Multi-unmanned aerial vehicle route planning method in multi-wind area environment

Abstract

The invention discloses a multi-unmanned aerial vehicle route planning method in a multi-wind area environment. The method mainly comprises the following steps: establishing a three-dimensional map of typical task areas of unmanned aerial vehicles, dividing the areas, setting different wind directions and wind speeds, setting a plurality of unmanned aerial vehicles, and setting a fixed starting point and a fixed target point; randomly distributing a target point for each unmanned aerial vehicle by using an evolutionary algorithm, calling an improved RRT * algorithm for each unmanned aerial vehicle, and searching a feasible leg from the starting point to the target point in the three-dimensional map; iterating the distribution schemes of different target points by using an evolutionary algorithm and a fitness function to obtain an optimal target point distribution scheme, and further generating a flight path of the unmanned aerial vehicle; and performing path stretching and cubic B-spline interpolation on the initial route of each unmanned aerial vehicle to obtain route planning of the multiple unmanned aerial vehicles.

Description

Technical Field

[0001] This invention belongs to the field of shipborne unmanned aerial vehicle (UAV) route planning technology, specifically relating to a multi-UAV route planning method in a multi-wind zone environment. Background Technology

[0002] With the widespread application of shipborne UAVs in maritime combat and support missions, how to quickly generate safe and economical multi-aircraft cooperative flight schemes in complex and ever-changing sea and air weather, obstacle and multi-target environments has become a hot research issue.

[0003] While traditional RRT* algorithms can achieve asymptotic optima through iterative rewiring, they often suffer from slow convergence, path redundancy, and high energy consumption in three-dimensional multi-wind zone environments due to unbiased sampling, fixed step size, and failure to consider real-time wind field effects. Furthermore, simple evolutionary allocation strategies struggle to balance path feasibility and fuel economy. Existing technologies typically optimize allocation individually or plan for single drones, lacking a holistic and coupled consideration of multi-UAV, multi-target flight paths, thus failing to meet the demands of efficient and reliable planning in shipboard environments.

[0004] Therefore, there is an urgent need for a new method for multi-UAV multi-target point route planning. Summary of the Invention

[0005] Purpose of the invention: This invention aims to provide a method for collaborative flight path planning for multiple UAVs performing multi-target tasks in complex sea and air environments. It takes into account both global search efficiency and local convergence quality, and can also consider wind speed, wind direction and fuel consumption in real time, so as to achieve efficient, safe and multi-target collaborative flight in low-altitude three-dimensional environment.

[0006] Technical solution:

[0007] This invention proposes a multi-UAV route planning method in a multi-wind zone environment, including:

[0008] A 3D map of the UAV mission area is established, and multiple UAVs are set in the 3D map, with several starting points and several target points; a starting point is assigned to each UAV; different wind vectors w, including wind direction and wind speed, are set at different locations on the 3D map.

[0009] An evolutionary algorithm is used for target point allocation. The target point allocation scheme is set as an individual of the evolutionary algorithm. An improved RRT* algorithm is used to generate feasible routes based on each individual. The fitness function of the evolutionary algorithm is the negative value of the sum of the path length of the feasible route and the fuel consumption. The evolutionary algorithm terminates when the fitness function converges, thus obtaining the optimal target point allocation scheme. The fuel consumption is a fuel consumption function with a wind correction factor introduced, and the wind correction factor is related to the wind vector w.

[0010] Based on the optimal target point allocation scheme, the improved RRT* algorithm is used to generate the flight paths of all UAVs from their starting point to the target point. The flight paths are then stretched and interpolated using cubic B-spline to obtain the route planning for multiple UAVs in a multi-wind environment.

[0011] Furthermore, different wind directions and speeds are set in the three-dimensional map, including:

[0012] The drone mission area is divided into several wind zones according to different wind directions and speeds. Each wind zone corresponds to a wind vector w. When the improved RRT* algorithm searches the 3D map, it determines the wind zone where the current position is located based on the coordinates and obtains the corresponding wind vector w.

[0013] Furthermore, the step of using an evolutionary algorithm for target point allocation includes: setting the target point allocation scheme as an individual of the evolutionary algorithm, where each individual is a one-dimensional array, and the i-th element j in the array represents the allocation of target point j to UAV i.

[0014] Furthermore, the improved RRT* algorithm includes:

[0015] (1) Set the starting point and target point according to the individual, initialize the tree of the search space, and each node contains coordinates (x, y, z), parent node information and cumulative cost;

[0016] (2) The sampling point is obtained by using the target bias sampling method. The nearest node of the sampling point is found in the tree. The nearest node is moved one step in the direction of the potential field attraction to obtain a new node. The potential field attraction is a linear combination of the direction from the nearest node to the target point and the direction from the nearest node to the sampling point. After each sampling is completed, the step size is updated by a function with an inertia factor.

[0017] (3) Extract nodes within a certain distance range of the new node to the extended set, find the ancestor node of each candidate node in the extended set and add it to the extended set;

[0018] (4) Check the collision between each candidate node and the new node one by one from the extended set, calculate the cost, and select the candidate node with the lowest cost as the parent node of the new node;

[0019] (5) Repeat steps (2) to (4) until the sampling point is the target point, and generate feasible flight segments based on all nodes in the tree.

