Unmanned aerial vehicle path planning method for emergency material delivery task
Through neural network prediction of node edge connection probability and division and treatment strategy combined with ant colony algorithm, the problem of large scale and local optimality in drone path planning is solved, and efficient and fast emergency material delivery path planning is achieved, which meets the emergency rescue Timeliness requirements.
Patent Information
- Application Number
- CN202411029721.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-15
AI Technical Summary
The existing technology lacks effective strategies to guide search in the UAV path planning in emergency material delivery tasks. The problem is large and fixed, and it is easy to fall into local optimality, the solution speed is slow, and it is difficult to meet the timeliness requirements of emergency rescue.
The neural network is used to predict the probability of connecting nodes as the pheromone matrix initialization of the ant colony algorithm, and the rescue points are grouped through the partitioning and treatment strategy, combined with the ant colony algorithm and the local search algorithm, and optimized using k-means clustering and periphery neighborhood descent algorithm. Solution: Adaptively adjust the dividing and consolidation scale to break out of local optimality.
It improves the timeliness and efficiency of path planning, effectively shortens the waiting time for delivery of rescue materials, improves the algorithm's global exploration ability and local exploration performance, and achieves efficient and rapid solution.
Smart Images

Figure CN118966496A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of path planning, and in particular to a method for unmanned aerial vehicle (UAV) path planning for emergency material delivery tasks. Background Art
[0002] After a disaster occurs, timely rescue is the key to reducing disaster losses, and immediate and sufficient supply of relief materials is the lifeline. Traditional emergency relief material distribution methods include ground transportation and manned aircraft transportation. After a disaster occurs, the disaster-stricken areas often have damaged and interrupted roads, making it difficult for ground transportation to enter directly. Although manned aircraft transportation is not subject to ground traffic restrictions, it is still subject to weather factors and is accompanied by safety risks. In recent years, drone technology has flourished. Compared with traditional material distribution methods, drones have the advantages of low cost, strong flexibility, no terrain constraints, and low distribution safety risks. They can fully meet the timeliness, flexibility, and safety requirements of post-disaster material distribution. In the process of emergency material distribution, multiple drones are usually dispatched to complete the delivery of rescue materials through collaborative cooperation. Drone path planning technology is the key to realizing multi-drone collaborative decision-making and control. According to the mission requirements of the rescue point and the current configuration of the drone, the order and route of drone material delivery are reasonably planned to minimize the waiting time for material distribution at the disaster site. To solve this problem, a variant of the vehicle routing planning (VRP) problem is introduced, called the cumulative UAV routing problem with capacity constraints at multiple stations. This problem has multiple UAV stations and requires finding the routes of each UAV. The goal is to minimize the total waiting time for material delivery to all rescue points. It also requires that all rescue points be served and only served once. UAVs have limited capacity and flight distance. Considering that the algorithm needs to plan the departure base station of the UAV, which expands the scale of the problem search, and multiple constraints are tight, and the emergency material delivery scenario has high requirements on the timeliness of the solution algorithm, it is worth paying close attention to find an efficient and fast solution method.
[0003] At present, domestic and foreign scholars have conducted systematic research on the solution methods for the cumulative UAV path planning problem with capacity constraints at multiple stations. The existing solution methods can be divided into two categories: exact solution algorithms and heuristic algorithms. The exact solution algorithm refers to the mixed integer linear programming model (MILP) of the problem to be optimized based on the optimization objectives, decision variables, and constraints, and then uses mathematical programming methods to search for the optimal solution to the problem. Commonly used exact algorithms are mainly branch and bound method, column generation method, dynamic programming method, etc. The advantage of the exact algorithm is that it can obtain the optimal solution or approximate optimal solution to the problem. The disadvantage is that the solution scale is small. When solving large-scale problems, its computational complexity is extremely high and the solution time is very long, which is difficult to meet the requirements of emergency rescue.
