A Material Delivery Method and System for Overloaded Drones

Abstract

The present invention relates to the technical field of material delivery, and specifically to a method and system for material delivery using heavy-load unmanned aerial vehicles, including: obtaining a rescue request within a disaster-stricken area, determining the location of the rescue point according to the rescue request, and determining the material transportation path of the heavy-load unmanned aerial vehicle cluster according to the location of the rescue point; controlling the heavy-load unmanned aerial vehicle cluster to transport materials to the air above the location of the rescue point according to the material transportation path, and performing image acquisition on the location of the rescue point by the heavy-load unmanned aerial vehicle to obtain an environmental image of the rescue point. By obtaining the rescue request within the disaster-stricken area and determining the location of the rescue point, the present invention can ensure that the unmanned aerial vehicle cluster accurately transports materials to the required area, thereby improving the efficiency of material distribution. Moreover, controlling the heavy-load unmanned aerial vehicle cluster to execute tasks according to the preset material transportation path can reduce the empty flight and unnecessary deviation of the unmanned aerial vehicles.

Description

Technical Field

[0001] The present invention relates to the technical field of material delivery, and specifically relates to a method and system for material delivery of heavy-lift drones. Background Art

[0002] A heavy-lift drone is a type of drone that can carry and transport heavier items. Different from conventional consumer drones, heavy-lift drones have a stronger power system, a larger airframe, and a higher load capacity, and are suitable for situations where a relatively large weight needs to be carried during transportation and mission execution.

[0003] Traditional material delivery methods, such as relying on transportation tools like trucks and helicopters, usually require a long transportation time and multiple transfers, resulting in low efficiency. In post-disaster rescue, traditional methods may slow down the response time of material delivery due to problems such as road blockages and traffic restrictions; moreover, traditional methods lack intelligent scheduling and path optimization, which may lead to empty flights and deviation from the path during the material distribution process, increasing the time cost of transportation and reducing efficiency; furthermore, traditional methods usually lack the ability to recognize the environment in real time and cannot make quick judgments based on the actual situation on site. Even in complex terrains or environments where landing is not suitable, traditional methods often cannot identify them in advance, resulting in failed material delivery or malfunctions of the drones; in addition, traditional material delivery methods rely on manual operation or fixed-point landing, and may not be able to accurately deliver materials due to misjudgment or unsuitable landing environments, causing the materials to not be delivered to the required places on time. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a method and system for material delivery of heavy-lift drones.

[0005] The technical solution adopted to solve the above technical problem is: A method for material delivery of heavy-lift drones, comprising:

[0006] Obtain rescue requests within the disaster-stricken area, determine the location of the rescue point according to the rescue requests, and determine the material transportation path of the heavy-lift drone cluster according to the location of the rescue point;

[0007] Control the heavy-lift drone cluster to transport materials to the air above the location of the rescue point according to the material transportation path, and collect images of the rescue point by the heavy-lift drones to obtain environmental images of the rescue point;

[0008] Perform environmental recognition on the environmental images according to a pre-trained environmental recognition model to obtain the environmental state of the rescue point, wherein the environmental state includes a state where landing is possible and a state where landing is not possible;

[0009] When the environmental state is a landable state, the heavy-duty drone is controlled to land at the rescue point according to a pre-trained landing control model;

[0010] When the environmental state is a non-landable state, the heavy-duty drone is directly controlled to land at a preset height, and the supplies are dropped at the rescue point.

[0011] Preferably, the position of the rescue point is determined according to the rescue request, and the material transportation path of the heavy-duty drone cluster is determined according to the position of the rescue point, including:

[0012] Obtain a preset list of material transportation paths, where the material transportation path includes at least one rescue point position and a heavy-duty drone base;

[0013] Randomly select a material transportation path from the list of material transportation paths and add it to a preset current path list, and add the remaining material transportation paths in the list of material transportation paths to a preset remaining path list;

[0014] Calculate the distance similarity between the material transportation path in the current path list and each material transportation path in the remaining path list;

[0015] Select the material transportation path with the smallest distance similarity, add the material transportation path with the smallest distance similarity to the current path list, and delete the material transportation path with the smallest distance similarity from the remaining path list;

[0016] Repeat the above operations until the number of rescue point positions in the current path list reaches a preset quantity threshold to obtain the path grouping of the heavy-duty drone cluster, and add the path grouping to a preset grouping list;

[0017] Judge whether the remaining path list is empty. If it is empty, output each path grouping in the grouping list, where one heavy-duty drone in the heavy-duty drone cluster is responsible for one path grouping in the grouping list. Otherwise, repeat the above operations.