[0020] Furthermore, the target bias sampling method is expressed as follows:

[0021]

[0022] In the formula For sampling points, For the target point, Let be the target bias probability, when In the case of sampling, the sampling point is the target point; otherwise, the global range is maintained. Random sampling is implemented within the facility.

[0023] Furthermore, the cost is a fuel consumption function that takes into account a wind correction factor. , represented as:

[0024]

[0025] in Basic fuel quantity:

[0026]

[0027] In the above formula, k cruise The cruise fuel consumption coefficient is d, where d is the parent node q. p With child node q c The distance between them, k turn The turning fuel consumption coefficient is given by Δθ, where Δθ is the turning angle and k is the turning angle. climb It is the fuel consumption reduction coefficient for climbing, k descend It reduces the fuel consumption coefficient;

[0028] The wind vector at the location of the parent node is w, and the angle between the unit vectors of the flight directions of the parent node and the child node is θ. Wind fitness correction is then applied. The wind correction factor was obtained. :

[0029]

[0030] In the formula, β is the adjustment parameter, and β>0.

[0031] Furthermore, the fitness function includes:

[0032] Divide the entire feasible flight segment into N segments, with the length of the i-th segment being... , Let W be the wind efficiency weighting factor for the i-th segment, and let W be the wind speed fitness index:

[0033]

[0034] Let f climb / desend To increase / decrease fuel consumption, f turn For fuel consumption during cornering, f cruise For cruising fuel consumption, fuel consumption Represented as:

[0035]

[0036] The fitness function W 总 The total path length and fuel consumption of the feasible flight segment. The negative value of the sum is represented as:

[0037] W 总 = - (path length + F total ).

[0038] Furthermore, an evolutionary algorithm is used for target point allocation, including:

[0039] Individuals in the evolutionary algorithm are evaluated based on the fitness function. The top three individuals with the best fitness are selected. While retaining the best individuals, new individuals are generated through crossover and mutation operations to gradually optimize the population.

[0040] Select two parent individuals and generate new offspring individuals by exchanging some of their genetic information; randomly change a gene of an individual to introduce a new solution space and avoid the algorithm getting stuck in local optima.

[0041] Repeat the above process until the optimal solution of the fitness function tends to stabilize after multiple iterations.

[0042] Furthermore, the path stretching includes: checking non-adjacent nodes in the flight path one pair at a time; if the pair of non-adjacent nodes satisfies a collision-free direct connection, removing the intermediate node of the pair of non-adjacent nodes.

[0043] Furthermore, the cubic B-spline interpolation generates a smooth curve based on the path obtained after path stretching, using a set of control points and piecewise cubic polynomial basis functions.

[0044] Beneficial effects:

[0045] 1. Global and Local Co-optimization. This invention embeds an improved RRT* path planning mechanism into the evolutionary allocation framework. This not only injects multi-objective heuristics into population initialization, crossover, and mutation stages, but also achieves an organic integration of multi-UAV task allocation and flight path search through mechanisms such as target bias sampling, potential field attraction, adaptive step size, and ancestor node rewiring. This co-optimization strategy enables the algorithm to achieve a good balance between global breadth exploration and local depth optimization, significantly improving convergence speed and effectively avoiding the risk of getting trapped in local optima, thereby significantly improving the quality and stability of the overall scheduling solution.

[0046] 2. Balancing Fuel Economy and Wind Field Adaptability. To address varying wind field conditions, this invention couples the local wind speed vector with the base fuel cost in real time during path expansion and node rewiring, enabling the search process to proactively maximize benefits and minimize risks: prioritizing expansion in downwind areas and cautiously selecting routes in upwind areas. Simulation results show that this fuel-wind field coupling model significantly reduces overall fuel consumption while ensuring safety, and achieves higher wind adaptability scores in multi-wind field tests.

[0047] 3. Streamlined, smooth, and executable flight paths. For discrete node flight paths generated by the improved RRT* algorithm, this invention first removes redundant intermediate nodes by stretching the path to eliminate unnecessary turns. Then, it employs cubic B-spline smoothing interpolation to eliminate angles while maintaining curvature continuity. This process not only improves the smoothness and executability of the flight path but also further shortens the actual flight distance, reduces turning energy consumption, and achieves a seamless transition from theoretical planning to actual flight. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the method of the present invention;

[0049] Figure 2 This is a scene layout diagram of the present invention;

[0050] Figure 3 This is a schematic diagram of the improved RRT* process of the present invention;

[0051] Figure 4 This is a comparison chart of the running results of different algorithms and different indicators in Embodiment 2 of the present invention;

[0052] Figure 5 This is a bar chart comparing the average results of multiple runs of different algorithms and different indicators in Embodiment 2 of the present invention;

[0053] Figure 6 This is a schematic diagram of the basic RRT planning results of Embodiment 2 of the present invention;

[0054] Figure 7 This is a schematic diagram of the basic RRT* planning results of Embodiment 2 of the present invention;

[0055] Figure 8 This is a schematic diagram of the Adaptive Step-size RRT* results in Embodiment 2 of the present invention.

