A method for regional coverage of starling drone cluster cooperative route obstacle avoidance

Abstract

The application discloses a kind of for regional coverage's imitation European starling unmanned aerial vehicle cluster cooperative route obstacle avoidance method, including the following implementation steps: step one: given region is divided into corresponding number of sub-region;Step two: planning closed path in sub-region;Step three: based on neighbor selection method and random line of sight obtainable interactive unmanned aerial vehicle;Step four: obstacle avoidance function based on distributed unmanned aerial vehicle cluster control method;Step five: output route obstacle avoidance planning result.The application simultaneously considers external security and cluster internal security, improved collision cone obstacle avoidance method and limited cluster interaction method are used, and distributed unmanned aerial vehicle cluster control framework is generated.In addition, a new type of cooperative coverage method is proposed to solve the problem of multi-unmanned aerial vehicle sustained coverage with complex constraints, effectively minimizing the difference between each unmanned aerial vehicle route.

Description

Technical Field

[0001] This invention is a collaborative obstacle avoidance method for a swarm of European starling-inspired UAVs used for regional coverage, belonging to the field of autonomous UAV control. Background Technology

[0002] Unmanned aerial vehicles (UAVs) are widely used in surveying, transportation, and resource exploration. In recent years, UAV swarms and swarm behavior have received increasing attention, and herd behavior in nature, such as that of birds and fish, has been extensively studied. Compared to flocks of pigeons and geese, starling flocks are enormous, with individual swarms reaching thousands of individuals, exhibiting highly ordered behavior and rapid responses to external stimuli. A single starling may only react to information from 7-8 of its neighbors, yet hundreds to tens of thousands can circle and tumble together, spontaneously and synchronously avoiding predator attacks. This invention maps the starling flock behavior mechanism to UAV swarm control, utilizing a topological distance neighbor selection method to restrict interaction between individuals to a fixed number of objects. To simultaneously consider external and internal swarm security, this invention employs obstacle avoidance and limited swarm interaction methods. Since most organisms in nature have blind spots in their vision, a limited perspective is introduced into the neighborhood selection method to generate a distributed UAV swarm control framework, thereby enhancing the coordination and flexibility of UAV swarm control. Related research indicates that swarms can be guided by incorporating informed individuals. Therefore, to better control and guide autonomous swarms without compromising their distributed nature, some drones can be configured as informed drones with the desired orientation.

[0003] In route planning, multi-UAV cooperative coverage has wide applications in many fields. The goal of the multi-UAV coverage problem is to ensure continuous coverage of a target area using a group of UAVs with limited sensing capabilities. To ensure efficient coverage, many factors need to be considered, such as the overlap between different areas, the difference in path length between each UAV, the coverage time and coverage cycle when achieving full coverage. In addition, irregular obstacles and task areas increase the complexity of the problem. Compared with a single UAV, using multiple UAVs for cooperative coverage has many advantages. Due to the division of workload, using multiple UAVs can significantly reduce the time cost of completing the coverage task, and a single UAV can further utilize its teammates to reduce positioning errors. Furthermore, using multiple UAVs can improve robustness, because the failure of some members in the cluster can be compensated by other members. Graph-based methods and spanning tree-based methods cannot handle coverage problems in complex environments, while clustering-based methods do not take into account the limitations of UAVs and practical applications. Therefore, this invention proposes a novel cooperative coverage method to solve the multi-UAV continuous coverage problem with complex constraints. First, sensor areas are deployed, then a clustering method is used to divide the area, and finally, the route of each UAV can be calculated by solving the Traveling Salesman Problem. The main advantages of the strategy proposed in this invention are reflected in the following three aspects:

[0004] (1) Traditionally, sensor deployment and region segmentation problems are separated, resulting in low efficiency in problem-solving. This invention combines the sensor deployment and region segmentation problems. Through this combination, the region segmentation problem is transformed into a clustering problem. Furthermore, traditional clustering-based methods using adjacent square grid nodes are cumbersome, while the sensor nodes used in this invention are more adaptable and practical.

[0005] (2) Existing route planning methods mainly consider the optimal or shortest path, but the route lengths of different UAVs may vary significantly. Drawing on control theory, this invention introduces a feedback mechanism and weighted distance. Through an improved K-means clustering method, the differences between the routes of each UAV can be effectively minimized.

[0006] (3) Although traditional control methods can guarantee the convergence of the swarm, the final direction and speed of the swarm are randomly determined. This invention introduces an informed UAV guidance mechanism to ensure that the UAV swarm moves at the expected speed and in the desired direction. Summary of the Invention

[0007] The purpose of this invention is to provide a collaborative obstacle avoidance method for starling-inspired UAV swarms for regional coverage. This method aims to solve the problems of regional division and sensor allocation in route planning, and achieves multi-UAV collaborative obstacle avoidance and regional coverage search by introducing an informed UAV guidance mechanism and an obstacle avoidance mechanism based on the collision cone method.

