Layered decoupling multi-unmanned aerial vehicle task planning method

Abstract

The invention provides a hierarchical decoupling multi-unmanned aerial vehicle cooperative task planning method. According to the method, task planning is layered and decoupled into three stages for a multi-unmanned aerial vehicle task planning problem in a complex environment after an earthquake disaster. The method comprises the following steps: firstly, converting a post-disaster environment model into a weighted undirected graph comprising unmanned aerial vehicle nodes and task point nodes; performing feature extraction on the weighted undirected graph by adopting a graph neural network to generate global environment features; generating an initial scheme of task allocation based on the global environment features; then, based on a track prediction result and a post-disaster environment model, performing task allocation by adopting a particle swarm algorithm; and then generating an action path according to the distributed task and planning the action of the unmanned aerial vehicle, finally carrying out action re-planning, and carrying out path and action adjustment in the task execution process. According to the method, effective decoupling of task allocation and action planning is realized, the calculation complexity of task planning is remarkably reduced, and the task planning efficiency and post-disaster complex environment adaptability of a multi-unmanned aerial vehicle system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-unmanned aerial vehicle cooperative task planning, and particularly relates to a multi-unmanned aerial vehicle cooperative task planning method based on hierarchical decoupling. BACKGROUND

[0002] As a sudden natural disaster, earthquakes have strong destructive power, wide range of influence and strong suddenness, and have caused great threat to human society. After an earthquake occurs, traffic paralysis and communication interruption often occur in the disaster area, making it difficult for traditional rescue methods to quickly reach the disaster area. With the continuous development of unmanned aerial vehicle technology, unmanned aerial vehicles have become an important means of post-disaster rescue and material delivery due to their characteristics of rapid response, flexibility and independence from ground transportation. In particular, in the field of post-disaster emergency rescue and material distribution, how to effectively plan the tasks of a multi-unmanned aerial vehicle system has become a hot research topic.

[0003] Multi-unmanned aerial vehicle cooperative task planning involves multiple unmanned aerial vehicles reasonably allocating tasks, planning flight trajectories, avoiding conflicts and obstacles, and optimizing task execution efficiency and system performance. For multi-unmanned aerial vehicle cooperative planning problems, common planning methods include A* algorithm, genetic algorithm, ant colony algorithm, and deep Q network method. With the increase in task size and the number of unmanned aerial vehicles, the computational complexity of existing methods increases significantly, making real-time performance difficult to meet. In addition, when facing complex post-disaster environments and multiple task demands, the coupling of task allocation and action planning in task planning increases the complexity of solving the overall problem, and existing methods lack efficient decoupling strategies, resulting in reduced real-time performance and flexibility of task planning, making it difficult to coordinate and resolve conflicts among multiple unmanned aerial vehicles. SUMMARY

[0004] Therefore, the present application provides a multi-unmanned aerial vehicle task planning method based on hierarchical decoupling, which realizes effective decoupling of task allocation and action planning by dividing the task planning process into three stages: task allocation, action planning, and action re-planning.

[0005] To solve the above technical problems, the present application is implemented as follows.

[0006] A multi-unmanned aerial vehicle task planning method based on hierarchical decoupling, comprising: a task allocation stage A, an action planning stage B, and an action re-planning stage C;

[0007] In the task allocation stage A: a post-disaster environment model is formed based on a map of the post-disaster environment, unmanned aerial vehicle states, and task demands; trajectory prediction is performed based on the post-disaster environment model to obtain flight trajectories from the starting point to the task point and between the task points of the unmanned aerial vehicles, forming global predicted trajectories;

[0008] The post-disaster environment model is converted into a weighted undirected graph including unmanned aerial vehicle nodes and task point nodes; a graph neural network is used to extract features of the weighted undirected graph, and global environment features are generated;

[0009] A task allocation model is established, and a decision variable is an allocation relationship between an unmanned aerial vehicle and a task; a task allocation probability is determined based on the global environment features, and a task allocation scheme is initialized according to the task allocation probability; a target function is calculated based on the global predicted track and the post-disaster environment model, and the task allocation scheme is optimized by using a particle swarm algorithm;

[0010] In the action planning stage B: an action planning model is established, and a decision variable is unmanned aerial vehicle path selection; on the basis of the global predicted track and the optimized task allocation scheme, a particle swarm algorithm is used to optimize the unmanned aerial vehicle path selection scheme;

[0011] In the action re-planning stage C: path conflicts of the optimized unmanned aerial vehicle path selection scheme are checked according to the weighted undirected graph, and the conflicts are adjusted to complete action re-planning.

