cluster target search algorithm based on the evolution law of plant population distribution enhances the diversity of drone cluster search, which is more suitable for performing tasks in unknown complex dynamic environments. The drone optimal path selection strategy and drone cluster obstacle avoidance strategy proposed by the present invention effectively reduce the energy consumption of the drone cluster and ensure that the drone cluster completes the task safely. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 It is a flow chart of the method for searching for target in a swarm of drones in a complex dynamic environment according to the present invention;

[0117] Figure 2 Schematic diagram of the contradiction between the coordination index and the signal strength index;

[0118] Figure 3 Schematic diagram of the process of fitting the incremental random forest model;

[0119] Figure 4 This is a schematic diagram of the UAV's mobile path planning, where the coordinate axes represent the geographic coordinate system of the two-dimensional plane space;

[0120] Figure 5 This is a schematic diagram of the drone's repulsive obstacle avoidance;

[0121] Figure 6 A top-down view of the main elements in the scene for the drone target search series;

[0122] Figure 7a Searching for initial state graphs for drone swarms in three target scenarios;

[0123] Figure 7b The first target state diagram was discovered for the drone swarm search in three target scenarios;

[0124] Figure 7c State diagram of the first target search and destruction for a drone swarm in three target scenarios;

[0125] Figure 7d A second target state diagram was discovered for the three-target scenario drone swarm search;

[0126] Figure 7eState diagram of the drone swarm searching and destroying the second target for a three-target scenario;

[0127] Figure 7f State diagram of a drone swarm searching for three targets and attempting to destroy the third target. DETAILED DESCRIPTION

[0128] The present disclosure is further described below with reference to the accompanying drawings and embodiments:

[0129] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs.

[0130] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof. In the absence of conflict, the embodiments in this disclosure and the features in the embodiments can be combined with each other.

[0131] Example 1

[0132] This embodiment provides a method for searching for drone cluster targets in complex dynamic environments. Figure 1 As shown, the method includes the following steps:

[0133] S1. Obtain a synergy index based on the adaptability surface, and use signal strength and synergy as parallel optimization targets.

[0134] In order to adapt to the complex dynamic environment, the collaborative working ability of the drone cluster is abstracted as the "adaptability surface" by drawing on the evolution law of plant population distribution, and the rationality of the group distribution is quantified through mathematical modeling. By analyzing the contradiction between the exploration of high signal intensity areas by individual drones and the collaborative distribution of the group, a coordination index based on the adaptability surface is constructed, and the signal strength and coordination are used as parallel optimization targets to achieve multi-objective optimization of the group drones, while ensuring high group coordination and high individual signal strength. In this embodiment, the signal strength is used as the first target of the target search, and the collaborative working ability of the drone cluster distribution pattern - coordination, is used as the second target of the target search. By quantifying the contradiction between the individual and the group and constructing a dual-objective function of signal strength and coordination, the limitations of a single target can be avoided, and the overall efficiency of the cluster can be improved, thereby achieving efficient search of drone clusters in complex dynamic environments.

[0135] like Figure 2 As shown in the figure, there is an inherent contradiction between the coordination index and signal strength index of the drone cluster. This contradiction essentially stems from the conflict between the individual behavior of drones and the coordination needs of the group. Specifically, the individual drones are driven by the signal strength objective function and tend to gather at locations with higher signal strength to improve the local target search accuracy and achieve individual optimization, which may lead to dispersed cluster distribution and decreased coordination; while the cluster as a whole tends to maintain a distribution pattern with higher coordination and needs to maintain a uniform distribution pattern through the coordination objective function to expand the global search coverage and achieve group optimization, which may sacrifice the signal strength of some individuals.

[0136] To eliminate the above contradictions, the present invention constructs a dual-objective optimization model, taking signal strength and coordination as parallel optimization objectives, while maximizing signal strength (individual efficiency) and minimizing the variance of drone spacing (group coordination). Through a multi-objective algorithm, a distribution pattern that simultaneously satisfies high signal strength and high coordination is screened, thereby achieving a dynamic balance between individuals and groups, and a dynamic balance between "local exploration" and "global coverage." Through the dual-objective balance, the present invention enables drones to maintain group collaboration while exploring high-value areas, making it suitable for complex scenarios such as dynamic obstacles and multi-target escape.

[0137] Furthermore, in step S1, the construction of the synergy index based on the adaptability surface, taking signal strength and synergy as parallel optimization targets, specifically includes the following steps:

[0138] S11. Construct an objective function based on signal strength, and take signal strength as the first optimization objective.

[0139] The signal strength-based objective function includes an individual signal strength function and a group signal strength function.

[0140] The individual signal strength function is:

[0141] f1(x i ) = signal_strength(x i ) (1);

[0142] The group signal strength function is:

[0143]

[0144] Among them, f1(x i ) represents the signal strength of the position of the i-th UAV, x i represents the position of the i-th drone, NP represents the number of drones in the drone cluster, the signal_strength() function is used to obtain the signal strength, f1(x1,x2,…,x NP) represents the sum of the signal strengths of each individual drone in the drone group.

[0145] The individual signal strength function quantifies a single drone's ability to perceive target signals; the group signal strength function optimizes by maximizing the sum of signal strengths, guiding drone clusters toward high-probability target areas. The signal strength-based objective function corresponds to the fitness function in intelligent optimization algorithms, optimizing for maximizing signal strength at the drone's location. By quantifying individual drones' responsiveness to target signals, it guides drones toward high-probability areas, improving search accuracy. This objective function is suitable for prioritizing exploration of high-probability areas during dynamic target search, preventing drones from missing critical target signals.

[0146] S12. Construct an objective function based on the degree of synergy, and use the degree of synergy as the second optimization goal.

[0147] The coordination objective represents the collaborative working ability of the drone swarm distribution pattern. It is specific to the drone cluster and is unaffected by changes in the search environment. This paper defines coordination as the variance of the distances between drones. The goal is to minimize the distance deviations between drones, ensuring harmonious flight within the swarm. By minimizing the variance, drones are forced to maintain appropriate spacing, avoiding excessive concentration or dispersion, and improving global coverage efficiency.

[0148] The objective function based on the degree of coordination is:

[0149]

[0150] Among them, x i represents the position of the i-th UAV, x j represents the position of the jth UAV, NP represents the number of UAVs in the UAV cluster, represents the average distance between the positions of the drone cluster, f2(x1,x2,…,x NP ) is the variance of the distance between drones, which is used to indicate the degree of coordination among the drones in the group.

[0151] The objective function based on coordination degree is suitable for avoiding the "clustering" or "dispersion" of drones in complex dynamic environments, balancing individual exploration and group collaboration, and preventing global search blind spots caused by local concentration.

[0152] S2. Construct a cost model based on ecological heritage to dynamically predict the target signal strength in unknown areas.