[0004] Due to the characteristics of emergency rescue scenarios, such as a large number of rescue points (greater than 200) and high timeliness requirements for solving problems, precise algorithms are difficult to meet actual needs. The research on heuristic algorithms has gradually become the focus. Heuristic algorithms can find a better solution in a short time. Heuristic algorithms can be specifically divided into local search algorithms and swarm intelligence optimization algorithms. The local search method refers to generating a feasible solution given an initialization strategy, using various local search operators to traverse the neighborhood of the current solution space, and iterating continuously until all local search operators cannot improve the performance of the solution. Representative methods include the neighborhood descent (VND), neighborhood search (VNS), and large neighborhood search algorithm (ALNS). The advantage of local search is that it converges quickly and can give a better solution to a large-scale problem in a very short time. However, it requires a feasible initialization strategy to support it, and it is easy to fall into the local optimum. It is necessary to design a specific perturbation strategy to help the algorithm jump out of the local optimum. Swarm intelligence optimization methods generate a group of solutions during optimization, called a population. Some better solutions are selected from the population to enter the next round of iteration. The entire population evolves and learns through continuous iterations, and the overall performance is constantly improving, thus constantly approaching the optimal solution. Representative swarm intelligence optimization methods include ant colony algorithm (ACO), genetic algorithm (GA), particle swarm (PSO) algorithm, etc. Compared with local search algorithms, swarm intelligence optimization algorithms have strong global exploration capabilities, but as the scale of the problem increases, their convergence deteriorates. At the same time, there is a lack of effective strategies to guide the search, and the worst case degenerates into random search.
[0005] In summary, the existing technology lacks effective strategies to guide the search, the problem scale is large and fixed, it is easy to fall into local optimality, and the problem solving speed is slow and inefficient. Summary of the invention
[0006] In view of the above problems, the present invention provides a UAV path planning method for emergency material delivery tasks, which solves the problems in the prior art of lack of effective strategies to guide search, large and fixed problem scale, easy to fall into local optimality, slow and inefficient problem solving speed.
[0007] The present invention provides a UAV path planning method for emergency material delivery tasks, comprising the following steps:
[0008] S1. Extract the features of the first type of nodes from the optimal solution of the vehicle routing problem with capacity constraints at multiple stations for training the neural network;
[0009] S2. In the UAV path planning scenario for emergency material delivery missions, the features of the second-category nodes are extracted, the edge probability between each second-category node is predicted using the trained neural network, and the pheromone matrix is constructed;
[0010] Use the divide-and-conquer strategy to divide the rescue points in the second type of nodes into multiple rescue point groups, each rescue point group corresponds to a sub-problem;
[0011] Based on the pheromone matrix, the ant colony algorithm is used to solve the sub-problems to optimize each rescue point group, and the optimization results of each rescue point group are pieced together to form the solution to the overall problem;
[0012] The rescue point group is continuously updated and optimized until the ant colony algorithm converges.
[0013] Furthermore, S1 specifically includes:
[0014] S101. Extracting the characteristics of each first-class node of the optimal solution of the vehicle routing planning problem with multiple stations and capacity constraints;
[0015] S102. Construct a training data set using the features of the first-class nodes, where each piece of data in the training data set represents an edge of a directed graph formed by the first-class nodes, the edge feature is the concatenation of the features of the two first-class nodes contained therein, and the edge label is used to indicate whether the edge between the two first-class nodes is included in the optimal solution of the vehicle routing planning problem with capacity constraints at multiple stations;
[0016] S103. Use the binary cross entropy loss function to train the neural network BpNet, and train the neural network BpNet until convergence.
[0017] Furthermore, the structure of the neural network BpNet is a multi-layer perceptron model including two hidden layers.
[0018] Furthermore, the input layer of the neural network BpNet includes neurons, the number of which is the same as the number of edge features; the relu activation function is used between the two hidden layers; the output layer is activated using the sigmoid function to output the probability of the existence of the edge.
[0019] Furthermore, S2 specifically includes:
[0020] S201. Extract the features of the second type of nodes, including the information of the rescue point nodes and the information of the drone station nodes, and input them into the trained neural network BpNet; set the outermost loop stop condition of the ant colony algorithm, initialize the divide-and-conquer scale L0, the maximum number of iterations of the ant colony algorithm and the number of ants in the ant colony algorithm, set the divide-and-conquer scale tabulist to empty, set the minimum change a and the maximum change b of the divide-and-conquer scale, and the divide-and-conquer scale L=L0; where L is the current divide-and-conquer scale;
[0021] S202. Use the trained neural network BpNet to predict the edge probability p between all the second type nodes ij , as the initialization pheromone matrix p0, and set the current pheromone matrix P = p0;
[0022] S203. Construct an initialization solution and set the initialization solution to the optimal solution sbest;
[0023] S204. Use k-means clustering method to split the optimal solution sbest into L rescue point groups {g1, g2,,, g L};
[0024] S205. Use the improved ant colony algorithm and the current pheromone matrix P to independently optimize each rescue point group in turn, and obtain the solution {s1,s2,,,s L}, and the optimized pheromone matrix of each rescue point group {ph1,ph2,,,ph L};
[0025] S206. Merge s1, s2,,,s L Get the new solution snew of this iteration, and use the neighborhood descent algorithm to optimize snew;
[0026] S207.Merge ph1, ph2,,, ph L Get the current pheromone matrix P of this round of iteration;
[0027] S208. If the performance of snew is better than sbest, then update sbest and clear the tabulist. Otherwise, add the current partition scale L to the tabulist and randomly add the tabulist in the range [α*L o , b*L o ] select an integer as the new current partition scale L, and require that the new current partition scale L is not in the taboo table. If the taboo table contains the range [α*L o , b*L o ], then randomly select an integer in the range as the new current partition scale L;
[0028] S209. Determine whether the stopping condition is met. If yes, finally output the optimal solution sbest; otherwise, proceed to S204.