[0056] Figure 9 This is a schematic diagram of the improved RRT* planning results of Embodiment 2 of the present invention. Detailed Implementation

[0057] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. The present invention provides a multi-UAV multi-target route planning method based on an improved RRT* algorithm in complex sea and air environments. The flow of the multi-UAV multi-target route planning method of the present invention is as follows: Figure 1 As shown, the specific steps are as follows:

[0058] Step (1): Arrangement of three-dimensional space with multiple wind speeds and directions.

[0059] Based on the three-dimensional spatial range, the drone's starting point and target distribution, and the presence of obstacles, the entire space can be divided into four regions, each assigned different wind speeds and directions to simulate headwinds, tailwinds, and two different crosswind environments. The three-dimensional spatial model established in this embodiment is as follows: Figure 2 As shown, black dots represent the start point, red stars represent the goal point, and arrows with numbers next to them indicate wind direction and speed. The start point for each drone is manually specified.

[0060] The starting point of the drones is located in the lower left region, which in this embodiment is set near (-40, -40, 25), while most targets are located in the upper right region, indicating that the overall flight trend of the drones is from the lower left to the upper right. Therefore, a headwind zone is set in the lower left region, and a tailwind zone is set in the upper right region. To fully test the robustness of the algorithm under different wind fields, two crosswind zones are set in the remaining regions, so that when the drones enter these regions, they will be affected by lateral winds perpendicular to the main flight direction. In this embodiment, the two-dimensional distribution of space (xy plane) is used, and the entire region is divided by x=25 and y=25. 25 is an artificially set boundary value, which ensures that the lower left region includes the starting point and that the upper right region covers more targets. The three-dimensional space (the z-direction directly takes the full height 0~50) is divided into four sub-regions:

[0061] Region 1 (bottom left): x∈[-50,25], y∈[-50,25], is the headwind region;

[0062] Region 2 (top left): x∈[-50,25], y∈[25,100], is the crosswind region (crosswind direction 1);

[0063] Region 3 (bottom right): x∈[25,100], y∈[-50,25], is the crosswind region (crosswind direction 2);

[0064] Region 4 (top right): x∈[25,100], y∈[25,100], is the downwind region.

[0065] Step (2): Initialize the multi-target assignment task.

[0066] Multiple target points are rationally allocated to multiple drones to ensure optimal task assignment for each drone. The core of this stage is to use evolutionary algorithms for global optimization. In evolutionary algorithms, the population refers to a group of possible individuals, and each individual represents a target point allocation scheme, that is, the target point sequence for each drone.

[0067] This embodiment employs a random initialization strategy. First, several individuals are randomly generated. Each individual is encoded as a one-dimensional array of target-drone, where each element represents which drone a target point is assigned to. For example, the array [1, 2, 1, 3] indicates that the first target point is assigned to drone 1, the second target point to drone 2, the third target point to drone 1, and the fourth target point to drone 3.

[0068] Step (3): Set the fitness function of the evolutionary algorithm.

[0069] For each individual in the population, based on its target allocation scheme, the improved RRT* path planning algorithm is invoked to calculate the flight path of each drone. The path lengths and fuel consumption of all drones are summed to obtain the individual's total path length and total fuel consumption. In the evolutionary algorithm, a fitness function is used to evaluate the performance of each individual. The fitness function W... 总 The path length and fuel consumption of the feasible flight segment The negative value of the sum is represented as: W 总 = -(path length + F total Fitness value W 总 The higher the value (the smaller the absolute value of the negative value), the better the allocation scheme. The path length and fuel consumption are calculated by the RRT* path planning algorithm.

[0070] This invention improves upon the traditional RRT* algorithm by designing an enhanced RRT* path planning algorithm. It sets the location of the UAV as a node and searches for a path from the UAV's preset starting point to a target point contained within the individual. The traditional RRT* algorithm, given a starting point, expands the tree structure outwards by continuously sampling points randomly within the search space and connecting them to the nearest node in the tree. During expansion, it not only selects the lowest-cost parent node for each new node but also attempts to reconnect surrounding neighboring nodes through the new node, thereby continuously optimizing existing paths. After multiple iterations, the tree covers a wider area and, where feasible, gradually approaches the optimal path from the starting point to the target point until a termination condition is met.

[0071] The improved RRT* algorithm, based on the basic RRT* framework, incorporates several improved methods to better adapt to the path planning needs of UAVs in three-dimensional multi-wind field environments. These include target bias sampling, potential field gravity, adaptive step size, and Quick-RRT*. A new "fuel-efficient" cost model has also been added, enabling the algorithm to balance path feasibility and fuel economy when facing complex wind zones and three-dimensional obstacles. The improved RRT* algorithm flow is as follows: Figure 3 As shown, it specifically includes:

[0072] 1. Nodes and Data Structure: For 3D scenes, each node contains coordinates (x, y, z), parent node information, cumulative cost, and the direction vector required for improvement. The direction vector records the flight vector from the parent node to the current node to assist in calculating factors such as turning angles.