[0012] Preferably, the unmanned aerial vehicle state is S i =<P i ,T i ,E i ,SOC i ,V i ,G i ,H i >; wherein P i represents a current position of the unmanned aerial vehicle i, T i represents a set of task points allocated to the unmanned aerial vehicle i, E i and SOC i represent total power and remaining power of the unmanned aerial vehicle i, V i represents a maximum flight speed of the unmanned aerial vehicle i, G i represents a task execution speed of the unmanned aerial vehicle i, and H i represents a maximum task execution amount of the unmanned aerial vehicle i.

[0013] The task demand is D j =<P j ,Q j ,R j >; wherein P j represents a position of the task point j, Q j represents a task demand amount of the task point j, and R j represents a task priority of the task point j.

[0014] Preferably, in the weighted undirected graph, the unmanned aerial vehicle node describes a state S i of each unmanned aerial vehicle; and the task point node describes a state Dj ;

[0015] The weight of the edge between the UAV node and the task point node is defined as a task allocation and action planning cost, the task allocation cost is a task demand, and the action planning cost is a predicted path length; and the weight of the edge between the task point nodes is defined as an action planning cost.

[0016] Preferably, the flight path prediction based on the post-disaster environment model is to perform flight path prediction based on a hierarchical path prediction algorithm, including:

[0017] A grid map is constructed using a first resolution, and a global path is planned using a first path prediction algorithm; wherein obstacles with an occupancy grid cell less than a set threshold are ignored;

[0018] A grid map is constructed using a second resolution higher than the first resolution, and a local optimization is performed on the obstacle regions in the global path using a second path prediction algorithm;

[0019] The global path and the local optimization path are merged to obtain a global predicted flight path.

[0020] Preferably, the first path prediction algorithm uses an A* algorithm, and the second path algorithm uses an improved A* algorithm; the improved A* algorithm adds an obstacle avoidance cost in the cost function; the smaller the distance between the path point and the nearest obstacle, the higher the avoidance cost corresponding to the path point.

[0021] Preferably, the obstacle avoidance cost is constructed in the following manner:

[0022]

[0023] wherein C obs (n) represents the obstacle avoidance cost of path point n and surrounding obstacles; d min (n) represents the distance from path point n to the nearest obstacle, and d threshold is a safety distance threshold.

[0024] Preferably, the task allocation model is established as follows:

[0025]

[0026] The task allocation target is to minimize the total task completion time, minimize the energy consumption of the UAV, and balance the task allocation; the task allocation constraint conditions include UAV load constraint, UAV endurance time constraint, task unique allocation constraint, and task priority constraint; and the action planning model is established as follows:

[0027]

[0028] The action planning target is to minimize the total path length and the unmanned aerial vehicle action time; and the action planning constraint conditions include path feasibility constraint, unmanned aerial vehicle energy constraint, path safety constraint and flight speed constraint.

[0029] Preferably, the particle swarm algorithm used in optimizing the task allocation scheme and the unmanned aerial vehicle path selection scheme is a multi-objective particle swarm optimization (MOPSO) algorithm;

[0030] The improved MOPSO algorithm introduces a dynamic inertia weight in the speed update formula; the dynamic inertia weight is multiplied by the particle speed to be updated, and the dynamic inertia weight linearly decreases in the iteration process.

[0031] In the improved MOPSO algorithm, when selecting a Pareto optimal solution, the fitness value f pen The particles are selected and optimized;

[0032] f pen =f+λ·pen

[0033] wherein f is an original fitness function value composed of objective function values; λ is a penalty term coefficient; pen is a total constraint penalty, which is obtained by normalizing and summing constraint penalties of each constraint; for each constraint, the constraint penalty represents the degree of violation of the constraint, and if there is no violation of the constraint, the constraint penalty is 0 and does not affect the fitness value; if the constraint is violated, the constraint penalty value is greater than 0, the fitness value is poor, and the constraint penalty value is determined according to the degree of violation of the constraint.

[0034] Preferably, the task allocation probability is determined based on the global environmental features, and the task allocation scheme is initialized according to the task allocation probability as follows:

[0035] The node feature matrix output by the graph neural network is divided into an unmanned aerial vehicle node feature matrix H U and a task point node feature matrix H T according to node types; and the matching score matrix S of unmanned aerial vehicle-task pairs is calculated using H U and H T .

[0036] The constraint is introduced into the matching score matrix S, so that the unmanned aerial vehicles with sufficient power, the tasks with high priority and high demand will obtain higher matching scores, and the matching score matrix S' after introducing the constraint is:

[0037] The row of the matching score matrix S is normalized to obtain a task allocation probability matrix P, and the element P ij in the task allocation probability matrix P is the probability of selecting the task point j by the unmanned aerial vehicle i.

[0038] When the particles are initialized, the probability of selecting the task point j by the unmanned aerial vehicle i and P ijProportional to, and ensure that each task point is only assigned to one UAV, the initial task assignment scheme is generated.