[0029] Compared with the prior art, the present invention has at least the following beneficial effects:
[0030] (1) The UAV path planning method for emergency material delivery tasks of the present invention proposes a knowledge-guided strategy for the high timeliness requirements of the algorithm in the cumulative UAV path planning problem scenario with capacity constraints at multiple stations. The optimal solution of the known problem is learned by using a neural network. The training goal is to predict the edge probability between nodes. The trained neural network is used to predict the edge probability of each node in the new scenario. The pheromone matrix of the ant colony algorithm is initialized with the neural network prediction results as heuristic information, creating a good start for the search.
[0031] (2) The UAV path planning method for emergency material delivery tasks of the present invention proposes a taboo divide-and-conquer strategy based on the characteristics of the large number and scale of rescue points in the problem. The k-means method is used to divide the rescue points into multiple groups for solution, thereby reducing the scale of the problem. At the same time, the divide-and-conquer scale is adaptively changed according to the search status, effectively enhancing the diversity of divide-and-conquer and helping the algorithm to escape from the local optimum.
[0032] (3) The UAV path planning method for emergency material delivery tasks of the present invention adds a solution repair mechanism and a local search algorithm to the ant colony algorithm to address the complex coupling of problem constraints and the difficulty in constructing a solution for the ant colony algorithm, thereby effectively improving the local exploration performance of the ant colony algorithm and the utilization efficiency of the solution, thereby achieving efficient and rapid problem solving.
[0033] (4) The present invention focuses on the problem of drone path planning with multiple departure stations. Starting from minimizing the cumulative time of drone delivery, the paper studies the solution method to this problem, which is helpful to shorten the waiting time for rescue material delivery and efficiently plan the drone delivery path. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] The drawings are only for the purpose of illustrating particular embodiments and are not to be construed as limiting the invention.
[0035] Figure 1 A schematic diagram of the UAV route disclosed in the present invention;
[0036] Figure 2 It is a schematic diagram of the structure of the neural network disclosed in the present invention;
[0037] Figure 3 A training flow chart of a neural network disclosed in the present invention;
[0038] Figure 4 It is a flow chart of the improved ant colony algorithm disclosed in the present invention;
[0039] Figure 5 The overall algorithm flow chart of the UAV path planning for emergency material delivery mission disclosed in the present invention;
[0040] Figure 6 It is a flow chart of the variable neighborhood descent algorithm disclosed in the present invention;
[0041] Figure 7 It is a schematic diagram of the local search operator disclosed in the present invention. DETAILED DESCRIPTION
[0042] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein, and therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.
[0043] The present invention discloses a UAV path planning method for emergency material delivery tasks, comprising the following steps:
[0044] S1. Extract the features of the first type of nodes from the optimal solution of the vehicle routing planning problem with capacity constraints at multiple stations for training neural networks.
[0045] S101. Extract the features of each first-category node of the optimal solution to the vehicle routing planning problem with capacity constraints at multiple stations.
[0046] Specifically, the multi-station capacity-constrained vehicle routing planning (MDCVRP) problem instance contains multiple first-class nodes, including customer points and vehicle stations. Each customer point has a certain material demand. The locations of all customer points and vehicle stations are known. Each vehicle station has a certain number of vehicles. The optimal solution contains the path of the customer points served by each vehicle. The path is specifically the first-class node sequence from the vehicle station where the vehicle departs and the customer points it serves in turn and finally returns to the vehicle station. The features of each node of the public optimal solution of the multi-station capacity-constrained vehicle routing planning problem are extracted, and the feature F of a single first-class node i′ is extracted. i′ There are 6 nodes in total. The first type of node i′ is a customer point or a vehicle station. Then F i′ Specifically include:
[0047] The demand of the first-class node; if the current first-class node is a vehicle station, the demand is 0; if the current first-class node is a customer point, it is the demand for materials at the customer point.