[0073] 2. Sampling and Expansion: In each iteration, the algorithm performs target bias sampling or random sampling with a certain probability, expanding from the nearest node towards that sampling point. To account for the influence of the wind field, the true position of the new node is superimposed with local wind displacement, and the expansion segment is evaluated using a fuel consumption function.

[0074] The sampling process is represented as follows:

[0075]

[0076] in Let be the target bias probability. When In the case of a target point, the current sampling point is directly set to equal the target point, significantly increasing the search frequency of the target region; in other cases, global random sampling is maintained to ensure that the algorithm still has global exploration capabilities. Given the search area; in the three-dimensional scene of this invention, for Compared to the traditional RRT* algorithm which only uses random sampling, the improved algorithm combines bias sampling with random sampling, which can accelerate the convergence speed in the later stages of the evolutionary algorithm search.

[0077] 3. Adaptive step size and potential field attraction: In terms of expansion direction and step size, the synthesis of the target potential field and random direction is introduced to accelerate the convergence speed; at the same time, the step size is dynamically adjusted by the reward mechanism, so that the algorithm can balance efficiency and feasibility in the search process.

[0078] Each successful node expansion (collision-free and successful addition of a new node) awards a +1 reward, while a collision or failed expansion results in a -0.5 penalty, which is accumulated in the global variable R. This accumulated reward reflects the success / failure frequency of the current search process. After each iteration sampling, the step size is updated using a function with an inertia factor, expressed as... ,in,

[0079]

[0080]

[0081] and These are the upper and lower limits of the inertia factor, respectively; t is the current iteration number. This represents the maximum number of iterations. and R represents the upper and lower bounds of the step size; R is the cumulative reward value. When R > 0, the algorithm tends to increase the step size to perform a larger search more quickly; when R < 0, the step size will decrease accordingly to avoid excessive collisions. This adaptive strategy allows the algorithm to flexibly adjust the step size at different search stages, balancing feasibility and search efficiency.

[0082] The improved RRT* algorithm designed in this invention also introduces the concept of potential field gravity, which... → The direction of "and" → Linear synthesis is performed in the direction of "". Among them, The nearest node, For sampling points, Let the target point be [the point]. The direction of composition is represented as:

[0083]

[0084] Where α is the composition weight (e.g., 0.5), controlling the balance between the potential field gravity and the random sampling direction. Direction d final After normalization, the position is used to generate the pure maneuvering position. The pure maneuvering position is the ideal target position obtained in path planning by temporarily ignoring external disturbances and considering only maneuvering propulsion. It is an intermediate step in wind field modeling, used to calculate the true position after the wind's influence on the path. This method more realistically reflects the actual flight path and improves the executability of the path.

[0085] 4. Quick-RRT* and Ancestor Node Search: After a new node is added, the ancestor nodes are traced upwards to perform rewiring, and the parent node relationships and costs of adjacent nodes are updated, thereby continuously optimizing the overall cost of the search tree.

[0086] Find the set N of nodes that are within the range of search_radius from the new node. nearFor each node n, its ancestor node (with a maximum depth of quick_rrt_depth) is searched upwards and added to the extended set. Collisions between nodes connected to the new node are checked one by one in this extended set, the new total cost is calculated, and the node with the lowest cost is selected as the parent node. Preferably, after successfully inserting a new node, a "rewiring" check is performed on other nodes in the extended set: if these nodes can be connected to through the new node at a lower cost, their parent node is switched to the new node, thus re-optimizing the local structure.

[0087] By using Quick-RRT* ancestor node search and rewiring, this invention can more actively perform cost-optimal path search in three-dimensional multi-wind zone environments, reducing planning in areas with strong headwinds or frequent turns.

[0088] 5. Fuel Consumption Optimization: "Single-segment fuel consumption" is used as the core cost indicator in the tree search and rerouting process. This enables the path to proactively avoid high fuel costs and effectively utilize environmental conditions such as tailwinds in windy areas, achieving a balance between fuel economy and wind speed adaptability. Specifically, this includes:

[0089] Consider the parent node q p With child node q c Factors such as the distance d between them, the turning angle Δθ, and the difference in elevation gain / loss are used to determine the base fuel quantity: , where k cruise The cruise fuel consumption coefficient is d, where d is the parent node q. p With child node q c The distance between them, k turn The turning fuel consumption coefficient is given by Δθ, where Δθ is the turning angle and k is the turning angle. climb It is the fuel consumption reduction coefficient for climbing, k descend It reduces the fuel consumption coefficient.

[0090] In three-dimensional coordinate space, the region is divided into several wind zones according to different wind directions and speeds, and each wind zone corresponds to a steady wind vector w=(w x ,w y ,w z When the improved RRT* algorithm finds a node q, it determines the wind zone it is in based on its coordinates and obtains the corresponding wind vector. When the wind vector at the parent node's location is w, its magnitude is... Let f be the unit vector of the flight direction, and θ be the angle between them. Then we have If θ is small (tailwind), W i >0; if θ is close to π (headwind), W i <0. Therefore, the wind correction factor is defined. .

[0091] Combining the base fuel quantity and the corrected fuel consumption considering the wind field effect, the final single-segment fuel consumption function is obtained as follows: When W i >0 (tailwind) will appropriately reduce fuel consumption; W i When the value is less than 0 (headwind), the consumption will increase, and it is encouraged to improve the RRT* algorithm to avoid areas with severe headwinds as much as possible.