[0153] The cost model is used to build a model of the search environment based on previous mission data. This model is used to approximate target signal strength in unknown areas, eliminating the need for drones to fly directly into those areas for acquisition. This cost model increases the exploration range while effectively reducing the search time and energy consumption of the drone swarm. However, designing a cost model is challenging; an inappropriate model can severely impact the search efficiency of the drone swarm.

[0154] Based on the random forest model, this embodiment designs an incremental random forest model in combination with the concept of ecological heritage. Figure 3 As shown in the data batch input and decision tree incremental construction mechanism in , this embodiment quantifies the cost of environmental exploration through a cost model to achieve adaptive optimization of drone cluster target search tasks in complex dynamic environments. A cost model is used to handle drone target search tasks in complex dynamic environments. Specifically, the cost model dynamically updates the forest structure by reusing the historical data pool (corresponding to ecological heritage) and fusing it with real-time incremental data, thereby predicting the target signal strength in unknown areas at a low computational cost, guiding drones to prioritize exploring high-probability areas, and reducing time and energy consumption during the search process.

[0155] Furthermore, in step S2, a cost model based on ecological heritage is constructed to dynamically predict the target signal strength in the unknown area, specifically including the following steps:

[0156] S21. Obtain search environment information and construct a dynamically updated training set;

[0157] S22. Design an incremental random forest model, use the incremental random forest model to process the search environment information in batches, and dynamically predict the target signal strength in the unknown area.

[0158] Furthermore, the step S21 specifically includes the following steps:

[0159] S211, obtain the search environment information accumulated in the historical iteration to form the original data set D old ;

[0160] S212, using drone sensors to obtain current search environment information to form incremental data D new ;

[0161] S213, using formula (4) to convert the incremental data D new Compared with the original dataset D old Merge into a total data set D, which is used as the training set:

[0162] D=D new ∪D old (4);

[0163] Among them, D represents the entire data set obtained by the drone sensor under the current search environment information, D new Indicates the newly acquired incremental data under the current search environment information obtained by the drone sensor, D old Represents the historical dataset previously acquired by the drone sensor under the current search environment information.

[0164] The cost model adopts a batch learning method. Each iteration of the target search algorithm uses the drone sensor to obtain the current search environment information to form a new incremental data D new The incremental data includes: target signal strength measurement values ​​(such as RF signal and infrared radiation intensity), environmental characteristics (such as terrain elevation and obstacle distribution), and spatiotemporal context information (collection timestamp and geographic location coordinates). old The environmental information accumulated in historical iterations, namely ecological heritage data, historical dataset D old The data pool mechanism is used for persistent storage. new and the original dataset D old The merge operation preserves the temporal and spatial indexes of the data, supporting temporal correlation and feature alignment of incremental data. By merging new and old data to build a dynamically updated training set, retraining on the entire data set can be avoided.

[0165] Furthermore, the step S22 specifically includes the following steps:

[0166] S221. Design an incremental random forest model.

[0167] Traditional random forests need to retrain all decision trees when new data is added, while the incremental random forest model retains the D old The corresponding tree structure is only for incremental data D new Generate a new tree. Each time only for the newly added data D new Update the cost model to avoid rebuilding the entire forest.

[0168] S222. Perform incremental training of decision trees on the incremental random forest model, and determine a new decision tree through a double random sampling mechanism and error-driven splitting.

[0169] For each batch of incremental data D new, sample subsets and feature subsets need to be randomly selected when building a new decision tree. In the decision tree and random forest algorithms, the split point refers to the features and their values ​​selected when dividing the samples of the current node during the growth of the decision tree. The split point determines how the samples are assigned to the left subtree or the right subtree, and is a key step for the decision tree to achieve feature space division. The decision tree gradually divides the high-dimensional feature space into low-dimensional subspaces by selecting features from the total feature set F and their split points, thereby realizing the prediction of the target variable (such as signal strength). In the random forest, by randomly sampling the total feature set F and randomly selecting feature subsets, randomness can be introduced to avoid overfitting of a single tree.

[0170] Furthermore, the step S222 specifically includes the following steps:

[0171] S2221, through Bootstrap sampling, from the incremental data D new The random sample subset S is formed by sampling.

[0172] Step S2221 introduces randomness at the sample level, which can avoid overfitting of the model and provide a training data basis for subsequent decision tree splitting. The number of samples in the random sample subset S is the same as D new Same, but allows for sample duplication.

[0173] S2222, at each split point, randomly select a feature subset from the total feature set F of the random sample subset S The feature subset The size m is determined by formula (5):

[0174]

[0175] Where m is the size of the feature subset randomly selected at each split, that is, the number of features; the total feature set F refers to the set of all available features in the training data. In the drone search scenario, the total feature set F includes: environmental features, signal features, and drone status, etc.

[0176] By randomly selecting feature subsets and introducing randomness at the feature level, we can avoid the decision tree's dependence on specific features and improve the model's generalization ability. According to the dimension of the total feature set F, m features are randomly selected from the total feature set F without replacement to form a feature subset. This is used for subsequent split point calculations. Step S2221 introduces randomness at the sample level through Bootstrap sampling, and step S2222 further introduces randomness at the feature level, forming a dual random mechanism that jointly improves the diversity of the decision tree. Different trees can use different combinations of sample subsets and feature subsets.

[0177] S2223, at each decision tree node, based on the sample subset S and feature subset F generated in step S2221 rand By calculating the mean square error MES and weighted error Loss, the features and their thresholds that minimize the prediction errors of the left and right child nodes are selected as split points, thereby generating left and right subtrees and determining the newly added decision tree.

[0178] The mean square error MES is calculated using formula (6):

[0179]

[0180] Among them, MES(S) is used to measure the prediction error of the sample subset S; S represents the prediction error from the incremental data D new Sampling forms a random sample subset, |S| represents the number of samples in the sample set S, y i Indicates the true value of signal strength, Indicates the predicted signal strength value.

[0181] The weighted error Loss is calculated using formula (7):

[0182]

[0183] Among them, formula (7) is used to minimize the weighted error, and the split point is selected to minimize the sum of the weighted MSE of the left and right child nodes after the split; Loss (S) represents the weighted loss function after the split point is selected; S represents the weighted loss function from the incremental data D new Sampling forms a random sample subset, |S| represents the number of samples in the sample set S, S L is the sample set of child nodes of the left subtree after splitting, S R is the sample set of child nodes of the right subtree after splitting, MES(S L ) represents the mean square error of the left subset, MES(S R ) represents the mean squared error of the right subset.