[0048] The horizontal coordinate of the first-type node position; if the first-type node position is expressed in longitude and latitude, the horizontal coordinate of the first-type node position is the longitude.
[0049] The ordinate of the first-type node position; if the first-type node position is expressed in longitude and latitude, the ordinate of the first-type node position is the latitude.
[0050] The minimum distance between the first type of node and the vehicle station.
[0051] The maximum distance between the first type of node and the vehicle station.
[0052] The average distance between the first type of nodes and the vehicle depot.
[0053] S102. Use the features of the first-class nodes to construct a training data set, where each piece of data in the training data set represents an edge in a directed graph consisting of the first-class nodes. The edge feature is the concatenation of the features of the two first-class nodes it contains. The edge label is used to indicate whether the edge between the two first-class nodes is included in the optimal solution to the vehicle path planning problem with capacity constraints at multiple stations. If included, the value is 1, otherwise it is 0.
[0054] For example, for the first-type node i′ and the first-type node j′, the edge e formed by them i′j′ The characteristic is (F i′ ,F j′ ), the edge feature dimension is 12, and the data label is whether the optimal solution contains the edge e consisting of these two nodes i′j′ , which is 1 if included, 0 otherwise.
[0055] When constructing the training data set, we first extract the edge features of the optimal solution, whose label is 1, and then randomly sample some edges from the directed graph formed by the first type of nodes to ensure that the ratio of positive and negative samples is approximately 1:2, to prevent the problem of difficulty in training due to uneven ratio of positive and negative samples.
[0056] S103. Use the binary cross entropy loss function to train the neural network BpNet, where the structure of the neural network BpNet is a multi-layer perceptron model including two hidden layers, and train T rounds on the training data set until the neural network BpNet converges.
[0057] The structure of the neural network is shown in Figure 2 , where the input layer has 12 neurons, corresponding to the number of edge features, the first hidden layer has 64 neurons, the second hidden layer has 32 neurons, and the hidden layers use the relu activation function. The output layer size is 1 and is activated using the sigmoid function, that is, the output is the probability of the existence of an edge. The training of the neural network uses the binary cross entropy loss function, which is defined as follows:
[0058]
[0059] Where n is the number of samples, x index is the sample probability predicted by the neural network, y index is the binary label of the sample, and index represents the sample number.
[0060] Use Adam optimizer to train for T rounds until the neural network converges. The overall process of neural network training is shown in Figure 3 .
[0061] S2. The UAV path planning scenario problem instance for the emergency material delivery mission contains multiple second-class nodes. The features of the second-class nodes are extracted, and the edge probability between each second-class node is predicted using the trained neural network to construct a pheromone matrix.
[0062] The divide-and-conquer strategy is used to divide the rescue points in the second type of nodes into multiple rescue point groups, each rescue point group corresponds to a sub-problem.
[0063] Based on the pheromone matrix, the ant colony algorithm is used to solve and optimize each rescue point group, and the optimization results of each rescue point group are pieced together into the solution to the overall problem.
[0064] The rescue point group is continuously updated and optimized until the ant colony algorithm converges.
[0065] S201. Extract the features of the second type of nodes in the same way as the first type of nodes, including rescue point information and drone station information, and input the trained neural network BpNet; set the outermost loop stop condition of the ant colony algorithm, initialize the divide-and-conquer scale L0, the maximum number of iterations of the ant colony algorithm and the number of ants in the ant colony algorithm, set the divide-and-conquer scale taboo table tobulist to empty, set the minimum change a and the maximum change b of the divide-and-conquer scale, and the divide-and-conquer scale L=L0; where L is the current divide-and-conquer scale.
[0066] Specifically, in the UAV path planning scenario for emergency material delivery missions, the second-class nodes include C rescue point nodes and N UAV station nodes. Each UAV station contains M UAVs, the maximum flight distance and maximum material capacity of each UAV, the time required for UAV flight between nodes, the number of materials required for each customer point, and the coordinate position of each node. The feature F of a single second-class node i is i There are 6 nodes in total. The second type of node i is a rescue point or a drone station, then F i Specifically include:
[0067] The demand for the second-class nodes; if the current second-class node is a drone station, the demand is 0; if the current second-class node is a rescue point, it is the demand for materials at the rescue point.