[0018] Preferably, the calculation formula of the distance similarity is as follows:

[0019] ;

[0020] Wherein, represents the distance similarity, and represent path and path , and represent path and path The number of rescue points and heavy-lift UAV bases included therein, and representing the path and the path of the heavy-lift UAV base, representing a random number uniformly distributed in the range of 0-1, representing the rescue point and the rescue point the distance between.

[0021] Preferably, the environment recognition model adopts an improved YOLO model, and the backbone network of the YOLO model is improved according to the feature extraction module and the feature fusion module. Among them, the feature extraction module includes four branches, each branch corresponding to a different receptive field size. Among them, the first branch adopts a residual structure. Before executing the remaining three branches, the number of channels of the input feature map is reduced from 2C to C by using a 1×1 convolution. The second branch extracts small receptive field features through a 3×3 convolution to obtain a first output feature map. The third branch uses a DRB module with K = 7 to extract medium receptive field features to obtain a second output feature map. The fourth branch uses a DRB module with K = 15 to extract large receptive field features to obtain a third output feature map. The first output feature map is added element-wise to the output feature map of the first branch to obtain a fourth output feature map. The second output feature map is added element-wise to the output feature map of the first branch to obtain a fifth output feature map. The third output feature map is added element-wise to the output feature map of the first branch to obtain a sixth output feature map.

[0022] Preferably, the feature fusion module uses a parallel structure of max pooling and average pooling to downsample the fourth output feature map to obtain a seventh output feature map, uses the nearest neighbor interpolation method to upsample the sixth output feature map to obtain an eighth output feature map, divides the seventh output feature map, the fifth output feature map, and the eighth output feature map into three equal sub-feature maps in the channel dimension, applies the Sigmoid function to the three equal channel sub-feature maps of the seventh output feature map, the fifth output feature map, and the eighth output feature map respectively through a feature adaptive selection fusion mechanism to calculate the selection weights, and fuses the three equal channel sub-feature maps of the seventh output feature map, the fifth output feature map, and the eighth output feature map according to the selection weights to obtain a first fusion feature, a second fusion feature, and a third fusion feature, and splices the first fusion feature, the second fusion feature, and the third fusion feature in the channel dimension to obtain an output fusion feature.

[0023] Preferably, the landing control model includes a Markov decision model and a deep reinforcement learning model. Among them, the Markov decision model includes a global state space set, an action space set of the heavy-load UAV cluster, a state transition probability, a discount factor, and a reward value. Among them, the global state space set includes a first state, a second state, and a third state. The action space set includes a height control action and a steering angle control action. The state transition probability represents the state transition probability that the global state jumps to the next state when the heavy-load UAV cluster takes an action. The discount factor represents the influence of subsequent states on the current action decision. The reward value includes a first reward, a second reward, and a third reward. The deep reinforcement learning model is used to solve the Markov decision model to obtain an optimal landing strategy.

[0024] Preferably, the first state represents the relative position information between the heavy-load UAV and the rescue point. The first state includes the distance and relative direction angle between the heavy-load UAV and the rescue point. The second state represents the relative information of the neighboring heavy-load UAVs. The second state includes the distance and relative direction angle between the two heavy-load UAVs with the closest distance within the communication range. The third state represents the obstacle distance obtained by lidar. The expression of the steering angle control action is as follows:

[0025] ;

[0026] Among them, represents the steering angle of the th heavy-load UAV, represents the th heavy-load UAV, represents the absolute value of the maximum steering angle of the heavy-load UAV, represents the number of equal divisions of the steering range of the heavy-load UAV.