[0039] Preferably, the action re-planning is: the output nodes of the graph neural network include UAV nodes and task point nodes; each layer of the graph neural network aggregates the spatio-temporal states of neighbor nodes, and after L layers of message passing, the final features of UAV nodes i and j are recorded as And The node features are linearly combined by a fully connected layer And , and an activation function is used to generate a conflict probability matrix C;

[0040] When the element C ij of the conflict probability matrix C corresponding to the UAV nodes i and j is greater than a given threshold, it is determined that the UAV nodes i and j have a conflict that needs to be handled; according to the priorities S i and S j of the tasks to be performed by the UAV nodes i and j, if S i >S j , the UAV i has priority, and the UAV j adjusts the path.

[0041] Advantages:

[0042] (1) The present application provides a hierarchical decoupling strategy based on a graph neural network, which divides the task planning into three stages of task assignment, action planning and action re-planning. This decoupling method makes the task assignment and action planning independent of each other, and can be optimized separately for different task requirements and environmental conditions, thereby greatly reducing the computational complexity and improving the flexibility and efficiency of multi-UAV cooperative task planning.

[0043] (2) Through the feature extraction capability of the graph neural network, the global environmental features are obtained, and when the environment changes, new global environmental features are extracted in real time according to the new environment, effectively improving the task planning efficiency of the multi-UAV system and the adaptability to complex post-disaster environments. Moreover, the global environmental features output by the graph neural network provide an initial assignment scheme for task assignment, making the task assignment optimization of the particle swarm algorithm faster and more accurate.

[0044] (3) In a preferred scheme, the A* algorithm is improved, and the improved A* algorithm combines global path planning and local optimization through a hierarchical strategy, which can accurately avoid obstacles and avoid a sharp increase in computational complexity while ensuring planning efficiency.

[0045] (4) In a preferred embodiment, the MOPSO algorithm is improved, and the improved MOPSO algorithm can make the particle balance global search and local exploration more flexibly in the search process by introducing a dynamic inertia weight and a penalty function, avoid the problem of early convergence, and can more effectively handle multi-objective optimization tasks, and ensure the multiple optimization of the target. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The principle diagram of the layered decoupled multi-unmanned aerial vehicle task planning scheme of the application.

[0047] Figure 2 The flowchart of the layered decoupled multi-unmanned aerial vehicle task planning method of the application.

[0048] Figure 3 The flowchart of the layered A* algorithm;

[0049] Figure 4 The schematic diagram of the weighted undirected graph model;

[0050] Figure 5 The schematic diagram of the feature extraction process of the graph neural network;

[0051] Figure 6 The flowchart of the improved MOPSO algorithm. DETAILED DESCRIPTION

[0052] The application provides a layered decoupled multi-unmanned aerial vehicle task planning method, which divides the task planning into three stages of task allocation, action planning and action re-planning. This decoupling method makes the task allocation and action planning independent of each other, can be optimized separately according to different task requirements and environmental conditions, thereby greatly reducing the calculation complexity and improving the flexibility and efficiency of the multi-unmanned aerial vehicle cooperative task planning. At the same time, the feature extraction capability of the graph neural network is adopted, the adaptability of the system to the task requirements and the environment is enhanced, and the task planning efficiency of the multi-unmanned aerial vehicle system and the adaptability to the complex post-disaster environment are effectively improved. Moreover, the global environmental features output by the graph neural network provide an initial allocation scheme for the task allocation, so that the task allocation optimization of the particle swarm algorithm is faster and more accurate.

[0053] The application will be described in detail below with reference to the drawings and examples.

[0054] The layered decoupled multi-unmanned aerial vehicle task planning method of the application includes a task allocation stage A, an action planning stage B and an action re-planning stage C. The three stages will be described in detail below. Figure 1 and Figure 2 The three stages are described in detail. Steps 1-6 correspond to the task allocation stage A, steps 7-8 correspond to the action planning stage B, and step 9 corresponds to the action re-planning stage C.

[0055] The multi-UAV mission planning method of the present invention specifically includes the following steps:

[0056] Step 1: Construct a post-disaster environmental model.

[0057] In this step, a post-disaster environment model is formed based on the post-disaster environment map, drone status, and mission requirements.

[0058] The post-disaster environment model includes a geographic environment model, a drone model, and a mission requirement model, which can be used to initialize the target area map, intelligent agent, and mission model.

[0059] Geographic environment model, including the UAV’s departure point location, mission point location, and obstacle set.

[0060] The drone model includes the state S of each drone i = <P i ,T i ,E i ,SOC i ,V i ,G i ,H i >, where P i Indicates the current position of drone i, T i represents the set of mission points assigned to UAV i, E i and SOC i Represent the total power and remaining power of drone i, V i represents the maximum flight speed of drone i, G i represents the mission execution speed of UAV i, H i Indicates the maximum mission execution amount of UAV i.

[0061] Task requirement model, including the status D of each task point j = <P j ,Q j ,R j >, where P j represents the location of task point j, Q j represents the task demand at task point j, R j Indicates the task priority of task point j.