[0008] This invention is a collaborative obstacle avoidance method for a swarm of European starling-inspired UAVs for regional coverage. The specific implementation steps are as follows:

[0009] Step 1: Divide the given region into a corresponding number of sub-regions.

[0010] (1) Cover a given area

[0011] The target area is divided into a series of square grids, each with a side length of l. c Much smaller than the sensor's detection radius r s Define C f ={C1,C2,...,C m} represents the reachable area R a A set of square grids. Let S = [S1, S2, ..., S...]. n Let be the set of sensors, where n is the maximum number of sensors. This area must be completely covered by the sensors, i.e., any square C... i All four vertices are at sensor S j Within the detection range, and the area is free of obstacles.

[0012] (2) Design the fitness function

[0013] Let e i =[p(S i ),g i ] represents specific information about the deployment of sensor i, p(S) i ) represents the sensor position, g i For a valid marker, g i =1 indicates S i Deployed. Therefore, the vector [e1,e2,e3,...,e] is... n The ] represents one possible sensor deployment scenario, which is a feasible solution to the sensor deployment problem. Based on the sensor deployment requirements, the fitness function for this problem is designed as follows:

[0014] f([e1,e2,...,e n ])=λ1×n d +λ2×n u +λ3×n o (1)

[0015] Where n d n represents the number of sensors in the obstacle area. u n represents the number of areas in the reachable region not covered by sensors. o This represents the number of overlapping areas covered by sensors in the reachable region, with weights λ1>>λ2>>λ3>0. The goal is to find the optimal solution that minimizes the fitness function value.

[0016] An improved cooperative co-evolutionary particle swarm optimization (CCPSO2) algorithm is used to update sensor deployment locations to solve the sensor deployment problem. In the CCPSO2 algorithm, particle positions are updated using methods based on Cauchy and Gaussian distributions, increasing particle diversity and improving algorithm performance. The algorithm decomposes the n-dimensional solution vector into K components, each corresponding to an s-dimensional particle (e1, e2, ..., e...). s A group of s, where s = n / K. Assume P j ·x i and P j ·y i These are the current position of the particle and the local optimal position of particle i in the j-th cluster, respectively. This represents the historical global best position of cluster j.

[0017] Local optimal position P j ·y i and global optimal position The updated formula is as follows:

[0018]

[0019] Where f(·) represents the fitness function, and the vector b(j,P) j ·x i ) represents a vector

[0020]

[0021] The formulas for updating the speed and position of cluster j are as follows:

[0022]

[0023] P j ·x i (t+1)=P j ·y i (t)+c3|P j ·y i (t)-P j ·y' i (t)| (3)

[0024]

[0025] Where ω(t) is the inertia coefficient balancing the global and local searches, r1(t) and r2(t) are random numbers on [0,1], c1, c2, and c3 are acceleration constants, C(1) and N(0,1) are random numbers generated by Cauchy and Gaussian distributions, respectively, and P j ·v i Let P represent the velocity of the j-th cluster. To maintain particle diversity and avoid premature convergence to a local optimum, we use the local neighborhood optimum P.j ·y' i Replace global optimum Using ring topology to describe the neighborhood of a particle, therefore, P j ·y' i It can be described as:

[0026]

[0027] Based on the improved CCPSO2 algorithm described above, the fitness function is minimized to obtain the optimal deployment vector.

[0028] (3) Region partitioning based on K-means method

[0029] After solving the sensor deployment problem, a set of sensor deployments is obtained. Here, the K-means method is used to cluster the sensor deployment points. The basic workflow is as follows:

[0030] S1. Select initial cluster centers C = {C1, C2, ..., C...} k};

[0031] S2. Calculate the distance D(x,c) between data point x and cluster center cj. j This assigns data points to the nearest cluster.

[0032] S3. Recalculate the center of each cluster;

[0033] S4. Repeat S2 and S3 until the criterion function converges or the cluster centers no longer change.

[0034] This method typically uses the squared error criterion function as the criterion function, which can be expressed as:

[0035]

[0036] Where k is the number of clusters, n i Let x be the number of data points in the i-th cluster. ij For the j-th data point in the i-th cluster, m i Let be the mean of all data points in the i-th cluster.

[0037]

[0038] Traditional K-means methods typically use randomly selected initial cluster centers, which can lead to significant differences between clusters and sometimes even fail to converge. Therefore, a density-based method is used to select the initial cluster centers. The density of sensor point x in S is defined as the number of sensor points whose distance from x is not greater than a specific neighborhood radius r. For the density-based method, a suitable neighborhood radius r needs to be selected first to calculate the density D of each point. i(r), and take the data point with the highest density as the initial center c1 of the first cluster; then, delete the points within the radius r of the c1 domain, and take the point with the highest density among the remaining data points as the initial center of the second cluster, and so on to derive the initial centers of all clusters.