[0184] Step S2223 is the core optimization link of the random forest algorithm. By combining feature random sampling with error-driven splitting, it improves the generalization ability of the model while ensuring rapid response to dynamic environments (such as signal fluctuations in drone search scenarios). By finding the optimal split point, it can ensure better prediction accuracy of the left and right subtrees, making the decision tree more responsive to incremental data D. new The signal strength prediction error is minimized, and the model dynamically adapts to environmental changes (such as signal fluctuations and obstacle movement) to improve the local fitting accuracy of the model. At each decision tree node, the feature subset F generated in step S2222 is traversed. rand Among all possible splitting thresholds in , the feature and threshold that minimize the sum of the weighted mean square error (MES) of the left and right child nodes after splitting are selected as the optimal splitting point.

[0185] S223, integrate the newly added decision tree into the original random forest, and based on the incremental data D new Train additional decision trees to extend the random forest.

[0186] Original Random Forest old Contains T old A decision tree is added based on incremental data training to obtain T new The newly added decision trees are integrated into the original random forest using formula (8), and the random forest is expanded to obtain the expanded random forest Forest. new In step S223, the historical tree structure is retained to reuse the ecological heritage data.

[0187]

[0188] Among them, Forest new Represents the expanded random forest, Forest old represents the historical random forest, T new Indicates the number of newly added decision trees, Tree k represents the kth new decision tree.

[0189] Step S223 is used to incrementally update the random forest of the cost model. The training process of the newly added decision tree is the same as the original tree (the training process is as steps S2221 to S2223), but only the incremental data D is used. new Build to avoid retraining the full data and only capture incremental data D by adding new trees new The history tree preserves ecological heritage knowledge and achieves a balance between reusing old knowledge and learning new knowledge.

[0190] S224. Based on the expanded random forest, the original decision tree and the newly added decision tree are merged, and the predicted value of the input feature x is updated to the average value of all the merged trees using formula (9) to obtain the integrated signal strength prediction value

[0191]

[0192] in, Represents the signal strength prediction value, T new Indicates the number of newly added decision trees, T old Represents the number of decision trees in the historical random forest, Tree t Represents the tth new decision tree.

[0193] Step S224 is used to combine the prediction results of the new and old decision trees and generate the final signal strength prediction value by weighted average. That is, step S22 needs to dynamically predict the target signal strength in the unknown area to ensure the synergistic contribution of historical experience and real-time data. In this step, the prediction values ​​of all trees are averaged to suppress the overfitting risk of a single tree. The integrated signal strength prediction value Used for signal strength evaluation of cost model. Through ensemble learning, the variance of a single model can be reduced, the prediction stability can be improved, and the prediction value of the newly added tree can quickly reflect the incremental data D new The signal fluctuations in the sensor can dynamically respond to environmental changes.

[0194] S225, according to the incremental data D new ,The importance of features is calculated using Equation (10) to identify the features that contribute most to the signal strength prediction and dynamically adjust the model focus.

[0195]

[0196] Among them, Importance(j) represents the importance of the j-th feature, ∑ split on j △MES represents the sum of all MES improvements based on feature j, and Total Reduction in MES represents the sum of MES improvements brought by all features. MES represents the mean squared error of the current sample set.

[0197] Due to the addition of data D new The split contribution of the feature may change, so the feature importance needs to be recalculated. In formula (10), ΔMES is the error reduction when feature j is split; the numerator is the sum of the error reductions of feature j at all split points, and the denominator is the error reduction of the entire forest. Through step S225, not only can redundant features be filtered and features irrelevant to the current environment be suppressed, such as the importance of stable terrain features in dynamic target scenes may decrease; it can also adapt to feature drift. When the dominant factor of the environment changes, such as from terrain dominance to signal strength dominance, the model automatically adjusts the splitting strategy. For example, if D new The Importance(j) of the “dynamic obstacle position” feature is significantly improved, indicating that this feature is more critical to the current signal prediction and is preferred in subsequent splitting.

[0198] S226, using formula (11) to calculate the incremental data D new Make predictions and calculate the generalization error for the newly added data. Adjust the number or depth of trees based on the generalization error to modify the incremental random forest model:

[0199]

[0200] Among them, ε new represents the generalization error of the newly added data, y i Indicates the true value of signal strength, represents the predicted value of signal strength, Express Perform mathematical expectation operation.

[0201] Step S226 is used to evaluate the dynamic generalization error of the cost model and modify the cost model according to the error evaluation result. new Make predictions and calculate the generalization error. If the calculated generalization error exceeds the limit, the random forest structure needs to be adjusted to ensure model accuracy. If the error exceeds the threshold, the number of trees T can be increased. new Or increase the depth of each tree to improve model performance. Increasing the number of new trees can increase model complexity. Increasing the maximum depth of a single tree can allow for more complex feature combination splits. Step S226 drives model adaptive optimization through error evaluation, which can avoid blind growth of random forests. When underfitting (high error), structural adjustments are made to improve fitting ability and prevent overfitting.

[0202] S3. Based on the target signal strength prediction value and coordination index, a multi-target search framework for drone clusters is constructed, and the position distribution of drone clusters is dynamically adjusted through dual-objective optimization.

[0203] By initializing, generating, evaluating and selecting the multi-objective algorithm for the drone cluster, the location distribution of the drone cluster is gradually optimized.

[0204] Furthermore, the step S3, based on the target signal strength prediction value and the coordination index, constructs a multi-target search framework for the drone cluster and dynamically adjusts the position distribution of the drone cluster through dual-objective optimization, specifically including the following steps:

[0205] S31. Initialize the location set X of the drone cluster t , X t As the initial candidate solution for multi-objective optimization.

[0206] Initialize the drone cluster position set The location of each drone is a randomly generated point in the search space Ω, It can be expressed as:

[0207]

[0208] Among them, Ω represents the UAV cluster search space, and NP represents the number of UAVs in the UAV cluster.

[0209] Step S31 randomly initializes the drone positions to provide an initial position distribution for the drone cluster search task. Random generation avoids initial position deviation and ensures that the drone cluster can widely explore different areas at the beginning of the search task to achieve search space coverage. The position coordinates obey a uniform distribution within the search space Ω, ensuring that the initial position covers the entire area and avoids bias towards a specific area. In addition, by randomly generating a covered search space, a diverse starting point is provided for subsequent optimization, providing an initial solution for subsequent non-dominated sorting and distribution pattern evaluation, and gradually converging to a position distribution that takes into account both high signal strength and high coordination through iterative optimization, such as Figure 7a The initial dispersed state shown shows the state of the drone cluster taking off.

[0210] S32, based on a certain UAV cluster target search algorithm, from the current position set X t Generate NP*M f Next-generation positions, forming the next-generation drone cluster position set X t+1 :

[0211]

[0212] Among them, NP represents the number of drones in the drone cluster, and the parameter M f Determines the number of next generation drones to be generated, M f The larger it is, the more locations are generated. Indicates the location of the next generation drone No. i.

[0213] In formula (13), the parameter M f Determines the number of next generation drone positions to be generated, M f The larger the value is, the more positions are generated. The generated candidate positions are the basis for evaluating signal strength and coordination, and their distribution quality determines the effectiveness of subsequent non-dominated solutions.