[0062] Step 2: Predict flight paths based on the post-disaster environment model.

[0063] In this step, trajectory prediction is performed based on the post-disaster environmental model to obtain the flight trajectory from the starting point of the UAV to the mission point and between the mission points to form a global predicted trajectory.

[0064] In a preferred embodiment, track prediction is performed based on a hierarchical path prediction algorithm. Figure 3The specific steps of the flight path prediction process are shown as follows:

[0065] Step 2.1: Plan a global path: construct a grid map with a first resolution, and plan a global path with a first path prediction algorithm; wherein obstacles occupying less than a set threshold of grid cells are ignored.

[0066] In this step, the starting position of the UAV, the task point position, the three-dimensional environment map, and the UAV flight parameters are input; the flight environment is divided into grids, and the positions of obstacles and passable areas are marked in each grid cell. For global path planning with low grid resolution, a traditional A* algorithm is used to generate a rough path from the starting point to the target task point. In this process, obstacles occupying fewer grid cells can be regarded as small obstacles and ignored to quickly obtain a preliminary path. The cost function of the traditional A* algorithm is:

[0067] f(n) = g(n) + h(n) (1)

[0068] wherein g(n) represents the actual cost from the current position of the UAV to path point n, and h(n) is the heuristic estimated cost from path point n to the target task point, and the heuristic estimated cost adopts the Euclidean distance.

[0069] Step 2.2: Optimize a local path: construct a grid map with a second resolution higher than the first resolution, and use a second path prediction algorithm to locally optimize the obstacle regions in the global path.

[0070] In this step, on the basis of the global path, a grid map with higher resolution is constructed to accurately represent obstacles and passable areas. In a preferred scheme, an improved A* algorithm is used to locally optimize the obstacle regions in the path to ensure that the action path avoids small obstacles.

[0071] The improved A* algorithm adds an obstacle avoidance cost in the original cost function, and the distance between the node and the obstacle is used to determine the avoidance cost:

[0072] f(n) = g(n) + h(n) + a · C obs (n) (2)

[0073] wherein a is a constant; C obs (n) represents the obstacle avoidance cost of path point n and the nearest obstacle around, and the cost is used to measure the spatial relationship between path point n and the obstacle. The smaller the distance between the path point and the nearest obstacle, the higher the avoidance cost corresponding to the path point. An optimal obstacle avoidance cost considers a safety distance threshold, and the obstacle avoidance cost is represented as:

[0074]

[0075] wherein dmin (n) represents the distance from path point n to the nearest obstacle, d threshold is a safety distance threshold.

[0076] Step 2.3, merging the global path and the locally optimized path to obtain a global predicted track.

[0077] In this step, the rough path in the global path is merged with the refined path after local optimization. If the global path passes through an obstacle region, the result after local path optimization is replaced.

[0078] Step 3: Convert the post-disaster environment model into a weighted undirected graph G=(V,E).

[0079] Figure 4 An example of a weighted undirected graph is shown. The node set V of the weighted undirected graph includes unmanned aerial vehicle nodes and task point nodes. The unmanned aerial vehicle nodes describe the state S i of each unmanned aerial vehicle; the task point nodes describe the state D j of each task point. The edge set E represents the mutual relationship between nodes, including edges between unmanned aerial vehicle nodes and task point nodes, and edges between task nodes. The weight of the edge between the unmanned aerial vehicle node and the task point node is defined as the task allocation and action planning cost, the task allocation cost is the task demand, and the action planning cost is the predicted path length; the weight of the edge between the task nodes is defined as the action planning cost. The edge set includes edges between unmanned aerial vehicle nodes and task point nodes, and edges between task nodes.

[0080] When initializing the graph modeling, the edge weight is the initial value, and the edge weight will change through the processing of the graph neural network. The present application determines the initial allocation scheme according to the output data of the graph neural network to guide the task allocation work.

[0081] When a new task is added, a new task point node is dynamically added to the weighted undirected graph, and the edge weight is adjusted; when the task location changes, the corresponding task point node in the weighted undirected graph is modified; when the unmanned aerial vehicle fails, the failed unmanned aerial vehicle node is temporarily removed from the weighted undirected graph. Therefore, the post-disaster environment is expressed by a graph model, which is more easily adapted to new environments and new tasks.

[0082] Step 4: Feature extraction of the graph using a graph neural network to generate global environmental features.

[0083] The input of the graph neural network is the weighted undirected graph, and the output nodes are divided into unmanned aerial vehicle nodes and task point nodes. The output data is the global environmental feature, i.e., the node feature matrix; it is divided into unmanned aerial vehicle node feature matrix and task point node feature matrix according to node type.