[0039] This invention achieves this by adjusting the distance D(x,c) j The K-means method is improved by using weighted distance instead of Euclidean distance. ij To cluster center c i The distance metric between them is defined as:

[0040] D i (x ij ,c i ,ω i )=||x ij -c i || / ω i (8)

[0041] ω i For cluster center c i The distance weights. Therefore, the corrected criterion function is:

[0042]

[0043] The distance weight ω needs to be adjusted after each iteration. i This ensures that the path lengths of each sub-region tend to be consistent. This means that if the search path length of a sub-region is greater than the average length, the distance weight ω... i It needs to be reduced.

[0044] Assumption Let be the shortest closed path length through all data points in the i-th cluster. If the average value is the shortest path length, then the deviation of the path length is:

[0045]

[0046] Therefore, the distance weight ω i Can be updated to:

[0047] ω i =ω i0 +Δω i =ω i0 -k1Δl i (11)

[0048] In the formula ω i0 Let k be the distance weight at the final moment, and k1 be the feedback gain, satisfying k1>0. The algorithm terminates when the absolute value of the maximum route length deviation is not greater than a certain threshold ε; otherwise, it continues to iterate.

[0049] Step 2: Plan closed paths within the sub-region

[0050] After obtaining the clusters, solve the traveling salesman problem to obtain the paths through all sensor points in the clusters. Under the premise of complete coverage, the goal is to find the shortest possible path.

[0051] If there are no obstacles between the two sensors, calculate the Euclidean distance directly; if there are obstacles, use the A* algorithm to solve for the distance. Let L... i ·=[l i1 ,l i2 ,…,l ij ,…,l im ] is the distance set, l ij Indicates sensor s i and s j The distance between them is described in detail below:

[0052]

[0053] For the sensor deployment point set S m =[s1,s2,...,s m This yields an m×m total distance matrix D. m×m =[L1,L2,…,L m ·]. Let sq i (s1,sm)=[s a ,s b ,…,s c ], a,b,c∈[1,m] is the sequence of sensor deployment points, where s· is s m The sensor deployment points. Therefore, with s m and D m×m Given the conditions, the Traveling Salesman Problem can be solved, yielding the desired solution sq within the solution region. An improved genetic algorithm can be used to solve this problem, and the fitness function can be defined as:

[0054] f t (sq i )=1 / l(sq i (13)

[0055] In the formula l(sq i ) represents sq i The length.

[0056] Step 3: Acquire Interactive Drones Based on Neighbor Selection Method and Random Line of Sight

[0057] This invention aims to construct a distributed unmanned aerial vehicle (UAV) swarm control framework, enhance the swarm's speed consistency and rapid response capabilities during maneuvering, and designs an information interaction method between UAVs, inspired by the behavior mechanism of starling flocks.

[0058] (1) Neighbor selection method

[0059] To ensure safe flight of large-scale drone swarms, each drone should maintain a reasonable distance from its neighbors. Neighbor interaction is a key behavior for drone swarms to achieve collision avoidance or self-organizing planning. Starlings in nature typically interact with six or seven of their nearest and most visible neighbors. By mimicking the swarm interaction behavior of starlings, this invention proposes a finite interaction method based on distributed control. Figure 1 As shown, starlings can perceive their neighbors within a certain range by detecting the edges of shadows. Due to this perceptual characteristic, individuals can even interact with unobstructed individuals far from the center. Therefore, starling flocks possess high mobility and flexibility. Meanwhile, biological visual systems are limited by line of sight and field of view in most cases. The field of view can be defined as the angle β between the boundaries of the field of view. Under the influence of limited field of view and occlusion relationships between individuals… Figure 1 The central individual in the diagram can only interact with individuals within the shaded area.

[0060] (2) Random line of sight

[0061] Due to the limited field of view, it's impossible to consider all individuals in the vicinity during flight. For drone swarms, if each drone uses a fixed, limited field of view (i.e., its line of sight is fixed), it may lead to swarm dispersion in certain extreme cases. Referring to the behavior of large flocks of birds, birds can increase their sensing range by turning their heads, thus achieving better interaction.

[0062] A non-fixed field of view can be achieved using the random line-of-sight method. The essence of this method is to set the angle between the direction of movement and the line of sight as a random variable. Based on the range of head-turning angles of most birds, the angle between the line of sight and the direction of movement is set to 80°–180°. This invention, based on neighbor selection and the random line-of-sight method, applies a finite swarm interaction method to UAV swarm control. Let the angle α between the line of sight and the direction of movement be a uniformly distributed random variable, calculated as follows:

[0063] α~U[-80°, 80°] (14)

[0064] Where U[·] is the probability density function of a uniform distribution. For example... Figure 2 As shown, each drone can only interact with other drones within a limited area.

[0065] Step 4: Obstacle avoidance function based on distributed UAV swarm control method

[0066] (1) Artificial potential field

[0067] Define three spatial spheres centered on the UAV. The three spherical regions, from the inside out, are the repulsion region, alignment region, and attraction region, with radii r and r, respectively. rep r ali and r att These areas are adjusted by a limited field of view. To ensure a reasonable safe distance between drones, individuals should control their movement away from neighbors in the exclusion zone. Simultaneously, to prevent swarm fragmentation, individuals should control their movement closer to neighbors in their attraction zone. Furthermore, to achieve cooperative movement within the group, individuals must adjust their speed to align with the speed of their neighbors within the alignment zone.