[0027] Preferably, the expression of the first reward is as follows:

[0028] ;

[0029] Among them, represents the first reward, and represent preset weights, and , represents the set of rescue points near the heavy-load UAV, represents the distance between the heavy-load UAV and the target point at time t;

[0030] The expression of the second reward is as follows:

[0031] ;

[0032] Among them, represents the second reward, represents the minimum distance between the heavy-load drone and the center of the influence range of surrounding obstacles, represents the influence radius of the obstacle, represents the obstacle avoidance buffer distance, represents the safety distance of the heavy-load drone, represents the allowable radius of the obstacle, represents the allowable straight-line distance between the heavy-load drone and the obstacle, represents the first preset reward value, represents the second preset reward value;

[0033] The expression of the third reward is as follows:

[0034] ;

[0035] Among them, represents the third reward, represents the minimum distance between the heavy-load drone and other heavy-load drones, represents the minimum distance that should be maintained between heavy-load drones.

[0036] Preferably, the deep reinforcement learning model is used to solve the Markov decision model to obtain an optimal landing strategy, including:

[0037] Initialize the departure position and target position of the drone;

[0038] Select an action uniformly at random from the set of global state spaces with a preset probability;

[0039] Execute the action, and at the same time obtain the next moment state and reward;

[0040] Put the current trajectory into the experience replay buffer, randomly and uniformly collect multiple trajectories, and update the main network parameters according to the loss function for each trajectory, where the trajectory includes the current state, the current action, the next moment state and the reward;

[0041] Update the target network parameters using the main network parameters every preset number of training steps.

[0042] The technical solution adopted to solve the above technical problems is: A material delivery system for a heavy-load drone, which is applicable to the above-mentioned material delivery method for a heavy-load drone, including:

[0043] A path planning unit, which is used to obtain rescue requests within the disaster area, determine the location of the rescue point according to the rescue requests, and determine the material transportation path of the heavy-load UAV cluster according to the location of the rescue point;

[0044] An environment acquisition unit, which is used to control the heavy-load UAV cluster to transport materials to the air above the rescue point location according to the material transportation path, and collect images of the rescue point location by the heavy-load UAV to obtain the environmental image of the rescue point;

[0045] An environment recognition unit, which is used to perform environment recognition on the environmental image according to a pre-trained environment recognition model to obtain the environmental state of the rescue point, where the environmental state includes a landable state and a non-landable state;

[0046] A landing control unit, which is used to control the heavy-load UAV to land at the rescue point according to a pre-trained landing control model when the environmental state is the landable state;

[0047] A material delivery unit, which is used to directly control the heavy-load UAV to land at a preset height and deliver the materials to the rescue point when the environmental state is the non-landable state.

[0048] The beneficial effects of the present invention are as follows: (1) By obtaining rescue requests within the disaster area and determining the location of the rescue point, the present invention can ensure that the UAV cluster accurately transports materials to the required area, thereby improving the efficiency of material distribution. Moreover, controlling the heavy-load UAV cluster to execute tasks according to the preset material transportation path can reduce the empty flight and unnecessary deviation of the UAVs, optimize the path planning, thereby saving time and improving the accuracy of material distribution; (2) By collecting images of the rescue point environment and applying a pre-trained environment recognition model, the system can identify and judge the environmental state of the rescue point. This intelligent environment recognition and judgment can avoid failures or accidents caused by unsafe landings of UAVs, and improve the safety and reliability of material delivery; (3) When the environmental state is the "landable state", the heavy-load UAV can land safely and complete the material delivery; while in the "non-landable state", the UAV will choose to land directly at the preset height and deliver the materials, which avoids the risk that the UAV cannot complete the task due to unsuitable landing environment, and ensures that the materials can be successfully delivered under different environmental conditions. Whether it can land or not, it can ensure that the materials reach the required rescue point in time. Description of the Drawings

[0049] Figure 1 It is a schematic diagram of the step flow of the overall method in an embodiment proposed by the present invention;

[0050] Figure 2 Schematic diagram of the system architecture of the overall system in an embodiment proposed by the present invention.