[0214] In step S32, based on the current cluster position, a diverse next-generation position candidate set is generated through an algorithm to provide sufficient solution space for multi-objective optimization. By performing operations such as mutation and crossover on the current position through methods such as particle swarm optimization and genetic algorithm to generate new positions, it can ensure that the candidate positions cover different areas of the search space and avoid premature convergence to the local optimum. In this embodiment, the drone cluster target search algorithm can adopt the existing plant population distribution evolution algorithm. By controlling M f Balance computational complexity and search coverage, provide sufficient candidate solutions for subsequent dual-objective optimization based on signal strength and coordination, and promote iterative optimization of cluster location distribution.

[0215] S33, based on the next generation of drone cluster position collection X t+1, combined generation distribution patterns to form a distribution pattern set. It represents the number of all possible drone cluster location distribution schemes combined from all generated candidate locations.

[0216] In the distribution pattern set, each distribution pattern consists of NP different drone positions. These distribution patterns provide candidate solutions for subsequent selection and optimization. Step S33 generates cluster distribution patterns by combination, converting the continuous position candidate set into a discrete cluster distribution pattern. The diversity of candidate solutions is ensured through mathematical combination, providing sufficient evaluation objects for the subsequent dual-objective optimization based on signal strength and coordination, and ultimately promoting the iterative optimization of cluster position distribution. From NP×Mf candidate positions, NP different positions are selected to form a distribution pattern. Each distribution pattern represents a possible drone cluster position allocation scheme, which contains NP unique positions. By combining and generating a variety of possible cluster distributions, it is ensured that multi-objective optimization can capture different search strategies, such as Figure 7b The pattern of drones gathering in high signal areas coexists with the uniform distribution pattern. Figure 7b In the video, the drone swarm found its first target.

[0217] S34. Based on the distribution pattern set and the cost model, the signal strength of each distribution pattern is calculated using formula (14) to obtain a signal strength value set for each distribution pattern.

[0218] With the distribution pattern set {y1,y2,…,y NP} as an example, the cost model signal strength is defined as the signal strength at the optimal position in the swarm drone:

[0219] f1(y1,y2,…,y NP )=max(CM(y1),CM(y2),…,CM(y NP )) (14);

[0220] Among them, CM is the current cost model, (y1,y2,…,y NP ) represents the location set of the drone cluster in the current distribution pattern, f1(y1,y2,…,y NP ) represents the predicted value of the strongest signal strength in the current distribution pattern.

[0221] In formula (14), CM(yi) is a cost model constructed by incremental random forest, which is used to predict the signal strength at position y; the signal strength at the optimal position in the cluster is taken as the evaluation value of the entire pattern, reflecting the cluster's coverage capability of high-value areas. Step S34 is to convert environmental cognition (the cost model in step S2) into specific decision-making indicators, which together with the coordination index drive the adaptive search behavior of the drone cluster in a complex environment. Specifically, the cost model is used to convert the area where the target may exist into a quantitative indicator, guiding the cluster to move to the area with high signal strength, such as Figure 7b There is a trend of drones gathering in hot spots.

[0222] S35. Based on the distribution pattern set, calculate the coordination degree of each distribution pattern to obtain a coordination degree value set of each distribution pattern.

[0223] With the distribution pattern set {y1,y2,…,y NP} as an example, its coordination degree is defined as:

[0224]

[0225] Among them, NP represents the number of drones in the drone cluster, d represents the average distance between the positions of the drone cluster, (y1,y2,…,y NP ) represents the location set of the drone cluster in the current distribution pattern, f2(y1,y2,…,y NP ) represents the variance prediction value of the distance between drones in the current distribution pattern, which is used to indicate the degree of coordination of the group of drones.

[0226] S36. Compare the cost model signal strength f1 and coordination degree f2 of all distribution patterns, and use the non-dominated sorting method to filter out the non-dominated solution as the optimal solution to form a non-dominated solution set. For each pattern Yi, check whether there is another pattern Yj that dominates it; if there is no Yj < Yi, then Yi is retained as a non-dominated solution. The patterns in the non-dominated solution set achieve the optimal trade-off between signal strength and coordination degree, such as Figure 7c In the various equilibrium states from dispersion to aggregation, Figure 7c In [1], the drone swarm searches for the second target. The initial K solutions are compressed into a more streamlined set of non-dominated solutions, reducing the complexity of subsequent decisions.

[0227] S37. In the search for non-dominated solutions, randomly select a distribution pattern as the location set X of the next generation UAV cluster. t+1 , determine the final distribution pattern. The updated position set in step S37 is used to drive the actual movement of the drone cluster, such as Figure 7d Dynamically adjust the medium cluster to the high signal area. Select an execution plan from the optimal solution set to iterate the cluster position towards a more optimal distribution.

[0228] S4. Determine the optimal path for the drone based on the cost model and the drone cluster multi-objective search framework.

[0229] Furthermore, in step S4, determining the optimal path of the drone based on the cost model and the drone cluster multi-objective search framework specifically includes the following steps:

[0230] Based on the moving distance and moving time of the drone as the key costs, a cost minimization goal is designed based on this, and the moving path of each drone in the cluster is optimized to ensure that the total cost from the current position to the next target position is minimized.

[0231] S41. Design a cost function based on the movement distance of the drone cluster and the time required to complete the task.

[0232] The cost function is shown in equations (16) to (20):

[0233] F d =min{a*γ+(1-a)*Δ} (16);

[0234]

[0235] Tspan(t)=max τ i (t),1≤i≤n (20);

[0236] Among them, F d represents a cost function based on the movement distance of the drone cluster and the time required to complete the task. a represents a weight parameter, a∈[0,1], which is used to optimize the weight to balance the total movement distance and maximum time required in the drone cluster path planning. Specifically, γ represents the total movement distance of the drone cluster after completing the entire task; Δ represents the maximum movement time of the drone cluster after completing the task; Z(t) is the cumulative movement distance of all drones in the tth generation; represents the moving distance of the i-th UAV in the t-th generation; Tspan(t) is the maximum moving time spent by the t-th generation UAV cluster; τ i (t) represents the specific moving time of the i-th UAV of the t-th generation.

[0237] In step S41, the cost function is a cost function that minimizes the distance the drone moves and the time required to complete the task. In a drone swarm target search task, individual drones typically need to move from their current position X(t) to the next target position X(t+1). However, the optimization method proposed in this invention considers the drone's movement distance and movement time as key costs and designs a cost minimization objective based on these costs. Figure 4A drone scheduling method based on the drone's moving distance and moving time as key costs is demonstrated, which effectively improves the multi-drone path coordination effect and reduces the drone's flight distance.

[0238] S42: Initialize the state of the drone cluster, calculate the comprehensive cost of each drone moving to each target location, and determine the optimal path for the drone based on the comprehensive cost.