[0084] In this step, the multi-layer message passing mechanism of the graph neural network is used to update and enhance the node features. For example,Figure 5 As shown, the graph neural network updates the feature of the target node by aggregating the information of neighbor nodes in each layer.

[0085] For node v i , the feature update formula in the k+1 layer is:

[0086]

[0087] where denotes the feature of node v i in the k layer, N(i) is the neighbor node set of node v i , W (k) is the learnable weight matrix in the k layer, the learning process is realized by supervised learning, where the label of the supervised learning training sample is the optimal solution generated by the classical optimization algorithm (such as mixed integer programming); σ is a nonlinear activation function, and AGG(·) is an aggregation function used to combine the feature information of neighbor nodes together.

[0088] Through continuous iterative update, the feature of each node gradually integrates the global information of the whole graph while retaining the local characteristics of the neighborhood, and finally outputs the updated node feature matrix.

[0089] Step 5: Establish a task allocation model.

[0090] In this step, a task allocation model is established, and the decision variable is the allocation relationship between the unmanned aerial vehicle and the task:

[0091]

[0092] The objective function includes the total task completion time, the energy consumption of the unmanned aerial vehicle, and the balanced allocation of tasks. The constraint conditions include the unmanned aerial vehicle load constraint, the unmanned aerial vehicle endurance time constraint, the task unique allocation constraint, and the task priority constraint, which specifically include the following contents:

[0093] 1) Task allocation objective function:

[0094] (a) Total task completion time f 11 (x): the maximum value of the time for a single unmanned aerial vehicle to complete all tasks, and the time for the unmanned aerial vehicle to complete all tasks is the sum of the time for the unmanned aerial vehicle to fly to the target task point and the time to complete the task;

[0095]

[0096] where d ij is the flight path length of the unmanned aerial vehicle i from the current position to the task point j, calculated according to the track prediction result, v i is the average flight speed of the unmanned aerial vehicle i, Q j is the task demand, and reff For the efficiency of the task execution of the UAV i, r eff is a constant; M is the total number of tasks.

[0097] (b) Energy consumption of the UAV f 12 (x): The total energy consumption of all UAVs completing the corresponding tasks is the sum of the energy consumption of the UAV flying to the target task point and the energy consumption of completing the task;

[0098]

[0099] wherein k f is the energy consumption coefficient per unit distance, k w is the energy consumption coefficient per unit task, both are constants; M is the total number of tasks, and N is the total number of UAVs.

[0100] (c) Task balanced allocation f 13 (x): Minimize the variance of the task completion time so that the task load of all UAVs is as uniform as possible;

[0101]

[0102] wherein, is the average of the task completion time of all UAVs;

[0103] The overall task allocation optimization goal is to minimize the total task completion time, minimize the energy consumption of the UAV, and balance the task allocation.

[0104] 2) Constraint conditions:

[0105] (a) UAV load constraint: the total load of the UAV cannot exceed its maximum carrying capacity W i ;

[0106]

[0107] (b) UAV endurance time constraint: the total energy consumption of the UAV when performing tasks cannot exceed its total power E i ;

[0108]

[0109] wherein k is the power consumption per unit distance of the UAV, and k is a constant.

[0110] (c) Task unique allocation constraint: each task can be performed by at most one UAV;

[0111]

[0112] (d) Task priority constraint: first allocate UAVs to high priority high tasks;

[0113]

[0114] Step 6: Task allocation scheme optimization based on particle swarm algorithm.

[0115] In this step, an initial task allocation scheme is generated based on the global environment features; a target function is calculated based on the global predicted track and post-disaster environment model, and a particle swarm algorithm is used to optimize the task allocation scheme.

[0116] In this step, the improved MOPSO algorithm is used for task allocation to determine the matching relationship between the unmanned aerial vehicle and the task point, and the corresponding unmanned aerial vehicle is allocated to each task point for material delivery. Figure 6 The specific steps of the improved MOPSO algorithm are as shown in

[0117] Step 6.1: Initialize the particle swarm and initialize the external archive.

[0118] In this step, the decision variables of the task allocation problem are represented as particles p = [x 11 ,...,x 1M ,...,x NM ], and the particle position and velocity of the population are initialized.

[0119] In the present application, the initialization of particles is not randomly generated, but is generated according to the global environment features output by the graph neural network. The specific generation process is as follows: the node feature matrix output by the graph neural network is converted into a task allocation probability matrix after feature alignment and constraint embedding, and each particle generates an initial task allocation scheme according to the task allocation probability matrix.

[0120] The node feature matrix output by the graph neural network is aligned by a weight matrix , which is the weight matrix of the last layer of the graph neural network in step 4, where D' is the feature dimension of the graph neural network output. The node feature matrix output by the graph neural network is divided into an unmanned aerial vehicle node feature matrix and a task point node feature matrix The row in the node feature matrix represents each node, the column represents the feature dimension of the node, and the element h ij is the jth feature value of the ith node. The matching score matrix of the unmanned aerial vehicle-task pair is calculated using H U and H T .