[0051] Reference numerals: 1, path planning unit; 2, environment acquisition unit; 3, environment recognition unit; 4, landing control unit; 5, material delivery unit. Detailed implementation manners

[0052] Embodiment 1, as Figure 1 shown, a material delivery method for a heavy-duty unmanned aerial vehicle proposed by the present invention includes:

[0053] S1. Obtain a rescue request within the disaster area, determine the location of the rescue point according to the rescue request, and determine the material transportation path of the heavy-duty unmanned aerial vehicle cluster according to the location of the rescue point;

[0054] S2. Control the heavy-duty unmanned aerial vehicle cluster to transport the materials to the air above the location of the rescue point according to the material transportation path, and perform image acquisition on the location of the rescue point by using the heavy-duty unmanned aerial vehicle to obtain an environmental image of the rescue point;

[0055] S3. Perform environmental recognition on the environmental image according to a pre-trained environmental recognition model to obtain the environmental state of the rescue point, where the environmental state includes a landable state and a non-landable state;

[0056] S4. When the environmental state is a landable state, control the heavy-duty unmanned aerial vehicle to land at the rescue point according to a pre-trained landing control model;

[0057] S5. When the environmental state is a non-landable state, directly control the heavy-duty unmanned aerial vehicle to land at a preset height and deliver the materials to the rescue point.

[0058] In the present invention, first, rescue requests within the disaster-stricken area need to be collected. These requests usually come from rescue personnel or relevant departments in the disaster area, indicating the need to urgently transport supplies to specific locations. After receiving the rescue requests, the system will determine the location of the rescue points where the supplies need to be delivered according to the content of the requests. This location is usually calibrated by positioning technologies such as GPS coordinates to accurately determine the specific location where the supplies need to be delivered. The drone swarm will start to execute the transportation task according to the supply transportation route, carrying rescue supplies and flying to the designated rescue points. During the transportation process, the drone swarm will fly along the predetermined route to ensure timely and accurate arrival at the target location. When the drone swarm reaches above the rescue point, the drones will conduct image acquisition, obtaining the environmental images around the rescue point through the on-board cameras or sensors. If the environmental recognition model determines that the conditions in the rescue point area are suitable for landing (for example, there are no obstacles, the ground is flat, and there are no power lines, etc.), it is regarded as a state where landing is possible. If it is determined that the environmental conditions are not suitable for landing (for example, the ground is uneven, there are obstacles, or there are other safety hazards), it is regarded as a state where landing is not possible. If it is determined that the environment is suitable for landing, a pre-trained landing control model will be used to control the landing process of the heavy-load drone. The landing control model will adjust the attitude, speed, etc. of the drone according to the real-time environmental information to ensure that the drone can safely land at the rescue point. If the environmental recognition model determines that the environment is not suitable for landing, the system will control the drone to land at a preset safe altitude. Then, the heavy-load drone will execute the supply delivery at the preset altitude, and the supplies will be accurately delivered to the rescue point area by means of aerial delivery. In this way, even if landing is not possible, the drone can still deliver the required supplies.

[0059] Embodiment 2. A method for delivering supplies using a heavy-load drone proposed by the present invention. Compared with Embodiment 1, this embodiment further includes: determining the location of the rescue point according to the rescue request, and determining the supply transportation route of the heavy-load drone swarm according to the location of the rescue point, including:

[0060] A1. Obtain a preset list of supply transportation routes, where the supply transportation route includes at least one rescue point location and a heavy-load drone base;

[0061] A2. Randomly select a supply transportation route from the list of supply transportation routes and add it to the preset current route list, and add the remaining supply transportation routes in the list of supply transportation routes to the preset remaining route list;

[0062] A3. Calculate the distance similarity between the supply transportation route in the current route list and each supply transportation route in the remaining route list;

[0063] A4. Select the supply transportation route with the smallest distance similarity, add the supply transportation route with the smallest distance similarity to the current route list, and delete the supply transportation route with the smallest distance similarity from the remaining route list;

[0064] A5. Repeat the above operations until the number of rescue point positions in the current path list reaches a preset quantity threshold to obtain path groups of the heavy-load UAV cluster, and add the path groups to a preset group list;

[0065] A6. Determine whether the remaining path list is empty. If it is empty, output each path group in the group list, where one heavy-load UAV in the heavy-load UAV cluster is responsible for one path group in the group list. Otherwise, repeat the above operations.