[0239] Step S42 transforms the drone path planning problem into a structured optimization problem through state initialization and cost matrix calculation. Multi-dimensional cost quantification and historical parameter reuse provide a scientific basis for subsequent path optimization, ensuring that the drone swarm can move at minimal cost in complex environments. This is the key computational link between swarm position optimization and actual execution.

[0240] Furthermore, the step S42 specifically includes the following steps:

[0241] S421, initialize the time vector τ of the tth generation [1×m] (t) = {0,…,0}, where m represents the number of drones and each element corresponds to the cumulative movement time of a drone.

[0242] S422, based on the time vector, normalize the distance cost μ ij and normalized time cost θ ij .

[0243] The normalized distance cost μ ij Using formula (21), we can calculate:

[0244] μ ij =D(S i (t),X j (t+1)) / γ max (twenty one);

[0245] The normalized time cost θ ij Using formula (22), we can calculate:

[0246] θ ij =(τ i +D(S i (t),X j (t+1)) / v) / Δ max (twenty two);

[0247] Among them, D(S i (t),X j (t+1)) represents the Euclidean distance between the current position and the target position, using Calculated; γ maxIndicates the maximum distance between the previous generation of drones and the target location, used for normalization to avoid the distance dimension affecting the cost comparison; τ i represents the cumulative time of the i-th UAV in the current iteration, which is initially 0; x represents the average moving speed of the UAV; Δ x Indicates the maximum time consumption of the previous generation of drones to complete the task, which is used to normalize the time cost. max and Δ ma Normalize costs across generations to ensure comparability of cost calculations across different iteration cycles.

[0248] S423, according to the normalized distance cost μ ij and normalized time cost θ ij , calculate the comprehensive cost F(i,j) required for each drone in the drone cluster to move to the next target point, and form the cost matrix F [m×n] The cost matrix F [m×n] Each element F(i,j) in represents the i-th drone moving to the next target point X j The comprehensive cost required for (t+1).

[0249] The comprehensive cost F(i,j) is calculated using formula (23):

[0250] F(i,j)=a*μ ij +(1-a)*θ ij (twenty three);

[0251] Among them, a represents the weight parameter, which is a∈[0,1] and is used to balance the optimization weight of the two costs of total moving distance and maximum time spent in the path planning of UAV clusters; μ ij represents the normalized distance cost of the i-th UAV moving to the j-th target point; θ ij represents the normalized time cost of the i-th UAV moving to the j-th target point, and F(i,j) represents the comprehensive cost of the i-th UAV moving to the j-th target point. Parameter a is used to adapt to different mission requirements, such as search priority distance and monitoring priority time.

[0252] Specifically, the current position matrix S(t) = [s1(t), s2(t), ..., sm(t)], S i (t)=(x i (t),y i(t)) is the coordinate of the i-th UAV in the t-th generation. The target position matrix X(t+1) = [x1(t+1),…,xn(t+1)] is determined by the next generation position set X(t+1) output in step S37 (n = NP×Mf, including all candidate target positions). In the path planning process, positions p1 and p2 represent two two-dimensional coordinate points, whose specific coordinates are (x1, y1) and (x2, y2). For the i-th UAV, its initial position in the t-th generation is denoted by S i (t) represents, and from position S i (t) Move to target position X j The relevant parameters at (t+1) include the normalized distance cost μ ij and normalized time cost θ ij Among them, μ ij It is used to represent the normalized distance of the moving path, which can intuitively reflect the length characteristics of the path; θ ij It represents the time cost required to complete the path movement, which is mainly determined by the moving speed v of the UAV.

[0253] S43, record the maximum distance extreme value parameter γ between the UAV and the target position calculated in the current generation max and the maximum time consumption parameter Δ for the UAV to complete the task max , for normalization use when calculating the comprehensive cost matrix in the next generation, and as the benchmark parameter for normalizing distance and time costs when calculating the comprehensive cost matrix in the next generation.

[0254] In each generation, the key performance indicators of the previous generation need to be recorded, including the maximum distance γ between the previous generation drone and the target location max And the maximum time consumption Δ when the drone cluster completes the task max (The maximum time consumed by a single drone in this iterative position update) is used to balance the difference in distance and time measurement units and achieve normalization. Step S43 is used to eliminate the dimensional difference between distance and time costs, ensure the rationality of the comprehensive cost function, and achieve cross-generation cost comparability through historical parameters to support dynamic optimization of path planning. Calculate the maximum moving distance of all drones in the current generation and update it as the γ value of the next generation. max ; Calculate the maximum movement time of all drones in the current generation and update it to the next generation Δ max By reusing historical extreme values, we can ensure that μ ij and θ ij Having the same dimensional benchmark can avoid cost fluctuations caused by environmental changes.

[0255] S5. Design obstacle avoidance strategies for drone swarms to perform real-time obstacle avoidance.

[0256] The artificial potential field method is used to establish repulsive relationships between drones and between drones and obstacles, ensuring that the swarm avoids collisions during movement and safely completes the search mission. The artificial potential field method controls the movement of drones by constructing a virtual repulsive field, thus avoiding collisions and obstructions.

[0257] Furthermore, in step S5, the design of the drone cluster obstacle avoidance strategy to perform real-time obstacle avoidance specifically includes the following steps:

[0258] S51. Define a safety zone: Define a safety zone with a radius of d around each drone to monitor potential collisions and obstacles in real time. The value of d must be larger than the drone's body size, typically 2-3 times the radius, or adjusted based on the communication radius and sensor detection range.

[0259] S52: Establish a virtual force field, and based on the virtual force field, the UAV performs real-time obstacle avoidance.

[0260] When other drones or obstacles enter the safe area, they generate repulsive force on drone i. The repulsive force function is as follows:

[0261]

[0262] Among them, d ij represents the distance between drone i and obstacle / drone j; Represents the repulsive force scale factor, which controls the repulsive force strength, and the unit is N·m 2 ;U repij =((i, j) / (i)) / (i) indicates that drone i is repelled by obstacle / drone j. The repulsive force is directed from the obstacle / drone towards drone i, forcing drones away from the threat source. The real-time obstacle avoidance results influence the cluster distribution pattern, indirectly affecting the multi-objective optimization in step S3.

[0263] S53. Use formula (25) to update the position of the drone after being affected by the repulsive force:

[0264] x_new=x+U repi (25)

[0265] Among them, x is the position of the UAV before being affected by the repulsive force, x_new is the position of the UAV after being affected by the repulsive force, and U repi is the repulsive force on drone i. The adjusted position is fed back to step S4, which may trigger path replanning.