[0068] The horizontal coordinate of the second type node position; if the second type node position is expressed in longitude and latitude, the horizontal coordinate of the second type node position is the longitude.
[0069] The ordinate of the second-type node position; if the second-type node position is expressed in longitude and latitude, the ordinate of the second-type node position is the latitude.
[0070] The minimum distance between the second type of node and the drone station.
[0071] The maximum distance between the second type of node and the drone station.
[0072] The average distance between the second type of nodes and the drone station.
[0073] Exemplarily, specifically, the ant colony algorithm loop stop conditions are set, including the maximum number of iterations Maxiter and the maximum number of no-improvement iterations Noimproveiter. The loop ends if the optimal solution is not updated after Noimproveiter cycles. The initial divide-and-conquer scale L0 is set, that is, the number of rescue point groups obtained by divide-and-conquer. The maximum number of iterations Maxacoiter and the number of ants in the ant colony algorithm Numant are set. The optimal solution sbest is set to empty. The divide-and-conquer scale tabulist is set to empty. The divide-and-conquer scale change a=0.5, b=1.5, the divide-and-conquer scale L=L0, the current number of iterations iter is set to 1, and the current number of no-improvement iterations niter is set to 0.
[0074] S202. Use the trained neural network BpNet to predict the edge probability p between all the second type nodes ij , as the initialization pheromone matrix P0, and set the current pheromone matrix P = P0.
[0075] Specifically, the trained neural network BpNet is used to predict the probability P of all nodes containing the optimal solution edge ij :
[0076]
[0077] in, Represents the second type of node edge e ij For a graph G consisting of C rescue points and N drone stations, the neural network BpNet is used to predict the probability P of each edge of the graph G being the optimal solution edge. ij , as the initialization pheromone matrix P0, the pheromone matrix contains the probability of transfer between each second-type node. The higher the transfer probability, the greater the possibility of the optimal solution edge. The pheromone matrix generated by the neural network creates a start for the search based on prior knowledge, reducing the blindness of the early search.
[0078] S203. Construct an initialization solution and set the initialization solution to the optimal solution sbest.
[0079] Specifically, an initialization solution is constructed, which includes M*N paths and a total of M*N drones. The initialization solution is set as the optimal solution sbest. The specific construction process of the initialization solution is as follows:
[0080] S203-1. First, set the paths of all drones to be empty, that is, only include drone stations. For example, the solution of the drone starting from drone station 0 is represented as {d0, d0}, and set the remaining rescue points to be inserted to all rescue points.
[0081] S203-2. Calculate the insertion cost of the remaining rescue points to be inserted relative to all M*N paths. The insertion cost is defined as the additional cost of inserting the rescue point to the end of the path. Let R rp is the difference between the minimum insertion cost of the rescue point rp relative to all paths and the second minimum value. If there is no second minimum, return the minimum value and record the R rp The rescue point rp with the largest value.
[0082] S203-3. Add the rescue point i to the tail of the path with the minimum insertion cost, and delete the rescue point i from the remaining rescue points to be inserted.
[0083] S203-4. Repeat steps S203-2 to S203-3 until all rescue points are inserted.
[0084] S203-5. Return the constructed initialization solution.
[0085] S204. Use k-means clustering method to split the optimal solution sbest into L rescue point groups {g1, g2,,, g L}.
[0086] Specifically, we calculate the centroids of each path in the optimal solution sbest {b1,b2,,,b M*N}. The calculation method of the center of gravity is as follows:
[0087]
[0088] Among them, |V index | represents the length of the path index, x v ,y v Represents the horizontal and vertical coordinates of node v. There are M*N paths in total, forming M*N data points. The k-means method is used to divide the M*N paths into L groups.
[0089] It should be noted that since each rescue point is only visited once, each set of questions is independent of each other and has no overlap.
[0090] S205. Use the improved ant colony algorithm and the current pheromone matrix P to independently optimize each rescue point group in turn, and obtain the solution {s1,s2,,,s L}, and the optimized pheromone matrix of each rescue point group {ph1,ph2,,,ph L}.