[0121]

[0122] where ReLU is an activation function.

[0123] Constraints are introduced into the matching score:

[0124]

[0125] where max(E) is the maximum energy of all drones, max(R) is the highest priority of all task points, and max(Q) is the highest demand of all task points. Drones with sufficient energy, high priority, and high demand will get higher matching scores.

[0126] The rows of the matching score matrix S are normalized to form the task assignment probability matrix P. The probability of drone i choosing task point j is P ij .

[0127] For example, there are 3 drones and 5 task points. After feature alignment and constraint embedding, the matching score matrix S' is:

[0128]

[0129] After row normalization, the probability of drone 1 executing task 1 is:

[0130]

[0131] When constructing particles, the roulette method is used to select task points according to the probability distribution. The probability of task point j being chosen by drone i is proportional to P ij , and it is ensured that each task point is assigned to only one drone. Each particle represents an initial task assignment scheme.

[0132] Step 6.2: Evaluate the fitness of the initial particle swarm and initialize the Pareto solution set.

[0133] Here, the objective function in the task assignment model is defined as three optimization objectives, which are the minimization of total task completion time, the minimization of drone energy consumption, and the balanced allocation of tasks.

[0134] Step 6.3: Select the guiding particles in the external archive according to the crowding degree.

[0135] In this step, the original fitness value f of each particle p is calculated, and the fitness is modified by a penalty function to dynamically adjust the fitness of solutions that do not meet the constraint conditions.

[0136] f pen = f + λ · pen (12)

[0137] where f penFor the fitness value of the added penalty term, f is the original fitness function value, is the target vector composed of the target function value; λ is the penalty term coefficient, pen represents the degree of violation of the constraint, and is the total constraint penalty. The constraint penalty of each constraint is normalized and summed to obtain, for each constraint, the constraint penalty represents the degree of violation of the constraint, and if there is no violation of the constraint, the constraint penalty is 0, which does not affect the fitness value; if the constraint is violated, the constraint penalty value is greater than 0, the fitness is poor, and the constraint penalty value is determined according to the degree of violation of the constraint. The degree of constraint violation is preferably increased in the form of a linear function to increase the value of the objective function.

[0138] For example, for the unmanned aerial vehicle load constraint, the linear penalty is:

[0139]

[0140] If , the constraint is satisfied, pen1=0, which does not affect the fitness value, otherwise the constraint is violated and pen1>0, the fitness is poor. The linear penalty of each constraint is normalized and summed to obtain the total constraint penalty pen.

[0141] Step 6.4: update the individual optimal solution p best and the global optimal solution g best .

[0142] Step 6.5: adaptively adjust the inertia weight according to the current iteration number.

[0143] The present application introduces a dynamic inertia weight in the speed update formula, so that the particle has strong global search ability in the early exploration stage, and gradually converges to the Pareto optimal front in the later stage, and the speed update formula is:

[0144]

[0145] Where ω (t) is the dynamic inertia weight of the tth generation, which decreases linearly in the iteration process, is the speed of particle p in the tth generation, is the position of particle p in the tth generation, c1 and c2 are the individual learning factor and the group learning factor respectively, and r1 and r2 are random numbers.

[0146] Step 6.6: update the particle position, recalculate the fitness of the particle, update the individual optimal solution p best and the global optimal solution g best .

[0147] Step 6.7: add the particles satisfying the Pareto optimal condition to the external archive, and delete too many solutions through the crowding degree screening mechanism to keep the archive size fixed.

[0148] In step 6.8, repeat steps 6.3 to 6.7. When the maximum number of iterations is reached, stop the iteration, output the Pareto frontier solution set, and obtain the final task allocation result.

[0149] Step 7: Build an action planning model.

[0150] In this step, an action planning model is established, and the decision variable is the drone path selection:

[0151]

[0152] The action planning objective function includes the total track length and the UAV action time; the constraints include path feasibility constraints, UAV energy constraints, track safety constraints, and flight speed constraints.

[0153] Specifically include the following:

[0154] 1) Objective function:

[0155] (a) Total track length f 21 : The sum of the trajectory lengths of all UAVs completing all tasks;

[0156]

[0157] Among them, d ijk is the trajectory length of UAV i from the current position j to the mission point k; E is the edge set.

[0158] (b) UAV operation time f 22 : The time it takes for the drone to complete all tasks;

[0159]

[0160] 2) Constraints:

[0161] (a) Path feasibility constraint: This constraint ensures the flow balance of the task points, i.e., ensures the path continuity;

[0162]

[0163] This constraint ensures that the number of times a drone enters mission point j is equal to the number of times it leaves mission point k.

[0164] (b) UAV energy constraint: the energy consumption of the UAV during mission execution shall not exceed the initial power;

[0165]

[0166] Where k is the power consumption of the drone per unit distance, and k is a constant.