[0266] Step S5 builds the real-time obstacle avoidance capability of the UAV swarm through the definition of safe areas, repulsive field modeling and position update mechanism. By combining the artificial potential field method with cluster collaborative control, it not only ensures the safety of individual UAVs, but also maintains the overall search efficiency of the cluster. It is a key guarantee for the reliable execution of UAV swarm missions in complex dynamic environments. Figure 5 Middle,U repij is the repulsive force exerted by UAV j on UAV i, U repio is the repulsive force generated by obstacle o on drone i, where is the repulsive force scale factor, d ij is the distance between drone i and drone j, and d is the maximum distance of repulsion. A virtual repulsion field is generated based on the positions of other drones or obstacles, forcing the flying drones to deviate from the danger zone to avoid collision. Figure 5 The resultant force of the obstacle on drone i or the repulsion on drone j is shown, which achieves the effect of drone avoiding obstacles. The virtual repulsion field is used to prevent drones from entering dangerous areas, such as Figure 5 As shown in , when an obstacle enters the safe area, the drone is pushed away. The repulsive force between drones ensures a reasonable spacing, avoids the "clustering" phenomenon, and maintains the uniform distribution required for collaborative search, as shown in Figure 7f The swarm remains dispersed as it approaches the target. It responds in real time to moving obstacles (such as dynamic rocks and trees) and the relative motion of the drones within the swarm, ensuring the timeliness of the obstacle avoidance strategy.

[0267] Complex dynamic environments include search scenarios for multiple moving obstacles and dynamic targets (with escape capabilities). Taking into account the complexity of the real target search environment, the conditions of multiple obstacles, dynamic obstacles, multiple targets, dynamic targets, and intelligent targets do not necessarily appear in the same search environment at the same time. Since the current drone cluster target search method does not have a systematic search environment combination, the present invention proposes a multi-obstacle, multi-target intelligent dynamic series search environment based on the characteristics of the real target search environment: a fixed drone cluster flight starting point; a 1000*1000 search environment grid map, where the grid color represents the target signal strength / probability of occurrence at the corresponding position; two types of moving obstacles; three targets that need to be searched, and the targets have the ability to avoid exploring drones in real time, such as Figure 6 shown. Figure 6 This is the initial environment for the drone cluster to search for targets, which is used to demonstrate the effect of the experimental scenario. The green and brown rectangles in the figure are obstacles, and the yellow area is the potential target area.

[0268] Figure 7 shows the change of the distribution of drones with the iteration of the algorithm proposed in this invention when the number of drones is 20, and shows the process of the drone cluster searching for the target and destroying it in turn under three target scenarios. Figure 7a As shown, the drone cluster takes off from a fixed starting point and searches for the target. Figure 7b and Figure 7c Shows the status diagram of the drone cluster searching, finding and destroying the first target. Figure 7d and Figure 7e The diagram shows the status of the drone cluster searching, finding and destroying the second target. Figure 7f The following figure shows the state diagram of the drone swarm searching, discovering and attempting to destroy the third target. Figure 7a-7f It can be seen that by optimizing the degree of coordination through the method described in the present invention, the drone cluster can gradually cover the entire area in iterations, discover and destroy targets in sequence, and verify the effectiveness of the method described in the present invention. Figure 7a-7f In the figure, the x-axis and y-axis represent the geographic coordinate system of the two-dimensional plane space, and the z-axis represents the signal strength. Figure 7a-7f In the environment shown, there are three targets in total. These figures show the experimental schematic diagrams of the drone cluster from finding the first target to finding the second target and the third target. The drones maintain high coordination to complete the task. Figure 7b and 7d It can be seen that by guiding the drone to fly to the high signal area first through the signal strength target, the target can be quickly located; Figure 7f As can be seen, the coordination objective ensures that the cluster is evenly distributed and covers the entire area to detect subsequent targets. Through dual-objective optimization, the drone cluster achieves rapid and accurate target positioning in complex environments, verifying the effectiveness of the coordination indicator.

[0269] Example 2

[0270] This embodiment provides a drone swarm target search device for complex dynamic environments, which can apply the drone swarm target search method for complex dynamic environments. The device includes a coordination index acquisition module, a target signal strength prediction module, a drone swarm position distribution adjustment module, a drone optimal path determination module, and an obstacle avoidance module.

[0271] The synergy index acquisition module is used to obtain the synergy index based on the adaptability surface, and uses signal strength and synergy as parallel optimization targets; the target signal strength prediction module is used to construct a cost model based on ecological heritage and dynamically predict the target signal strength in an unknown area; the drone cluster position distribution adjustment module is used to construct a drone cluster multi-target search framework based on the target signal strength prediction value and the synergy index, and dynamically adjust the drone cluster position distribution through dual-target optimization; the drone optimal path determination module is used to determine the drone optimal path based on the cost model and the drone cluster multi-target search framework; the obstacle avoidance module is used to design a drone cluster obstacle avoidance strategy and perform real-time obstacle avoidance. The specific implementation method of each module is as described in Example 1 and will not be repeated here.

[0272] Example 3

[0273] This embodiment also provides an electronic device, comprising: at least one processor; and a memory, wherein the memory stores instructions, and when the instructions are executed by the at least one processor, the at least one processor executes the drone cluster target search method for complex dynamic environments as described above.

[0274] In this embodiment, electronic devices may include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile computing devices, smart phones, tablet computers, cellular phones, personal digital assistants (PDAs), handheld devices, messaging devices, wearable computing devices, consumer electronic devices, and the like.

[0275] Example 4

[0276] This embodiment also provides a machine-readable storage medium storing executable instructions, which, when executed, enable the machine to execute the drone cluster target search method for complex dynamic environments as described above.

[0277] Specifically, a system or device equipped with a readable storage medium can be provided, on which software program codes that implement the functions of any of the above-mentioned embodiments are stored, and a computer or processor of the system or device can read and execute instructions stored in the readable storage medium.

[0278] In this case, the program code itself read from the machine-readable medium can implement the functions of any one of the above embodiments, and thus the machine-readable code and the machine-readable storage medium storing the machine-readable code constitute part of this specification.

[0279] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (e.g., CD-ROMs, CD-Rs, CD-RWs, DVD-ROMs, DVD-RAMs, DVD-RWs, DVD-RWs), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code may be downloaded from a server computer or a cloud via a communication network.

[0280] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0281] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0282] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0283] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0284] The above-described embodiments are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for searching for target in a swarm of drones in a complex dynamic environment, characterized in that: The method comprises the following steps: S1. Obtain a synergy index based on the adaptability surface, and use signal strength and synergy as parallel optimization targets; S2. Build a cost model based on ecological heritage to dynamically predict target signal strength in unknown areas; S3. Based on the predicted value of target signal strength and the coordination index, a multi-target search framework for drone clusters is constructed, and the position distribution of drone clusters is dynamically adjusted through dual-objective optimization. S4. Determine the optimal path for the UAV based on the cost model and the UAV cluster multi-objective search framework; S5. Design obstacle avoidance strategies for drone swarms to perform real-time obstacle avoidance.