[0092] The improved RRT* algorithm of this invention adds the single-segment fuel consumption of edgeFuel to the cost of the parent node to obtain the cost of the child node. During rewiring, it compares the total cost of each candidate parent node to determine the optimal parent node, iterating continuously until a complete path is obtained. After obtaining the complete path, the improved RRT* algorithm performs a comprehensive calculation of fuel and wind speed fitness for the entire path, comparing the merits of different paths (corresponding to an individual in the evolutionary algorithm).

[0093] Step (4): Perform population optimization to obtain the optimal target point allocation scheme.

[0094] Based on the individuals generated in step (2) of the fitness function evaluation, the three individuals with the highest fitness are selected as excellent individuals. While retaining the excellent individuals, new individuals are generated through crossover and mutation operations to gradually optimize the population. The crossover operation includes: selecting two parent individuals and generating new offspring individuals by exchanging some gene information. The mutation operation includes: randomly changing a gene of an individual to introduce a new solution space and avoid the algorithm getting trapped in local optima.

[0095] After multiple iterations of the evolutionary algorithm, the optimal fitness value of the population tends to stabilize, the evolutionary algorithm terminates, and the optimal target allocation scheme is output. The optimal target allocation scheme is the optimal sequence of target points assigned to each UAV, effectively distributing multiple target points to multiple UAVs.

[0096] Step (5): Calculate the preliminary flight path for each UAV using the improved RRT* algorithm, followed by coupling optimization and smoothing.

[0097] The obtained preliminary flight paths are then subjected to coupling optimization and smoothing processes, including path stretching and cubic B-spline interpolation. Path stretching involves checking non-adjacent nodes in the path pair by pair; if a pair of non-adjacent nodes satisfies a collision-free direct connection, the intermediate node of that pair is removed. Cubic B-spline interpolation uses a set of control points and piecewise cubic polynomial basis functions to generate a smooth curve based on the path obtained after path stretching, ultimately yielding flight path plans for multiple UAVs within the same map.

[0098] Example 2

[0099] To comprehensively evaluate the performance of multi-UAV multi-target path planning, the following four path planning algorithms were selected for comparison: (1) RRT: Random Fast Extended Tree Algorithm; (2) RRT*: Added parent node reselection and rewiring operation on the basis of RRT; (3) Adaptive Step-size RRT*: Compared with the basic RRT*, some changes were made in the sampling strategy or local optimization, which can improve the path quality or convergence speed to a certain extent, but multiple optimization mechanisms have not yet been introduced; (4) The RRT* designed and improved in this invention: introduced multiple improvements such as target bias, adaptive step size, potential field gravity, Quick-RRT* ancestor search and path post-processing.

[0100] Each experiment used a fixed set of 3 drones, 20 target points, and the same obstacle and wind conditions, and was repeated 8 times to observe the algorithm's stability and overall performance. The comparison of the eight metrics for the four algorithms across the eight runs is presented below. Figure 4 As shown. The main experimental parameters are set as follows:

[0101] A. Total Path Length: The total path length is obtained by summing the lengths of all trajectories after the three drones have completed their assigned targets. A smaller value generally indicates higher overall efficiency. The performance of the four algorithms on this metric is as follows: Figure 4 As shown in (a).

[0102] B. Average Path Length: The total path length is divided by the number of drones to measure the average path cost in multi-drone collaboration. The performance of the four algorithms on this metric is as follows: Figure 4 As shown in (b).

[0103] C. Longest Flight Path per Drone: A drone may be assigned to distant or numerous targets, causing its flight path to far exceed that of other drones. This value reflects the situation of the "worst performer" and also reflects whether the allocation is balanced. The performance of the four algorithms on this indicator is as follows: Figure 4 As shown in (c).

[0104] D. Allocation Algorithm Fitness Value: In this algorithm, it is defined as the negative of the sum of total path length and fuel consumption. A larger value (i.e., a smaller absolute negative value) indicates better performance. The performance of the four algorithms on this metric is as follows: Figure 4 As shown in (d).

[0105] E. Total Number of Nodes: The sum of the number of path nodes used by all drones. The performance of the four algorithms on this metric is as follows: Figure 4 As shown in (e).

[0106] F. Number of Planning Iterations: The actual number of iterations performed during path planning, reflecting the algorithm's complexity or convergence cost. The performance of the four algorithms on this metric is as follows: Figure 4 As shown in (f).

[0107] G. Route Planning Fuel Consumption: This metric calculates the total fuel cost required to complete the planned route. A smaller value indicates lower energy consumption under the same environmental conditions and initial fuel load assumptions. The performance of the four algorithms on this metric is as follows: Figure 4 As shown in (g).

[0108] H. Path Planning Wind Fitness: The fitness value obtained after incorporating the influence of the wind field can approximately reflect the degree of conformity or safety of the flight path with the wind field. A higher value (closer to or greater than 0) indicates better wind resistance or utilization. The performance of the four algorithms on this indicator is as follows: Figure 4 As shown in (h).