[0066] In this embodiment, the rescue point position is the target area, the place where materials need to be delivered; the heavy-load UAV base refers to the starting point of material transportation, usually the take-off location of the heavy-load UAV; randomly select a path from the material transportation path list and add it to the current path list, and the remaining paths (unselected paths) will be added to the remaining path list; calculate the distance similarity between each path in the current path list and each path in the remaining path list, and this similarity may be based on factors such as geographical distance, flight time, and path complexity; select the path with the minimum similarity to the paths in the current path list from the remaining path list, which means selecting the path that is most different from the paths in the current path list. Usually, this is to avoid path repetition or optimize resource scheduling; repeat the above steps: each time select the path with the minimum similarity to the paths in the current path list from the remaining path list, add it to the current path list and delete it. This process will continue until the preset number of rescue points is included in the current path list; when the number of paths in the current path list reaches the preset quantity threshold, regard the current path list as a path group and add it to the group list; determine whether the remaining path list is empty. If it is empty, it means that all paths have been allocated and the path grouping is completed; if the remaining path list is empty, output the group list, where each path group can be assigned to a UAV in the heavy-load UAV cluster.

[0067] In an alternative embodiment, the calculation formula for the distance similarity is as follows:

[0068] ;

[0069] Wherein, represents the distance similarity, and represent paths and path , and represent paths and path include the number of rescue points and heavy-load UAV bases, and represent a path and a path of the overloaded UAV base represents a random number uniformly distributed in the range of 0 - 1 represents a rescue point and a rescue point the distance between them

[0070] In an optional embodiment, the environment recognition model adopts an improved YOLO model, and improves the backbone network of the YOLO model according to the feature extraction module and the feature fusion module. Among them, the feature extraction module includes four branches, each branch corresponding to a different receptive field size. Among them, the first branch adopts a residual structure. Before executing the remaining three branches, the number of channels of the input feature map is reduced from 2C to C by using a 1×1 convolution. The second branch extracts small receptive field features through a 3×3 convolution to obtain a first output feature map. The third branch uses a DRB module with K = 7 to extract medium receptive field features to obtain a second output feature map. The fourth branch uses a DRB module with K = 15 to extract large receptive field features to obtain a third output feature map. The first output feature map is added element-wise to the output feature map of the first branch to obtain a fourth output feature map. The second output feature map is added element-wise to the output feature map of the first branch to obtain a fifth output feature map. The third output feature map is added element-wise to the output feature map of the first branch to obtain a sixth output feature map.

[0071] It should be noted that the goal of the feature extraction module is to extract useful features from the input image to assist in environment recognition. To improve the effect of feature extraction, this module uses four branches, each branch targeting a different receptive field size (i.e., the size of the image information area captured by each branch); DRB with K = 7 represents a dynamic convolution module with a convolution kernel size of 7×7; the DRB module with K = 15 has a convolution kernel size of 15×15.

[0072] In an optional embodiment, the feature fusion module downsamples the fourth output feature map using a parallel structure of max pooling and average pooling to obtain a seventh output feature map, upsamples the sixth output feature map using the nearest neighbor interpolation method to obtain an eighth output feature map, divides the seventh output feature map, the fifth output feature map, and the eighth output feature map into three equal sub-feature maps in the channel dimension, applies the Sigmoid function to the three equal channel sub-feature maps of the seventh output feature map, the fifth output feature map, and the eighth output feature map through a feature adaptive selection and fusion mechanism to calculate the selection weights, fuses the three equal channel sub-feature maps of the seventh output feature map, the fifth output feature map, and the eighth output feature map according to the selection weights to obtain a first fused feature, a second fused feature, and a third fused feature, and concatenates the first fused feature, the second fused feature, and the third fused feature in the channel dimension to obtain an output fused feature.

[0073] It should be noted that max pooling is used to extract the local strongest response in the feature map, while average pooling retains the overall information of the feature map. The combination of the two can effectively retain different types of feature information; applying the Sigmoid function to each sub-feature map to calculate the selection weights, the output of the Sigmoid function will be a value between 0 and 1, which is used to represent the importance or weight of each channel sub-feature map.

[0074] In an optional embodiment, the landing control model includes a Markov decision model and a deep reinforcement learning model. Among them, the Markov decision model includes a global state space set, an action space set of the heavy-load UAV cluster, a state transition probability, a discount factor, and a reward value. Among them, the global state space set includes a first state, a second state, and a third state. The action space set includes a height control action and a steering angle control action. The state transition probability represents the state transition probability that the global state jumps to the next state when the heavy-load UAV cluster takes an action. The discount factor represents the influence of subsequent states on the current action decision. The reward value includes a first reward, a second reward, and a third reward. The deep reinforcement learning model is used to solve the Markov decision model to obtain an optimal landing strategy.