[0068] Assuming that under the designed neighbor selection method, drones in the cluster can obtain necessary information, including their own and other drones' states, through their own sensors. Assume there are N locations... speed is There are drones, each of which is influenced by its neighboring drones. The interaction force between drone i and its neighboring drones is... The calculation is as follows:

[0069]

[0070] in These represent the repulsive force, alignment force, and attractive force acting on drone i, respectively.

[0071] If the distance between drones is small, flight safety is lower, requiring a greater repulsive force. The flight speed of the drones also affects flight safety; if drone i or drone j is faster, the flight safety will be lower. Furthermore, the calculation of the repulsive force must consider the influence of the drone's heading. When drone j flies towards drone i, the repulsive force should be large enough to suppress this flight path. Therefore, the repulsive force can be calculated as follows:

[0072]

[0073]

[0074] in Indicates the speed of drone i. Indicates the position of drone i. Indicates the location of neighboring drone j. Γ i This represents the set of neighboring drones of drone i.

[0075] To achieve rapid response from a drone swarm, the alignment force should be greater the closer the drones are to each other. Considering the influence of drone heading, the required alignment force is minimized when two drones are traveling in the same direction. The alignment force is calculated as follows:

[0076]

[0077] Similarly, an attractive force is introduced to prevent the drone swarm from diverging. The formula considers factors including the distance between drones, their speed, and their direction of velocity. The attractive force is calculated as follows:

[0078]

[0079] The interaction force is designed to ensure the drone achieves the desired motion. Therefore, mapped to drone control, the interaction force should be used as the desired velocity input for the drone. For the i-th drone, its velocity control variable is:

[0080]

[0081] (2) Informed drone guidance mechanism

[0082] Under the neighbor selection and drone interaction mechanism, drones with initial random positions and velocities will eventually achieve velocity polarization. In an unrestricted space, the drone swarm will move in a random direction and at a random speed. When performing tasks, the drone swarm will typically encounter external obstacles or be attacked. When some drones sense danger, they should lead the entire swarm to make a quick turn to achieve effective obstacle avoidance. To ensure the drone swarm moves at the expected speed and direction, informed drones are added to our proposed distributed drone swarm control framework. These informed drones have a predicted preferred speed. Furthermore, flying along a pre-planned closed flight path, it can influence other drones through interaction mechanisms. As part of the entire drone swarm, the informed drone shares the same interaction rules as other drones. Simultaneously, informed drones should be included in the neighbor set of other individuals. Since each informed drone has its expected velocity and desired direction, its interaction force should maintain a balance between the interaction force and its preferred velocity. Therefore, the interaction force of informed drone k is calculated as follows:

[0083]

[0084] in The interaction force generated by each informed drone's neighbors is denoted by λ, where λ ∈ [0, 1]. If λ = 1, the informed drone only considers the expected speed, ignoring the safety of its neighbors and the group. If λ = 0, the informed drone is actually a regular drone.

[0085] When an informed drone is a neighbor of a regular drone i, the weight μ also needs to be considered when calculating the interaction force between the informed drone and drone i. The interaction force of drone i is calculated as follows:

[0086]

[0087] in This represents the force exerted by the informed drone j on drone i. The resultant force of the forces exerted on drone i by other drones in the neighbor set of drone i is represented by μ, where μ represents the weight ratio. if The informed drone accounts for the largest proportion of the interaction forces. If μ→1, the informed drone is actually regarded as an ordinary drone.

[0088] In step two, the present invention has completed path planning within the closed area. This route information will be sent to a drone in each area; these drones with defined routes are now known drones.

[0089] (3) Obstacle avoidance mechanism based on collision cone method

[0090] Obstacle avoidance can be achieved by adjusting the velocity direction outside the collision cone. A higher safety margin should be provided for both the obstacle zone and the safe zone, and the required obstacle avoidance speed should be tangential to the safe zone. To achieve a higher obstacle avoidance speed, the drone should move forward at its maximum speed in the expected direction. If multiple obstacles pose a collision risk, the nearest obstacle should be considered first.

[0091] Given the desired obstacle avoidance speed Group interaction The final distributed swarm velocity of drone i The calculation is as follows:

[0092]

[0093] Assuming each drone is equipped with a low-level autopilot in the inner-loop controller, a six-degree-of-freedom model is used to drive the motion of the drone swarm. The kinematic model of the fixed-wing drone is as follows:

[0094]

[0095] Where [x] i ,y i ,h i ], V i , ψ i and λ i V represents the inertial position, airspeed, heading angle, and rate of change of altitude of UAV i, respectively. i c , These are the command inputs for horizontal speed, heading angle, and rate of change of altitude, respectively. τ v , τ ψ τλ and τλ are the time constants for velocity, heading angle, and rate of change of altitude, respectively.