[0109] The specific data from the above experiments are recorded in Tables 1 to 4.

[0110] Table 1. Results of the improved RRT* algorithm of this invention

[0111] experiment Total path length Average path length Longest section of road for a single machine fitness value Total number of nodes Planning iteration count Fuel consumption wind adaptability 1 722 240 339 -776 24 22 54 -0.74 2 705 235 319 -754 23 22 48 -0.32 3 730 243 272 -784 24 23 53 -0.68 4 746 248 460 -801 23 23 54 -0.75 5 747 249 334 -804 24 26 57 -0.66 6 737 245 309 -790 23 22 53 -0.82 7 765 255 280 -819 23 22 54 -0.76 8 639 213 337 -683 23 22 43 0.11

[0112] Table 2 Results of the Adaptive Step-size RRT* algorithm

[0113] experiment Total path length Average path length Longest section of road for a single machine fitness value Total number of nodes Planning iteration count Fuel consumption wind adaptability 1 811 270 346 -880 44 47 69 -2.22 2 757 252 322 -818 45 43 61 -1.05 3 703 234 306 -757 46 45 54 -1.05 4 747 249 484 -807 44 41 59 -1.52 5 752 250 450 -814 43 47 62 -0.54 6 788 262 409 -850 44 46 61 -1.32 7 703 234 289 -760 47 64 56 -1.06 8 722 240 440 -783 45 42 61 -0.60

[0114] Table 3 Results of the basic RRT* algorithm

[0115] experiment Total path length Average path length Longest section of road for a single machine fitness value Total number of nodes Planning iteration count Fuel consumption wind adaptability 1 860 286 465 -952 46 117 92 -1.37 2 772 257 287 -856 45 121 84 -1.18 3 806 268 333 -877 44 129 71 -1.75 4 898 299 442 -985 44 121 86 -1.05 5 807 269 406 -879 44 114 72 -1.17 6 728 242 267 -801 43 156 72 -1.18 7 903 301 467 -981 43 107 77 -0.96 8 862 287 360 -946 45 163 83 -1.61

[0116] Table 4 Results of the basic RRT algorithm

[0117] experiment Total path length Average path length Longest section of road for a single machine fitness value Total number of nodes Planning iteration count Fuel consumption wind adaptability 1 1031 343 483 -1156 71 151 125 -1.71 2 958 319 444 -1077 71 153 119 -1.16 3 961 320 692 -1117 77 162 155 -1.12 4 994 331 419 -1101 66 125 107 -1.48 5 1027 342 576 -1162 72 114 135 -1.02 6 977 325 506 -1108 70 136 130 -1.17 7 977 325 561 -1110 74 132 133 -1.03 8 1063 354 491 -1172 65 122 108 -1.79

[0118] A bar chart showing the average values ​​of the eight metrics for the four algorithms across eight runs, for example. Figure 5 As shown, the top of the three bar charts displays the percentage improvement of the improved RRT* algorithm compared to the other three algorithms. The improved RRT* algorithm disclosed in this invention was evaluated based on eight metrics, leading to the following conclusions:

[0119] The optimal RRT* algorithm achieves a value of 639 (8th iteration), which is significantly better than other algorithms. The optimal value of the Adaptive Step-size RRT* algorithm is approximately 703 (3rd and 7th iterations), the optimal value of the basic RRT* algorithm is 728 (6th iteration), and the optimal value of the basic RRT algorithm is 958 (2nd iteration).

[0120] like Figure 5As shown in (a), the improved RRT* algorithm has a lower average total path length than the other three algorithms, with differences of 3.20%, 12.74%, and 27.49%, respectively. This demonstrates that the algorithm can effectively reduce flight path distance, thereby reducing flight time and energy consumption, and exhibits significant optimization advantages.

[0121] like Figure 5 As shown in (b), the improved RRT* had the lowest average path length of 213 in the 8th experiment, demonstrating excellent multi-machine allocation performance. The lowest average path lengths of the Adaptive Step-size RRT*, the base RRT*, and the base RRT were 234, 242, and 319, respectively, with the improved RRT* showing an overall advantage.

[0122] In terms of average path length, the improved RRT* algorithm also demonstrates advantages of 3.20%, 12.74%, and 27.49%. By effectively shortening the flight distance of each segment, the algorithm reduces the number of consecutive turns, further improving mission completion efficiency. This fully demonstrates its superior performance even at the detailed level.

[0123] like Figure 5 As shown in (c), the longest single-unit length of the improved RRT* remained between 300 and 400 in most experiments (slightly higher or lower in some cases), with a minimum value of 272 (3rd experiment). The longest single-unit length of the Adaptive Step-size RRT* exceeded 450 in some experiments; the basic RRT* also recorded lengths of around 465 and 470; and the basic RRT even showed high values ​​such as 692 (3rd experiment) and 576. This indicates that the improved RRT* is more balanced and less prone to the problem of excessively long "worst-case" segments.

[0124] For the longest single-aircraft segment, the improved RRT* algorithm showed significantly lower average results than other methods, with differences of approximately 12.93%, 12.44%, and 36.42%. This means that the algorithm can better avoid excessively long segments, helping to improve path controllability and reduce flight risks, with particularly obvious advantages.