[0091] The improved ant colony algorithm process is as follows:
[0092] S205-1. Input current solution g index , pheromone matrix P, let the ant colony algorithm optimal solution gbest = s index , where index represents the sequence number of the group, extract the rescue point C contained in the current solution index and drone D index Set the current ant colony algorithm iteration number acoiter to 1; Set the number of solutions currently constructed by the ant colony algorithm K = 1;
[0093] S205-2. Based on rescue point C index 、UAV D index The collection of constructs the Kth solution sol K The specific process of solution construction is: maintain a set of current remaining rescue points RC = C index , the current set of remaining drones RD = D i , randomly selected from the drone station set D i Select a starting point of the drone, set it as the current position, and calculate the transition probability prob from the current position i to the node j in RC ij :If node i departs to serve node j, it can meet the capacity and power constraints of the drone, then Otherwise, prob ij =0, where P ij is the transition probability from node i to node j in the pheromone matrix, is the heuristic information, dist ij represents the distance between node i and node j, α is the weight of pheromone concentration, set to 2, β is the heuristic information weight, set to 1, if all node transfer probabilities are 0, it means that the current drone can no longer serve any node, insert the starting point of the drone to the end of the path, representing the return of the drone, save the path of the current drone in the solution, delete the current drone from RD, and then randomly select a drone from RD to continue to build the solution. Otherwise, use the probability based on p ij Use roulette to select a node j to insert into the path of the current drone, delete the node j from RC, and set node j as the current node. Repeat the node selection process until RC is empty (all rescue points have been visited) or RD is empty (all drones have been used).
[0094] S205-3. If RC is empty, skip it, otherwise it represents the current solution sol K There are still rescue points that have not been served. The following solution repair mechanism is used: Calculate the insertion cost of the remaining rescue points a relative to all positions in the solution In order to increase the objective function value by inserting rescue point a between nodes i and j, the position with the minimum insertion cost is selected, rescue point a is inserted, and a is deleted from RC. The insertion process is performed sequentially until RC is empty or the insertion of rescue point a at all current positions cannot meet the power and capacity constraints of the drone.
[0095] S205-4. If RC is not empty, skip it, otherwise use the variable neighborhood descent algorithm (VND) to optimize the solution sol K The specific process is as follows Figure 6 shown.
[0096] S205-4-1. Initial solution s, define k = 6 neighborhoods, denoted as Ω m (m=1,2,3…,6).
[0097] S205-4-2.index=1.
[0098] S205-4-3. Using the neighborhood structure Ω i The search continues until it reaches the local optimal solution s1.
[0099] S205-4-4. If s1 is better than s, let s=s1, .index=1; otherwise, index=index+1.
[0100] S205-4-5. If index≤k, return to S205-4-2.
[0101] S205-4-6. Output the optimal solution s.
[0102] The six neighborhood operations Ω1-Ω6 used in the neighborhood descent algorithm are as follows:
[0103] The operation of Ω1 (resetting different node positions) is: deleting a rescue point node from its original position and inserting it into a position on another path.
[0104] The operation of Ω2 (same node position reset) is: delete a rescue point node from its original position and select a position of the current path to insert it.
[0105] The Ω3 (node position exchange) operation is: exchanging the positions of two rescue point nodes in different paths.
[0106] The Ω4 (2-Opt) operation is: exchanging two segments containing two consecutive rescue point nodes in different paths.
[0107] The Ω5 (Or-Opt) operation is: select a segment containing three consecutive rescue points and reinsert it at another location.
[0108] The Ω6 (four-node position swap) operation is: select two rescue point nodes a and b on one path and two rescue point nodes m and n on another path, swap the positions of rescue point node a and rescue point node n, and swap the positions of rescue point node b and rescue point node m.
[0109] Figure 7 The specific operation methods of the Ω1-Ω6 operators are given.
[0110] S205-5. If sol K is a feasible solution (RC is empty) and sol K The performance of is better than the optimal solution gbest, then update the optimal solution gbest = sol K
[0111] S205-6. Let K=K+1. If K=Numant, it means that the construction of the solution for Numant ants has been completed. K=1, proceed to S205-7. Otherwise, proceed to S205-2 to continue constructing the solution.
[0112] S205-7. Update the pheromone matrix P. The updating method is: if gbest passes through the path (i, j), then
[0113]
[0114] Among them, Q is the pheromone increase strength, which is set to 100, and cost(gbest) represents the objective function value of gbest.
[0115] S205-8. Let acoiter = acoiter + 1. If acoiter = Maxacoiter, it means that the maximum number of iterations has been reached, and the optimal solution gbest and pheromone matrix P are output. The ant colony algorithm process is detailed in Figure 4 .