2. The method for searching for target in a swarm of drones in a complex dynamic environment according to claim 1, wherein: In step S1, the construction of a synergy index based on the adaptability surface, with signal strength and synergy as parallel optimization targets, includes: S11, constructing an objective function based on signal strength, taking signal strength as the first optimization objective; The signal strength-based objective function includes an individual signal strength function and a group signal strength function; The individual signal strength function is: f1(x i )=signal_strength(x i ) (1); The group signal strength function is: Among them, f1(x i ) represents the signal strength of the position of the i-th UAV, x i represents the position of the i-th drone, NP represents the number of drones in the drone cluster, the signal_strength() function is used to obtain the signal strength, f1(x1,x2,…,x NP ) represents the sum of the signal strengths of each individual drone in the drone group; S12, constructing an objective function based on synergy, and taking synergy as the second optimization objective; The objective function based on the degree of coordination is: Among them, x i represents the position of the i-th UAV, x j represents the position of the jth UAV, NP represents the number of UAVs in the UAV cluster, represents the average distance between the positions of the drone cluster, f2(x1,x2,…,x NP ) represents the variance of the distance between drones and is used to indicate the degree of coordination among drones in a group.

3. The method for searching for target in a swarm of drones in a complex dynamic environment according to claim 1, wherein: In step S2, a cost model based on ecological heritage is constructed to dynamically predict the target signal strength in the unknown area, including: S21. Obtain search environment information and construct a dynamically updated training set; S22. Design an incremental random forest model, use the incremental random forest model to process the search environment information in batches, and dynamically predict the target signal strength in the unknown area; The step S21 specifically includes the following steps: S211, obtain the search environment information accumulated in the historical iteration to form the original data set D old ; S212, using drone sensors to obtain current search environment information to form incremental data D new ; S213, using formula (4) to convert the incremental data D new Compared with the original dataset D old Merge into a total data set D, which is used as the training set: D=D new ∪D old (4); Among them, D represents the entire data set obtained by the drone sensor under the current search environment information, D new Indicates the newly acquired incremental data under the current search environment information obtained by the drone sensor, D old Represents the historical data set previously acquired by the drone sensor under the current search environment information; The step S22 specifically includes the following steps: S221. Design an incremental random forest model; S222, performing incremental training of decision trees on the incremental random forest model, and determining a new decision tree through a double random sampling mechanism and error-driven splitting; S223, integrate the newly added decision tree into the original random forest, and based on the incremental data D new Training new decision trees to expand the random forest; the original random forest old Contains T old A decision tree is added based on incremental data training to obtain T new The newly added decision trees are integrated into the original random forest using formula (8), and the random forest is expanded to obtain the expanded random forest Forest. new : Among them, Forest new Represents the expanded random forest, Forest old represents the historical random forest, T new Indicates the number of newly added decision trees, Tree k represents the kth new decision tree; S224. Based on the expanded random forest, the original decision tree and the newly added decision tree are merged, and the predicted value of the input feature x is updated to the average value of all the merged trees using formula (9) to obtain the integrated signal strength prediction value in, Represents the signal strength prediction value, T new Indicates the number of newly added decision trees, T old Represents the number of decision trees in the historical random forest, Tree t represents the tth new decision tree; S225, according to the incremental data D new , using formula (10) to calculate the importance of features and identify the features that contribute most to signal strength prediction, so as to dynamically adjust the model focus: Among them, Importance(j) represents the importance of the j-th feature, ∑ splitonj △MES represents the sum of all MES improvements based on feature j, Total Reduction in MES represents the sum of MES improvements brought by all features, and MES represents the mean square error of the current sample set; S226, using formula (11) to calculate the incremental data D new Make predictions and calculate the generalization error for the newly added data. Adjust the number or depth of trees based on the generalization error to modify the incremental random forest model: Among them, ε new represents the generalization error of the newly added data, y i Indicates the true value of signal strength, represents the predicted signal strength value, Express Perform mathematical expectation operations; The step S222 specifically includes the following steps: S2221, through Bootstrap sampling, from the incremental data D new Sampling forms a random sample subset S; S2222, at each split point, randomly select a feature subset from the total feature set F of the random sample subset S The feature subset The size m is determined by formula (5): Where m is the size of the feature subset randomly selected at each split, that is, the number of features; the total feature set F refers to the set of all available features in the training data; S2223, at each decision tree node, based on the sample subset S and feature subset F generated in step S2221 rand , by calculating the mean square error MES and weighted error Loss, select the features and their thresholds that minimize the prediction errors of the left and right child nodes as the splitting points, generate the left and right subtrees, and determine the newly added decision tree; The mean square error MES is calculated using formula (6): Among them, MES(S) is used to measure the prediction error of the sample subset S, S represents the prediction error from the incremental data D new Sampling forms a random sample subset, |S| represents the number of samples in the sample set S, y i Indicates the true value of signal strength, represents the predicted value of signal strength; The weighted error Loss is calculated using formula (7): Among them, formula (7) is used to minimize the weighted error, and the split point is selected to minimize the sum of the weighted MES of the left and right child nodes after the split; Loss (S) represents the weighted loss function after the split point is selected; S represents the weighted loss function from the incremental data D new Sampling forms a random sample subset, |S| represents the number of samples in the sample set S, S L is the sample set of child nodes of the left subtree after splitting, S R is the sample set of child nodes of the right subtree after splitting, MES(S L ) represents the mean square error of the left subset, MES(S R ) represents the mean squared error of the right subset.

4. The method for searching for target in a swarm of drones in a complex dynamic environment according to claim 1, wherein: The step S3, constructing a multi-target search framework for the drone cluster based on the target signal strength prediction value and the coordination index, and dynamically adjusting the position distribution of the drone cluster through dual-objective optimization, includes: S31. Initialize the location set X of the drone cluster t , X t As the initial candidate solution for multi-objective optimization; initialize the drone cluster position set The location of each drone is a randomly generated point in the search space Ω, Expressed as: Among them, Ω represents the search space of the drone cluster, and NP represents the number of drones in the drone cluster; S32, based on a certain UAV cluster target search algorithm, from the current position set X t Generate NP*M f Next-generation positions form the next-generation drone cluster position set X t+1 : Among them, NP represents the number of drones in the drone cluster, and the parameter M f Determines the number of next generation drones to be generated, M f The larger it is, the more locations are generated. Indicates the position of the next generation drone No. i; S33, based on the next generation of drone cluster position collection X t+1 , combined generation distribution patterns, forming a distribution pattern set; among them, Indicates the number of all possible drone cluster location distribution schemes combined from all generated candidate locations; S34. Based on the distribution pattern set and the cost model, the signal strength of each distribution pattern is calculated using formula (14) to obtain a signal strength value set for each distribution pattern: f1(y1,y2,…,y NP )=max(CM(y1),CM(y2),…,CM(y NP )) (14); Among them, CM is the current cost model; (y1,y2,…,y NP ) represents the location set of the drone cluster in the current distribution pattern, f1(y1,y2,…,y NP ) represents the predicted value of the strongest signal strength in the current distribution pattern; S35. Based on the distribution pattern set, the synergy degree of each distribution pattern is calculated using formula (15) to obtain a synergy degree value set for each distribution pattern; Among them, NP represents the number of drones in the drone cluster, represents the average distance of the drone cluster position, (y1,y2,…,y NP ) represents the location set of the drone cluster in the current distribution pattern, f2(y1,y2,…,y NP ) represents the variance prediction value of the distance between drones in the current distribution pattern, which is used to indicate the degree of coordination of the group of drones; S36, comparing the cost model signal strength f1 and the coordination degree f2 of all distribution patterns, and using the non-dominated sorting method to select the non-dominated solution as the optimal solution to form a non-dominated solution set; S37. In the search for non-dominated solutions, randomly select a distribution pattern as the location set X of the next generation UAV cluster. t+1 , determine the final distribution pattern.