[0125] It can be seen that the improved RRT* reaches -683 in the optimal case, the optimal Adaptive Step-size RRT* is -757, the optimal base RRT* is -801, while the base RRT values ​​are generally below -1000. The larger the value (the smaller the absolute value of the negative value), the better the overall path. Therefore, the fitness distribution of the improved RRT* reflects its better path optimization ability.

[0126] like Figure 5As shown in (d), the improved RRT* algorithm has a higher average fitness value, exceeding the other algorithms by 3.97%, 14.64%, and 30.99%, respectively. The higher fitness indicates that it has better overall navigation capabilities and stronger adaptability, further demonstrating its planning efficiency.

[0127] The improved RRT* typically uses 23 or 24 nodes in each experiment, with the optimal value being 23. The structure is very streamlined, and the fewer nodes indicate that the improved RRT* search tree generation is more efficient and has less redundancy.

[0128] like Figure 5 As shown in (e), the improved RRT* algorithm has a lower average number of nodes than other algorithms, with differences of 47.77%, 47.18%, and 66.96%. The reduction in the number of nodes significantly reduces the computational burden while maintaining path accuracy, indicating its excellent performance in node utilization efficiency.

[0129] like Figure 5 As shown in (f), the improved RRT* typically requires between 22 and 26 iterations, with an optimal value of 22, demonstrating extremely fast convergence. The Adaptive Step-size RRT* requires over 45 iterations in some experiments, with an optimal value of 41. The basic RRT* commonly requires over 100 iterations, with an optimal value of 107. The basic RRT even reaches over 100 iterations, sometimes exceeding 150, with an optimal value of 114. Compared to the basic RRT*, the improved RRT* reduces iterations by approximately 80% or more; it also shows a significant reduction compared to the Adaptive Step-size RRT*.

[0130] Regarding the total number of planning iterations, the improved RRT* algorithm also has a lower average level, with differences from other algorithms remaining at 51.47%, 82.30%, and 83.38%. This significantly reduces computation time and hardware resource consumption, providing more room for system stability and real-time planning.

[0131] like Figure 5 As shown in (g), the improved RRT* generally consumed fuel in the range of 43 to 57 in multiple rounds of experiments, with the optimal value being 43 (8th time), indicating that it has a shorter path and fewer detours; the Adaptive Step-size RRT* was mostly between 54 and 69, with the optimal value being 54; the basic RRT* was mostly around 71 to 92, with the optimal value being 71; while the basic RRT in some experiments reached as high as 125 or 155, with the optimal value being 107; indicating that the improved algorithm can effectively reduce energy consumption, which is particularly important for actual task requirements.

[0132] In terms of fuel consumption, the improved RRT* algorithm averaged lower than the other three methods, with differences of approximately 13.42%, 34.29%, and 58.56%, respectively. This reduction in fuel consumption not only translates to lower economic costs but also reduces flight load, further demonstrating the algorithm's energy-saving characteristics.

[0133] like Figure 5 As shown in (h), the improved RRT* only showed a positive value (0.11) in the 8th experiment, while the values ​​in other experiments were negative but small (-0.32 to -0.82). The Adaptive Step-size RRT* and the basic RRT* and basic RRT all showed lower fitness values ​​such as -1.0 or even -2.2 in many experiments, which means that the flight path may not be wind-friendly. Therefore, the improved RRT* can maintain better wind fitness in most cases, and in some cases (such as the 8th experiment), it can even achieve positive fitness with tailwind or lower wind resistance.

[0134] In terms of wind adaptability, the improved RRT* algorithm scored higher on average, with differences of approximately 50.42%, 54.78%, and 55.71% compared to other algorithms. A higher score indicates that it can maintain a more stable flight trajectory when facing complex wind fields, achieving better environmental adaptability.

[0135] In the experiment, the planning results of each algorithm were visualized and compared in three dimensions to verify the consistency between the numerical indicators and the actual flight path. Figures 6 to 9 The trajectory layouts of the four algorithms under the same initial and target settings are presented respectively. Figures 6 to 9 These are the basic RRT planning results, basic RRT* planning results, Adaptive Step-size RRT* results, and improved RRT* planning results, respectively. In each figure, (a) represents the trajectory before B-spline processing, and (b) represents the trajectory after B-spline processing. The red, green, and blue trajectories represent the path planning results of three different UAVs. Other legends are... Figure 2 Same. (Through comparison) Figures 6 to 9 The results can be observed Figure 9 The improved RRT* algorithm demonstrates its comprehensive advantages in obstacle avoidance, reducing redundant turns, and shortening the total mileage.