[0096] To simplify control inputs, the drone model can be based on distributed swarm velocity. Convert the three-axis components:

[0097]

[0098] For a realistic drone model, constraints should be considered. Command inputs should satisfy the following restrictions:

[0099]

[0100] Step 5: Output the route obstacle avoidance planning results

[0101] All UAVs performing the task are randomly deployed within the task area to obtain the initial sensor deployment positions. The fitness function value under the current deployment condition can be obtained from Equation (1). Based on the position update formulas of Equations (2) and (3), a better sensor deployment position is considered. Using the criterion function formula of Equation (9) and the distance weight update formula of Equation (11), the process is iterated multiple times until the absolute value of the maximum route length deviation is not greater than a certain threshold ε, thus completing the area division.

[0102] After obtaining the cluster division, the traveling salesman problem is solved to obtain the path through all sensor points in the cluster. Under the premise of complete coverage, the shortest path is required. The optimal closed path planning in each search area is calculated by Equations (12) and (13).

[0103] Each drone performs coverage tasks only within its designated area; therefore, the calculation of interaction forces between drones only considers drones within its own area. Based on the neighbor selection method in step three, and considering the random line-of-sight assumption, the set of neighboring drones for each drone at a given time is obtained.

[0104] After obtaining the neighboring drone set, based on the artificial potential field assumption, if the neighbor is in the repulsion zone, the repulsion force is calculated by equation (16); if the neighbor is in the alignment zone, the alignment force is calculated by equation (17); and if the neighbor is in the attraction zone, the attraction force is calculated by equation (18). The interaction force is designed to ensure that the drone achieves the expected motion. Therefore, when mapped to the drone control, the interaction force should be used as the desired speed input of the drone. The desired control input can be obtained from equation (19). If the drone is an informed drone, the formula for calculating its interaction force is equation (20), which aims to make the drone swarm move at the expected speed value and in the desired direction. If the drone is a regular drone, but its neighbor set includes informed drones, the formula for calculating its interaction force is equation (21), which aims to enhance the leadership role of informed drones in the drone swarm.

[0105] To achieve obstacle avoidance, an obstacle avoidance mechanism based on the collision cone method is adopted. The final distributed swarm velocity v of UAV i can be calculated from equation (22). c,i Based on the UAV model of Equations (23) and (24), the control input is calculated, and the closed flight path that completely covers the search area is output through simulation, while also having collision avoidance functions inside and outside the cluster.

[0106] The cooperative path avoidance method for starling-inspired UAV swarms used for regional coverage, as described in this invention, has the advantages of simultaneously considering external and internal swarm safety, employing an improved collision cone obstacle avoidance method and a limited swarm interaction method to generate a distributed UAV swarm control framework. Furthermore, a novel cooperative coverage method is proposed to solve the problem of continuous coverage for multiple UAVs under complex constraints, effectively minimizing the differences in routes between each UAV. Attached Figure Description

[0107] Figure 1 Neighbor selection method diagram

[0108] Figure 2 Schematic diagram of limited communication in a drone swarm

[0109] Figure 3 Flowchart of a collaborative route obstacle avoidance planning method for UAV swarm area coverage

[0110] Figure 4 Schematic diagram of instance area division and sensor deployment

[0111] Figure 5 Example route planning diagram

[0112] Figure 6 Instance cluster final state diagram

[0113] Figure 7 Instance Cluster Order Parameter Graph

[0114] Figure 8 Cluster obstacle avoidance diagram Detailed Implementation

[0115] The effectiveness of the proposed method is verified through a specific example below. In this example, 95 UAVs are used to perform coverage search and swarm obstacle avoidance over a 120km × 120km area. The simulation environment for this example is configured with an Intel i5 processor and MATLAB 2022a software.

[0116] A cooperative route avoidance planning method for UAV swarm area coverage, the implementation process of which is as follows: Figure 3 As shown, the specific practical steps for this example are as follows:

[0117] Step 1: Divide the given region into a corresponding number of sub-regions.

[0118] S11: Cover a given area

[0119] The area to be covered is divided into 3600 sections with sides of length l. c The sensor detection range is r, which is a 2km grid. s The range is 10km, and the number of drones, n, is 95. The initial deployment locations of the drones are random.

[0120] S12: Calculate the fitness function value and update the drone deployment location.

[0121] The initial deployment location of the UAV is known from step S11. Substituting this deployment vector into the formula,

[0122] f([e1,e2,...,e n ])=λ1×n d +λ2×n u +λ3×n o

[0123] Let λ1 = 500, λ2 = 10, and λ3 = 0.5. Calculate the fitness function values ​​for each UAV. Substitute these calculated fitness function values ​​into the position update formula.

[0124]

[0125] P j ·x i (t+1)=P j ·y i (t)+c3|P j ·y i (t)-P j ·y' i (t)|

[0126] The process is repeated multiple times until the fitness function value is minimized, at which point the optimal deployment vector is obtained.