5. The method for searching for target in a swarm of drones in a complex dynamic environment according to claim 1, wherein: In step S4, determining the optimal path of the drone based on the cost model and the drone cluster multi-objective search framework includes: S41. Design a cost function based on the movement distance of the drone cluster and the time required to complete the task; The cost function is shown in formula (16): F d =min{a*γ+(1-a)*Δ} (16); in, Tspan(t)=maxτ i (t),1≤i≤n (20); Among them, F d represents the cost function based on the moving distance of the drone cluster and the time required to complete the task; a represents the weight parameter, a∈[0,1], which is used to optimize the weight of the two costs of total moving distance and maximum time spent in the path planning of the drone cluster; γ represents the total moving distance of the drone cluster after completing the entire task; Δ represents the maximum moving time of the drone cluster after completing the task; Z(t) represents the cumulative moving distance of all drones in the tth generation; represents the moving distance of the i-th UAV in the t-th generation; Tspan(t) represents the maximum moving time spent by the t-th generation UAV cluster; τ i (t) represents the specific moving time of the i-th UAV in the t-th generation; S42, initializing the state of the drone cluster, calculating the comprehensive cost of each drone moving to each target location, and determining the optimal path for the drone based on the comprehensive cost; S43, record the maximum distance extreme value parameter γ between the UAV and the target position calculated in the current generation max and the maximum time consumption parameter Δ for the UAV to complete the task max , for normalization in the next generation of calculation of the comprehensive cost matrix.

6. The method for searching for target in a swarm of drones in a complex dynamic environment according to claim 1, wherein: The step S42 specifically includes the following steps: S421. Initialize the time vector τ of the tth generation [1×m] (t) = {0,…,0}, where m represents the number of drones and each element corresponds to the cumulative movement time of a drone; S422, based on the time vector, normalize the distance cost μ ij and normalized time cost θ ij ; The normalized distance cost μ ij Using formula (21), we can calculate: m ij =D(S i (t),X j (t+1)) / γ max (21); The normalized time cost θ ij Using formula (22), we can calculate: i ij =(t i +D(S i (t),X j (t+1)) / v) / Δ max (22); Among them, D(S i (t),X j (t+1)) represents the Euclidean distance between the current position and the target position, using Calculated; γ max represents the maximum distance between the previous generation UAV and the target location; τ i represents the cumulative time of the i-th UAV in the current iteration, which is initially 0; v represents the average moving speed of the UAV; Δ max Indicates the maximum time consumption of the previous generation of drones to complete the task, which is used to normalize the time cost; S423, according to the normalized distance cost μ ij and normalized time cost θ ij , calculate the comprehensive cost F(i,j) required for each drone in the drone cluster to move to the next target point, and form the cost matrix F [m×n] ; The cost matrix F [m×n] Each element F(i,j) in represents the i-th drone moving to the next target point X j The comprehensive cost required for (t+1); the comprehensive cost F(i,j) is calculated using formula (23): F(i,j)=a*μ ij +(1-a)*θ ij (23); Among them, a represents the weight parameter, a∈[0,1] is used to balance the optimization weight of the total moving distance and the maximum time spent in the path planning of the UAV cluster; μ ij represents the normalized distance cost of the i-th UAV moving to the j-th target point, θ ij represents the normalized time cost of the i-th UAV moving to the j-th target point, and f(i,j) represents the comprehensive cost of the i-th UAV moving to the j-th target point.

7. The method for searching for target in a swarm of drones in a complex dynamic environment according to claim 1, wherein: In step S5, the design of a drone cluster obstacle avoidance strategy to perform real-time obstacle avoidance includes: S51. Define safety zone: Define a safety zone with a radius of d around each drone to monitor potential collision and obstacle threats in real time. S52: Establish a virtual force field. Based on the virtual force field, the UAV performs real-time obstacle avoidance. When other UAVs or obstacles enter the safe area, a repulsive force is generated on UAV i. The repulsive force function is as follows: Among them, d ij represents the distance between drone i and obstacle / drone j; Represents the repulsive force scale factor, which controls the repulsive force strength, and the unit is N·m 2 ;U repij It means that drone i is repelled by an obstacle or drone j, and the repulsive force is directed from the obstacle / drone to the opposite direction of drone i, forcing the drone to move away from the threat source; S53. Use formula (25) to update the position of the drone after being affected by the repulsive force: x_new=x+U repi (25); Among them, x represents the position of the UAV before being affected by the repulsive force, x_new represents the position of the UAV after being affected by the repulsive force, and U repi is the resultant repulsive force acting on UAV No. i.

8. A UAV cluster target search device for complex dynamic environments, characterized in that: The device includes a coordination index acquisition module, a target signal strength prediction module, a UAV cluster position distribution adjustment module, a UAV optimal path determination module, and an obstacle avoidance module; The synergy index acquisition module is used to obtain a synergy index based on the adaptability surface, taking signal strength and synergy as parallel optimization targets; The target signal strength prediction module is used to construct a cost model based on ecological heritage and dynamically predict the target signal strength in unknown areas; the drone cluster position distribution adjustment module is used to construct a drone cluster multi-objective search framework based on the target signal strength prediction value and the coordination index, and dynamically adjust the drone cluster position distribution through dual-objective optimization; the drone optimal path determination module is used to determine the optimal drone path based on the cost model and the drone cluster multi-objective search framework; The obstacle avoidance module is used to design a drone cluster obstacle avoidance strategy and perform real-time obstacle avoidance.

9. An electronic device, characterized in that: include: at least one processor; and a memory storing instructions, which, when executed by the at least one processor, enable the at least one processor to execute the drone cluster target search method for a complex dynamic environment as described in any one of claims 1 to 7.

10. A machine-readable storage medium, characterized in that It stores executable instructions, which, when executed, enable the machine to perform the drone cluster target search method for complex dynamic environments as described in any one of claims 1 to 7.

Images