[0116] S206. Merge s1, s2,,,s L The new solution snew of this iteration is obtained, and snew is optimized using the Versatile Neighborhood Descent (VND) algorithm.
[0117] S207.Merge ph1, ph2,,, ph L Get the current pheromone matrix P of this round of iteration.
[0118] S208. If the performance of snew is better than sbeat, then update sbeat and clear the tabulist. Otherwise, add the current partition scale L to the tabulist and randomly add the tabulist in the range [α*L o , b*L o] select an integer as the new current partition scale L, and require that the new current partition scale L is not in the taboo table. If the taboo table contains the range [α*L o , b*L o ], then randomly select an integer in the range as the new current partition scale L.
[0119] S209. Determine whether the stopping condition is met. If yes, finally output the optimal solution sbest; otherwise, proceed to S204.
[0120] Specifically, iter=iter+1. If the current number of iterations iter is equal to the maximum number of iterations Maxiter or the current number of no-improvement iterations niter is equal to Noimproveiter, the algorithm stopping condition is met and the optimal solution sbest is output, otherwise S204 is performed.
[0121] The overall process of UAV path planning for emergency material delivery tasks is as follows: Figure 5 shown.
[0122] The mathematical model of the UAV path planning problem for emergency material delivery missions and the meaning of the symbols in the model are as follows:
[0123]
[0124]
[0125] Where: C represents the rescue point set, D represents the drone station set, V = C ∪ D represents the union of the rescue point node set and the drone station node set, R represents the available drone set, Q represents the capacity of the drone to carry supplies, and E represents the battery capacity of the drone; Indicates whether drone k starts from node i and visits node j. If yes, it takes 1, otherwise it takes 0; represents the time when UAV k arrives at the i-th node; represents the remaining power of UAV k when it reaches the i-th node; α ij represents the amount of power consumed by the drone from the i-th node to the j-th node; t ij represents the time it takes for the drone to reach the jth node from the i-th node; q j Represents the demand of the j-th node; M is an artificially set "infinite" number.
[0126] The meaning of the formula is:
[0127] Formula (1) represents the objective function of this problem, which is to minimize the total cumulative time required to serve all rescue points;
[0128] Formula (2) means that there can only be one path from node i to node j;
[0129] Formula (3) means that any UAV must leave after visiting a rescue point;
[0130] Formula (4) means that any UAV must start and end from the UAV station;
[0131] Formula (5) means that the demand for any rescue point served by a drone cannot exceed the maximum capacity of the supplies it carries;
[0132] Formula (6) means that if a drone visits node j from node i, then the time it takes for the drone to reach node i must be less than the time it takes to reach node j;
[0133] Formula (7) means that if a drone visits node j from node i, then the remaining power of the drone at node i must be greater than the remaining power at node j;
[0134] Formula (8) indicates that after visiting a rescue point, the UAV must have enough power left to return to the UAV station;
[0135] Formula (9) indicates that the time when the UAV arrives at the UAV station is 0 (the time when it departs is 0);
[0136] Formula (10) indicates that the time it takes for the drone to reach any rescue point should be greater than 0;
[0137] Formula (11) indicates that the starting power of the drone is its battery capacity (the drone is fully charged when it departs);
[0138] Formula (12) indicates that the remaining power of the drone when it reaches any rescue point should be less than its battery capacity;
[0139] Formula (13) represents the variable The value range is 0 or 1.
[0140] Compared with the prior art, the UAV path planning method for emergency material delivery tasks of the present invention proposes a knowledge-guided strategy for the high timeliness requirements of the algorithm in the cumulative UAV path planning problem scenario with capacity constraints in multiple stations. The optimal solution of the known problem is learned by using a neural network. The training goal is to predict the edge probability between nodes. The trained neural network is used to predict the edge probability of each node in the new scenario. The pheromone matrix of the ant colony algorithm is initialized with the neural network prediction result as heuristic information, creating a good start for the search. In view of the large number and scale of rescue points in the problem, a taboo divide-and-conquer strategy is proposed, using k-means The s method divides the rescue points into multiple groups for solution, reducing the scale of the problem. At the same time, it adaptively changes the division scale according to the search status, effectively enhancing the diversity of division and helping the algorithm to jump out of the local optimum. In view of the complex coupling of problem constraints and the difficulty in constructing the solution of the ant colony algorithm, a solution repair mechanism and a local search algorithm are added to the ant colony algorithm to effectively improve the local exploration performance and solution utilization efficiency of the ant colony algorithm, thereby achieving efficient and rapid solution to the problem. The present invention focuses on the path planning problem of drones with multiple departure stations. Starting from minimizing the cumulative time of drone delivery, a study on the solution method of this problem is carried out, which is conducive to shortening the waiting time for rescue material delivery and efficiently planning drone delivery paths.