[0127] S13: Region partitioning based on the K-means method

[0128] Based on the basic workflow of the K-means method, the density method is used to select the initial cluster centers, using the formula...

[0129]

[0130] As the criterion function, the distance weight ω is adjusted after each iteration. i This ensures that the path lengths of each sub-region tend to be consistent. If the search path length of a sub-region is greater than the average length, the distance weight ω is reduced. i Through the formula

[0131]

[0132] The path length deviation can be calculated. Iteration requires following the formula.

[0133] ω i =ω i0 +Δω i =ω i0 -k1Δl i

[0134] Update the distance weights. The algorithm terminates when the absolute value of the maximum route length deviation is no greater than a certain threshold ε; otherwise, it continues to iterate.

[0135] Step 2: Plan closed paths within the sub-region

[0136] After obtaining the clusters, the Traveling Salesman Problem is solved to obtain the paths through all sensor points in the clusters. Under the premise of complete coverage, the goal is to find the shortest possible paths. If there are no obstacles between two sensors, the Euclidean distance is calculated directly; if obstacles exist, the A* algorithm is used to solve for the distance. Specifically, the sensor points si and s are calculated. j The formula for the distance between them is:

[0137]

[0138] This invention utilizes an improved genetic algorithm to solve this problem, where the fitness function is defined as follows:

[0139] f t (sq i )=1 / l(sq i )

[0140] Step 3: Acquire Interactive Drones Based on Neighbor Selection Method and Random Line of Sight

[0141] S31: Neighbor Selection Method

[0142] like Figure 1 Considering the occlusion relationships between individual drones and the field of view of the drones, the interaction between a single drone and the swarm is somewhat limited. Based on the aforementioned neighbor selection mechanism, the set of neighboring drones corresponding to all drones at a given time is determined.

[0143] S32: Random line of sight

[0144] like Figure 2 Due to the limited field of view, it's impossible to consider all individuals in the surrounding area during flight. For drone swarms, if each drone uses a fixed, limited field of view (i.e., its line of sight is fixed), it may lead to swarm divergence in certain extreme cases. A non-fixed field of view can be achieved through a random line-of-sight method, where the angle between the direction of movement and the line of sight is set as a random variable α, and each drone can only interact with other drones within a limited area.

[0145] Step 4: Obstacle avoidance function based on distributed UAV swarm control method

[0146] S41: Artificial potential field

[0147] After obtaining the collection of neighboring drones, based on the artificial potential field assumption, if the neighbors are within the repulsion zone, the repulsion force is calculated using the following formula.

[0148]

[0149] If the neighboring property is within the alignment zone, the alignment force is calculated using the following formula.

[0150]

[0151] If the neighbor is within the attraction zone, the attraction is calculated using the following formula.

[0152]

[0153] The interaction force is designed to ensure that the drone achieves the intended motion. Therefore, when mapped to drone control, the interaction force should be used as the drone's desired velocity input.

[0154]

[0155] Based on the desired speed input obtained from the solution, the desired control input can be calculated.

[0156] S42: Informed Drone Guidance Mechanism

[0157] To ensure the drone swarm moves at the desired speed and direction, informed drones are introduced into each area. The interaction rules for informed drones are the same as for ordinary drones, and they should be included in the set of neighboring drones. The formula for calculating their interaction force is as follows:

[0158]

[0159] The impact of desired velocity on the motion of informed drones was considered, and the swarm motion was guided by informed drones. To enhance the leadership role of informed drones in the entire swarm, the interaction force calculation formula was used when an informed drone was in the neighbor set of a certain drone.

[0160]

[0161] Based on the closed route planning results obtained in step two, the route planning information is sent to a drone in each region. These drones are known as informed drones.

[0162] S43: Obstacle Avoidance Mechanism Based on Collision Cone Method

[0163] Given the desired obstacle avoidance speed and the calculated group interaction force By the following formula

[0164]

[0165] The final distributed swarm velocity of UAV i can be calculated.

[0166] Based on the corresponding UAV model, and considering actual limitations such as UAV speed, altitude change rate, and yaw rate, the final control input is calculated.

[0167]

[0168] Step 5: Output the route obstacle avoidance planning results

[0169] Based on the cluster control input calculated in the above steps, the simulation result diagram of the cluster obstacle avoidance route planning is obtained. It can complete the closed route planning that fully covers the search area, and at the same time has collision avoidance functions inside and outside the UAV cluster.

[0170] Figures 4 to 8 The simulation results for this example are presented, among which... Figures 4 to 7 The diagrams provided for this example include the deployment of UAV sensors, the route planning in a closed area, the final motion state of the swarm, and the swarm sequence parameter diagram. Figure 8 The obstacle avoidance behavior of the drone swarm throughout the entire simulation process is presented. Figure 7 and Figure 8Simulation results show that the method used in this invention effectively improves the convergence of swarm motion, emphasizing internal swarm coordination and external obstacle avoidance while completing the regional target coverage task. Through a distributed UAV swarm control framework and a novel cooperative coverage method, the problem of continuous coverage by multiple UAVs under complex constraints is solved.