Claims

1. A method for multi-UAV path planning in a multi-wind-zone environment, characterized in that, The application relates to a method for planning a flight path of a plurality of unmanned aerial vehicles (UAVs) in a multi-wind region, comprising the following steps: establishing a three-dimensional map of a UAV task region, setting a plurality of UAVs in the three-dimensional map, setting a plurality of start points and a plurality of target points; assigning a start point to each UAV; different positions of the three-dimensional map are provided with different wind vectors w, including wind direction and wind speed; using an evolutionary algorithm to assign target points, setting a target point assignment scheme as an individual of the evolutionary algorithm, using an improved RRT* algorithm to generate a feasible flight segment according to each individual, and setting a fitness function of the evolutionary algorithm as a negative value of a sum of path length and fuel consumption of the feasible flight segment; the evolutionary algorithm is terminated when the fitness function converges, and an optimal target point assignment scheme is obtained; the fuel consumption is a fuel consumption function with a wind correction factor, and the wind correction factor is related to the wind vector w; according to the optimal target point assignment scheme, the improved RRT* algorithm is used to generate flight paths of all UAVs from start points to target points, and path stretching and cubic B-spline interpolation are performed on the flight paths, so that the flight path planning of the plurality of UAVs in the multi-wind region is realized. The three-dimensional map is provided with different wind directions and wind speeds, comprising the following steps: dividing the UAV task region into a plurality of wind regions according to different wind directions and wind speeds, and corresponding to a wind vector w; when the improved RRT* algorithm searches the three-dimensional map, the wind region where a current position is located is determined according to coordinates, and the corresponding wind vector w is obtained. The method for assigning target points by using the evolutionary algorithm comprises the following steps: setting a target point assignment scheme as an individual of the evolutionary algorithm, and each individual is a one-dimensional array; the jth element in the array represents that the jth target point is assigned to the ith UAV. The improved RRT* algorithm comprises the following steps: (1) setting start points and target points according to the individual, initializing a tree of a search space, and each node contains coordinates (x, y, z), parent node information and cumulative cost; 2.The multi-UAV route planning method of claim 1, wherein, (2) using a target bias sampling method to obtain a sampling point, finding a nearest node of the sampling point in the tree, making the nearest node move in a potential field attractive direction by a step length to obtain a new node; the potential field attractive force is a linear combination of a direction from the nearest node to the target point and a direction from the nearest node to the sampling point; after each sampling is completed, a function with an inertia factor is used to update the step length; (3) extracting nodes within a certain distance range of the new node into an expansion set, finding ancestor nodes of each candidate node in the expansion set and adding the ancestor nodes into the expansion set; 3. The route planning method according to claim 2, characterized in that, (4) detecting collision conditions of each candidate node and a line connecting the new node one by one from the expansion set, calculating a cost, and selecting a candidate node with the lowest cost as a parent node of the new node; 4. The route planning method according to claim 3, characterized in that, (5) repeating steps (2) to (4) until the sampling point is the target point, and generating a feasible flight segment according to all nodes in the tree. The target bias sampling method is expressed as: wherein beta is an adjusting parameter, and beta>0. The fitness function comprises the following steps: The method for assigning target points by using the evolutionary algorithm comprises the following steps: evaluating individuals in the evolutionary algorithm according to the fitness function, screening out three excellent individuals with the highest fitness, generating new individuals through crossover and mutation operations on the basis of the excellent individuals, and gradually optimizing the population. ​ 5. The route planning method according to claim 4, characterized in that, ​ In the formula For sampling points, For the target point, Let be the target bias probability, when In the case of sampling, the sampling point is the target point; otherwise, the global range is maintained. Random sampling is implemented within the facility.

6. The route planning method according to claim 5, characterized in that, The cost is a fuel consumption function that takes into account a wind correction factor is expressed as: wherein is the base fuel quantity: k in the above equation cruise is a cruising fuel consumption coefficient, d is the distance between the parent node q p and the child node q c , k turn is a turning fuel consumption coefficient, Δθ is the turning angle, k climb is a climbing fuel consumption coefficient, k descend is a descending fuel consumption coefficient; The position of the parent node is a wind vector w, the angle between the parent node and the child node flight direction unit vector is θ, and a wind fitness correction is added , to obtain a wind correction factor : ​ 7. The route planning method according to claim 6, characterized in that, ​ The whole feasible voyage section is divided into N sections, the length of the ith section is , is the wind effect weight factor of the ith section, and the wind speed fitness index W is expressed as: Let f climb / desend f for climbing / descending fuel consumption turn f for cornering fuel consumption cruise f for cruising fuel consumption, fuel consumption is expressed as: The fitness function W 总 The sum of the total path length path length and the fuel consumption of the feasible flight segments The negative value of the sum of the total path length path length and the fuel consumption of the feasible flight segments, is represented as: W 总 = - (path length + F total ).

8. The route planning method according to claim 7, characterized in that, ​ ​ Two parent individuals are selected to generate a new child individual by exchanging part of the genetic information; a certain gene of a certain individual is randomly changed to introduce a new solution space and avoid the algorithm from falling into a local optimum; The above process is repeated until the optimal solution of the fitness function tends to be stable after multiple iterations.

9. The route planning method according to claim 8, characterized in that, The path stretching includes: checking each pair of non-adjacent nodes in the flight path, and removing the intermediate nodes of the pair of non-adjacent nodes if the pair of non-adjacent nodes meet the collision-free direct connection.

10. The route planning method according to claim 9, characterized in that, The cubic B-spline interpolation generates a smooth curve based on the path obtained after path stretching through a set of control points and a segmented cubic polynomial basis function.

Images