[0167] (c) Track safety constraint: the UAVs must maintain a minimum safety distance during flight;

[0168]

[0169] where d io is the distance between the UAV i and other UAVs or obstacles o min is the set minimum safety distance.

[0170] (d) Flight speed constraint: the flight speed of the UAVs does not exceed their maximum speed;

[0171]

[0172] The optimization goal of the overall action planning is to minimize the total track length and the UAV action time.

[0173] Step 8: Action planning scheme optimization based on particle swarm algorithm.

[0174] This step uses a particle swarm algorithm to optimize the UAV path selection scheme based on the global predicted track and the optimized task allocation scheme.

[0175] In a preferred scheme, this step uses an improved MOPSO algorithm for action planning to further optimize the specific flight path of the UAV based on the predicted track and the allocation scheme. The decision variables of the action planning problem are represented as particles p = [p 112 ,...,p 11M ,...,p 1(M-1)M ,...,p N(M-1)M ], and the objective function in the action planning model is defined as two optimization goals, which are to minimize the total track length and to minimize the UAV action time. The specific steps of the improved MOPSO algorithm are the same as steps 6.3-6.8, which are not repeated here.

[0176] Step 9: Action replanning.

[0177] This step perceives UAV path conflicts through a graph neural network. Each layer of the graph neural network aggregates the spatio-temporal states of neighboring nodes. After L layers of message passing through the graph neural network, the final features of UAV nodes i and j are and The node features are linearly combined through a fully connected layer, and the conflict probability matrix

[0178]

[0179] where W c is the weight matrix of the fully connected layer, and b cis the bias term, σ is the sigmoid function, and C is the cost function. ij If C > τ (τ is a given threshold), then the UAV pair (i, j) is determined to be a conflict that needs to be handled.

[0180] For the conflict UAV pair (i, j), according to the priority S i and S j of the task to be executed, if S i > S j , then the UAV i has priority to pass, and the UAV j needs to adjust the path. According to the path information of the conflict UAV pair (i, j) (including the path coordinate sequence and the corresponding time stamp), the intersection or the closest distance point of the paths of the UAVs is calculated as the conflict point P conflict . A circular region with a safety distance threshold as the radius is extended outward from the conflict point as the center, the obstacles and the existing paths are removed, and the uniform sampling in the polar coordinate system is performed in the region to generate a set of safe waypoints P safe . For each candidate waypoint P safe in the set P k , the detour distance of adding the waypoint to the original path is calculated, and the UAV j selects the waypoint P best with the shortest detour path to be added to the original path and the conflict point P conflict to be removed, so as to realize the action re-planning.

[0181] After all the UAVs complete the corresponding tasks, it is determined that the multi-UAV cooperative task planning is completed. During the task execution, the path and action can be adjusted according to the environmental changes.

[0182] The specific embodiments described above only describe the design principles of the present application, and the shapes and names of the components in the description can be different and are not limited. Therefore, the person skilled in the art of the present application can modify or equivalently replace the technical solutions described in the foregoing embodiments; and these modifications and replacements do not deviate from the purpose and technical solutions of the present application, and should all belong to the protection scope of the present application.

Claims

1. A hierarchical decoupling multi-UAV mission planning method, characterized in that: include: Task allocation phase A, action planning phase B, and action replanning phase C; In the task allocation phase A, a post-disaster environment model is formed based on the post-disaster environment map, the drone status, and the task requirements; a trajectory is predicted based on the post-disaster environment model to obtain the flight trajectory of the drone from the starting point to the task point and between the task points to form a global predicted trajectory; The post-disaster environmental model is converted into a weighted undirected graph consisting of drone nodes and mission point nodes. A graph neural network is used to extract features from the weighted undirected graph to generate global environmental features. Establish a task allocation model, where the decision variable is the allocation relationship between the UAV and the task; determine the task allocation probability based on the global environmental characteristics, and initialize the task allocation plan based on the task allocation probability; calculate the objective function based on the global predicted trajectory and post-disaster environmental model, and use the particle swarm algorithm to optimize the task allocation plan; In the action planning phase B: establishing an action planning model, with the decision variable being the UAV path selection; and using a particle swarm algorithm to optimize the UAV path selection scheme based on the global trajectory prediction and the optimized task allocation scheme; In the action replanning phase C: the path conflicts of the optimized UAV path selection scheme are checked according to the weighted undirected graph, and adjustments are made to the conflicting paths to complete the action replanning.

2. The method according to claim 1, wherein The drone status is: S i = <P i ,T i ,E i ,SOC i ,V i ,G i ,H i > Among them, P i Indicates the current position of drone i, T i represents the set of mission points assigned to UAV i, E i and SOC i Represent the total power and remaining power of drone i, V i represents the maximum flight speed of drone i, G i represents the mission execution speed of UAV i, H i represents the maximum mission execution amount of UAV i; The task requirements are: D j = <P j ,Q j ,R j > Among them, P j represents the location of task point j, Q j represents the task demand at task point j, R j Indicates the task priority of task point j.