Claims

1. A method for cooperative obstacle avoidance of a swarm of European starling-inspired UAVs for regional coverage, characterized in that: The implementation steps of this method are as follows: Step 1: Divide the given region into a corresponding number of sub-regions, specifically including: S11. Cover a certain number of sensor points within a given area to meet the coverage requirements; S12. Design the fitness function to obtain the deployment location of each sensor and ensure that the area is completely covered; S13. Based on the K-means method and feedback mechanism, complete the region division; Step 2: Plan closed paths within the sub-region After obtaining the clusters, solve the traveling salesman problem to obtain the path through all sensor points in the clusters; under the premise of complete coverage, the shortest path is required to be planned. Step 3: Identify interactive drones based on neighbor selection methods and random line-of-sight; Step 4: Implement obstacle avoidance functionality based on a distributed UAV swarm control method, specifically including: S41. Considering the interaction mechanism between unmanned aerial vehicles (UAVs), an artificial potential field is introduced. The specific implementation process is as follows: Define three spatial spheres centered on the UAV; the three spatial sphere regions, from the inside out, are the repulsion region, the alignment region, and the attraction region, with radii of respectively... , and These areas are adjusted by a limited field of view; Assuming that under the designed neighbor selection method, drones in the cluster acquire necessary information, including their own and other drones' states, through their own sensors; assuming there are N locations... Speed ​​is The drones, each affected by neighboring drones; drones i Interaction forces between it and its neighboring drones The calculation is as follows: ； in, , , These represent the effects on drones. i Repulsive forces, alignment forces, and attractive forces; Calculating the repulsive force requires considering the influence of the drone's heading. j Towards drones i During flight, the repulsive force must be large enough to suppress the flight state; therefore, the repulsive force is calculated as follows: ； in, Indicates the speed of drone i. Indicates drone i Location, Indicates neighboring drones j Location, , Indicates drone i A collection of neighboring drones; The alignment force is calculated as follows: ； Similarly, gravity is introduced to prevent the drone swarm from diverging, as calculated below: ； The interaction force is designed to ensure the drone achieves the desired motion; therefore, when mapped into drone control, the interaction force should be used as the drone's desired velocity input; for the first... i The speed control parameters for the drone are: ； S42. Based on the informed UAV guidance mechanism, ensure that the swarm converges to the planned flight path. The specific implementation process is as follows: To enable the drone swarm to move at the expected speed and in the desired direction, informed drones are added to the distributed drone swarm control framework; these informed drones have a desired preferred speed. Furthermore, it flies along a pre-planned closed flight path, influencing other drones through interaction mechanisms; as a member of the entire drone swarm, the informed drone has the same interaction rules as other drones; simultaneously, informed drones should be included in the neighbor set of other individuals; since each informed drone has its expected speed and desired direction, its interaction force should maintain a balance between the interaction force and its preferred speed; therefore, informed drones... k The interaction forces are calculated as follows: ； in, The interaction forces generated by each informed drone's neighbors. Indicates the weighting ratio. ;if = 1, Informed drones only consider the expected speed, ignoring the security of their neighbors and swarm; if = 0, the informed drone is actually an ordinary drone; When informed drones are used as ordinary drones i When calculating the neighbor information of a drone, the drone is informed of the drone's behavior. i Weights also need to be considered when dealing with interaction forces. drones i The interaction forces are calculated as follows: ； in, Informed drone j For drones i The force, Indicates drone i Other drones in the neighbor set i The resultant force of the forces, Indicates the weighting ratio. ;if In this case, informed drones account for the largest proportion of the interaction forces; if In this case, the informed drone is actually regarded as an ordinary drone; S43. Design an obstacle avoidance mechanism based on the collision cone method. The specific implementation process is as follows: Given the desired obstacle avoidance speed Group interaction drones i final distributed group speed The calculation is as follows: ； Assuming each drone is equipped with a low-level autopilot in the inner-loop controller, a six-degree-of-freedom model is used to drive the motion of the drone swarm; the kinematic model of the fixed-wing drone is as follows: ； in, , , and They represent drones i Its inertial position, airspeed, heading angle, and rate of change of altitude; , , These are the command inputs for horizontal speed, heading angle, and rate of change of altitude, respectively. , and These are the time constants for speed, heading angle, and rate of change of altitude, respectively. To simplify control inputs, the drone model is based on distributed swarm velocity. Convert the three-axis components: ； For a real-world drone model, constraints should be considered; the command input should meet the following limiting conditions: ； Step 5: Output the route obstacle avoidance planning results.