[0141] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A UAV path planning method for emergency material delivery tasks, characterized in that: The steps include: S1. Extract the features of the first type of nodes from the optimal solution of the vehicle routing problem with capacity constraints at multiple stations for training the neural network; S2. In the UAV path planning scenario for emergency material delivery missions, the features of the second-category nodes are extracted, the edge probability between each second-category node is predicted using the trained neural network, and the pheromone matrix is constructed; Use the divide-and-conquer strategy to divide the rescue points in the second type of nodes into multiple rescue point groups, each rescue point group corresponds to a sub-problem; Based on the pheromone matrix, the ant colony algorithm is used to solve the sub-problems to optimize each rescue point group, and the optimization results of each rescue point group are pieced together to form the solution to the overall problem; The rescue point group is continuously updated and optimized until the ant colony algorithm converges.
2. The method for UAV path planning for emergency material delivery mission according to claim 1 is characterized in that: S1 specifically includes: S101. Extracting the characteristics of each first-class node of the optimal solution of the vehicle routing planning problem with multiple stations and capacity constraints; S102. Construct a training data set using the features of the first-class nodes, where each piece of data in the training data set represents an edge of a directed graph formed by the first-class nodes, the edge feature is the concatenation of the features of the two first-class nodes contained therein, and the edge label is used to indicate whether the edge between the two first-class nodes is included in the optimal solution of the vehicle routing planning problem with capacity constraints at multiple stations; S103. Use the binary cross entropy loss function to train the neural network BpNet, and train the neural network BpNet until convergence.
3. The method for UAV path planning for emergency material delivery mission according to claim 2 is characterized in that: The structure of the neural network BpNet is a multi-layer perceptron model including two hidden layers.
4. The method for UAV path planning for emergency material delivery mission according to claim 3 is characterized in that: The input layer of the neural network BpNet includes neurons, whose number is the same as the number of edge features; the relu activation function is used between the two hidden layers; the output layer is activated by the sigmoid function and outputs the probability of the existence of an edge.
5. The method for UAV path planning for emergency material delivery mission according to claim 4, characterized in that: S2 specifically includes: S201. Extract the features of the second type of nodes, including the information of the rescue point nodes and the information of the drone station nodes, and input them into the trained neural network BpNet; set the outermost loop stop condition of the ant colony algorithm, initialize the divide-and-conquer scale L0, the maximum number of iterations of the ant colony algorithm and the number of ants in the ant colony algorithm, set the divide-and-conquer scale tabulist to empty, set the minimum change a and the maximum change b of the divide-and-conquer scale, and the divide-and-conquer scale L=L0; where L is the current divide-and-conquer scale; S202. Use the trained neural network BpNet to predict the edge probability p between all the second type nodes ij , as the initialization pheromone matrix P0, and set the current pheromone matrix P = P0; S203. Construct an initialization solution and set the initialization solution to the optimal solution sbest; S204. Use k-means clustering method to split the optimal solution sbest into L rescue point groups {g1, g2,,, g L }; S205. Use the improved ant colony algorithm and the current pheromone matrix P to independently optimize each rescue point group in turn, and obtain the solution {s1,s2,,,s L }, and the optimized pheromone matrix of each rescue point group {ph1,ph2,,,ph L }; S206. Merge s1, s2,,,s L Get the new solution snew of this iteration, and use the neighborhood descent algorithm to optimize snew; S207.Merge ph1, ph2,,, ph L Get the current pheromone matrix P of this round of iteration; S208. If the performance of snew is better than sbest, then update sbest and clear the tabulist. Otherwise, add the current partition scale L to the tabulist and randomly add the tabulist in the range [α*L o , b*L o ] select an integer as the new current partition scale L, and require that the new current partition scale L is not in the taboo table. If the taboo table contains the range [α*L o , b*L o ], then randomly select an integer in the range as the new current partition scale L; S209. Determine whether the stopping condition is met. If yes, finally output the optimal solution sbest; otherwise, proceed to S204.
Citation Information
Cited By
Park express low-altitude commissioning path planning method
CN119599237A
Helicopter rescue scheduling method and system based on large neighborhood search algorithm
CN120509700A