3. The method according to claim 2, wherein In the weighted undirected graph, the drone node describes the state S of each drone. i ; Task point node describes the status of each task point D j ; The weight of the edge between the UAV node and the task point node is defined as the task allocation and action planning cost, the task allocation cost is the task demand, and the action planning cost is the predicted path length; the weight of the edge between the task point nodes is defined as the action planning cost.

4. The method according to claim 1, wherein The track prediction based on the post-disaster environment model is: track prediction based on a hierarchical path prediction algorithm, including: A grid map is constructed using a first resolution, and a global path is planned using a first path prediction algorithm; obstacles that occupy fewer grid cells than a set threshold are ignored; A grid map is constructed using a second resolution higher than the first resolution, and a second path prediction algorithm is used to locally optimize the obstacle area in the global path; The global path and the local optimized path are combined to obtain a global predicted trajectory.

5. The method according to claim 4, wherein The first path prediction algorithm adopts the A* algorithm; the second path algorithm adopts the improved A* algorithm; the improved A* algorithm adds obstacle avoidance cost to the cost function; the smaller the distance between the path point and the nearest obstacle, the higher the avoidance cost corresponding to the path point.

6. The method according to claim 5, wherein The obstacle avoidance cost is constructed as follows: Among them, C obs (n) represents the obstacle avoidance cost of path point n and surrounding obstacles; d min (n) represents the distance from path point n to the nearest obstacle, d threshold is the safety distance threshold.

7. The method according to claim 1, wherein The task allocation model is established as follows: The task allocation objectives are to minimize the total task completion time, minimize the UAV energy consumption, and balance the task allocation; the task allocation constraints include UAV payload constraints, UAV endurance constraints, task unique allocation constraints, and task priority constraints; the action planning model is established as follows: The action planning goal is to minimize the total track length and the UAV action time; the action planning constraints include path feasibility constraints, UAV energy constraints, track safety constraints, and flight speed constraints.

8. The method according to claim 1 or 7, wherein: The particle swarm optimization algorithm used to optimize the task allocation plan and the UAV path selection plan is the improved multi-objective particle swarm optimization MOPSO algorithm; The improved MOPSO algorithm introduces a dynamic inertia weight into the velocity update formula; the dynamic inertia weight is multiplied by the particle velocity to be updated, and the dynamic inertia weight decreases linearly during the iteration process; When the improved MOPSO algorithm selects the Pareto optimal solution, it adds the fitness value f of the penalty term. pen Perform particle selection; f pen =f+λ·pen Among them, f is the original fitness function value, which is composed of the objective function value; λ is the penalty term coefficient; pen is the total constraint penalty, which is obtained by normalizing the constraint penalty of each constraint; for each constraint, the constraint penalty represents the degree of violation of the constraint. If the constraint is not violated, the constraint penalty is 0 and does not affect the fitness value; if the constraint is violated, the constraint penalty value is greater than 0, the fitness deteriorates, and the constraint penalty value is determined according to the degree of violation of the constraint.

9. The method according to claim 1, wherein The task allocation probability is determined based on the global environment characteristics, and the task allocation scheme is initialized according to the task allocation probability: The node feature matrix output by the graph neural network is divided into the drone node feature matrix H according to the node type U And the task point node feature matrix H T ; Using H U and H T Calculate the matching score matrix S of the drone-task pair; Introducing constraints into the matching score matrix S, so that UAVs with sufficient power and tasks with high priority and high demand will obtain higher matching scores. The matching score matrix with introduced constraints is S'; Normalize the rows of the matching score matrix S and transform it into the task assignment probability matrix P. The element P in the task assignment probability matrix P is ij The probability of selecting mission point j for drone i; When initializing particles, consider the probability of task point j being selected by drone i and P ij Proportional to each task point, and ensure that each task point is assigned to only one UAV, to generate the initial task allocation plan.

10. The method according to claim 1, wherein The action replanning is as follows: the output nodes of the graph neural network include drone nodes and task point nodes; each layer of the graph neural network aggregates the spatiotemporal states of neighboring nodes, and after L layers of message transmission, the final features of drone nodes i and j are recorded as and Through the fully connected layer, the node features and Perform linear combination and then generate the conflict probability matrix C through activation function; When the elements C corresponding to drone nodes i and j in the conflict probability matrix C ij If the value is greater than a given threshold, it is determined that there is a conflict between drone nodes i and j that needs to be handled; according to the priority S of the tasks to be performed by drone nodes i and j i and S j , if S i >S j , UAV i has priority and UAV j adjusts its path.

Images