2. The method for cooperative obstacle avoidance of a swarm of European starling-inspired UAVs for regional coverage according to claim 1, wherein the specific implementation process of step S12 is as follows: The target area is divided into a series of square grids, each with a side length of... Much smaller than the sensor's detection radius ;definition Accessible area Chinese grid set; let For a set of sensors, where n This represents the maximum number of sensors; the area must be completely covered by the sensors, i.e., any grid. All four vertices are on the sensor Within the detection range, and the area is free of obstacles; assuming Indicates sensor Specific deployment information, For sensor position, As a valid mark, express Deployed; therefore, vector This represents one possible sensor deployment scenario, which is a feasible solution to the sensor deployment problem. Based on the sensor deployment requirements, the fitness function for this problem is designed as follows: ； in, Indicates the number of sensors in the obstacle area. This indicates the number of areas within the reachable region that are not covered by sensors. This represents the number of overlapping areas covered by sensors within the reachable region, with weights. >> >> If the fitness function value is greater than 0, we need to find the optimal solution that minimizes the fitness function value. An improved cooperative co-evolutionary particle swarm optimization (CCPSO2) algorithm is used to update sensor deployment locations to solve the sensor deployment problem. The improvement involves proposing heuristic rules for updating particle positions and employing multiple rules during the position update process, increasing particle diversity and improving algorithm performance. This algorithm will... n The dimensional solution vector is decomposed into K Each component corresponds to a certain number of components. s 3D particles ( e 1, e 2,..., e s ) group, among which, s = n / K ; Assumption and These are the particle's current position and the number of... j Particles in a cluster i The local optimal position, and For cluster j The historical global optimal position; Local optimal position and global optimal position The updated formula is as follows: ； in f (·) represents the fitness function, a vector Representing vectors ; The formulas for updating the speed and position of cluster j are as follows: ； in, ; in, It is the inertia coefficient that balances global and local search. and It is a random number in the range [0,1]. , , It is the acceleration constant. C (1) and N (0,1) are random numbers generated by Cauchy and Gaussian distributions, respectively. Indicates the first j The speed of each cluster; to maintain particle diversity and avoid premature convergence to a local optimum, local neighborhood optima are used. Replace global optimum ; By using ring topology to describe the neighborhood of a particle, therefore, Described as: 。 3. The method for cooperative obstacle avoidance of a swarm of European starling-inspired UAVs for regional coverage according to claim 1, wherein the specific implementation process of step S13 is as follows: After solving the sensor deployment problem, a set of sensor deployments is obtained. The K-means algorithm is used to cluster the sensor deployment points. The workflow is as follows: (1) Select initial cluster centers C = {C1,C2, ...,C} k }; (2) Calculate data points x With cluster center Distance between The data points are assigned to the nearest cluster. (3) Recalculate the center of each cluster; (4) Repeat steps (2) and (3) until the criterion function converges or the cluster centers no longer change; The algorithm uses the squared error criterion function as the criterion function, expressed as: ； in, k For the number of clusters, Let i be the number of data points in the i-th cluster. For the first i The first cluster of the th cluster j Data points, For the first i The mean of all data points in each cluster ； The initial cluster centers are selected using the density method; S Medium sensor point x The density is defined as with x The distance is not greater than the radius of a specific neighborhood. r The number of sensor points; For the density method, the first step is to select an appropriate neighborhood radius. r To calculate the density at each point D i (r) The data points with the highest density are used as the initial centers of the first cluster. c 1; then, delete. c 1 domain radius r The point with the highest density among the remaining data points is taken as the initial center of the second cluster, and the initial centers of all clusters are derived from this. By distance We improve the K-means method by using weighted distances instead of Euclidean distances as data points. To the cluster center The distance metric between them is defined as: ； Cluster center The distance weights; therefore, the modified criterion function: ； The distance weights need to be adjusted after each iteration. This ensures that the path lengths of each sub-region are consistent; if the search path length of a sub-region is greater than the average length, then the distance weight is adjusted. Need to be reduced; Assumption To pass the first i The shortest closed path length of all data points in each cluster. If the average value is the shortest path length, then the deviation of the path length is: ； Therefore, distance weight Updated to: ； In the formula, The distance weight at the final moment. For feedback gain, satisfying > 0; the algorithm terminates when the absolute value of the maximum route length deviation is not greater than a certain threshold ε, otherwise it continues to iterate.

4. The method for cooperative path avoidance of a swarm of European starling-inspired UAVs for regional coverage, as described in claim 1, includes the following specific steps in step two of solving the traveling salesman problem: If there are no obstacles between the two sensors, calculate the Euclidean distance directly; if there are obstacles, use the A* algorithm to solve for the distance; let... For distance set, Indicates sensor and The distance between them is described in detail below: ； For the sensor deployment point set This yields an m×m total distance matrix. ;set up Let a, b, c ∈ [1, m] be the sequence of sensor deployment points, where for The sensor deployment points; therefore, with and Given the conditions, solve the traveling salesman problem and obtain the desired solution within the solution region. sq This problem is solved using an improved genetic algorithm, with the fitness function defined as: ； In the formula, express The length.

Images