[0075] It should be noted that the Markov decision process is a mathematical framework for modeling decision-making problems, especially in the presence of uncertainty; deep reinforcement learning is a technology that combines deep learning and reinforcement learning, which can automatically learn the optimal strategy in a complex environment. In this landing control model, deep reinforcement learning is used to solve the Markov decision process (MDP) and optimize the landing strategy of the UAV by interacting with the environment.

[0076] In an alternative embodiment, the first state represents the relative position information of the heavy-load UAV and the rescue point. The first state includes the distance and relative direction angle between the heavy-load UAV and the rescue point. The second state represents the relative information of the nearby heavy-load UAVs. The second state includes the distance and relative direction angle of the two heavy-load UAVs with the closest distance within the communication range. The third state represents the obstacle distance obtained by lidar. The expression of the steering angle control action is as follows:

[0077] ;

[0078] where, represents the steering angle of the th heavy-load UAV, represents the th heavy-load UAV, represents the absolute value of the maximum steering angle of the heavy-load UAV, represents the number of equal divisions of the steering range of the heavy-load UAV.

[0079] In an alternative embodiment, the expression of the first reward is as follows:

[0080] ;

[0081] where, represents the first reward, and represent the preset weights, and , represents the set of rescue points near the heavy-load UAV, represents the distance between the heavy-load UAV and the target point at time t, represents the first preset reward value, represents the second preset reward value;

[0082] The expression of the second reward is as follows:

[0083] ;

[0084] where, represents the second reward, represents the minimum value of the distance between the heavy-load UAV and the center of the influence range of the surrounding obstacles, represents the influence radius of the obstacle, represents the obstacle avoidance buffer distance, represents the safety distance of the heavy-load UAV, represents the allowable radius of the obstacle, represents the allowable straight-line distance between the heavy-load UAV and the obstacle;

[0085] The expression of the third reward is as follows:

[0086] ;

[0087] Among them, represents the third reward, represents the minimum distance between the heavy-lift drone and other heavy-lift drones, represents the minimum distance that should be maintained between heavy-lift drones.

[0088] In an alternative embodiment, a deep reinforcement learning model is used to solve the Markov decision model to obtain an optimal landing strategy, including:

[0089] B1. Initialize the departure position and target position of the drone;

[0090] B2. Select an action uniformly at random from the global state space set with a preset probability;

[0091] B3. Execute the action while obtaining the next moment state and reward;

[0092] B4. Put the current trajectory into the experience replay buffer, randomly and uniformly collect multiple trajectories, and update the main network parameters according to the loss function for each trajectory, where the trajectory includes the current state, the current action, the next moment state, and the reward;

[0093] B5. Update the target network parameters using the main network parameters every preset number of training steps.

[0094] It should be noted that in reinforcement learning, the main network and the target network are usually used for training. The parameters of the target network are a copy of the main network and are updated regularly to maintain the stability of training. After every preset number of training steps, the parameters of the target network are updated according to the parameters of the main network. This is done to avoid training instability caused by frequent updates.

[0095] Embodiment III, as Figure 2 shown, a material delivery system for heavy-lift drones proposed by the present invention, and a material delivery method for heavy-lift drones applicable thereto, include:

[0096] Path planning unit 1, which is used to obtain rescue requests within the disaster area, determine the rescue point location according to the rescue requests, and determine the material transportation path of the heavy-lift drone cluster according to the rescue point location;

[0097] Environment acquisition unit 2, which is used to control the heavy-lift drone cluster to transport materials to the air above the rescue point location according to the material transportation path, and perform image acquisition on the rescue point location by the heavy-lift drones to obtain the environmental image of the rescue point;

[0098] The environment recognition unit 3 is configured to perform environment recognition on the environment image according to a pre-trained environment recognition model to obtain the environmental state of the rescue point, where the environmental state includes a landable state and a non-landable state;

[0099] The landing control unit 4 is configured to, when the environmental state is the landable state, control the heavy-duty unmanned aerial vehicle to land at the rescue point according to a pre-trained landing control model;

[0100] The material delivery unit 5 is configured to, when the environmental state is the non-landable state, directly control the heavy-duty unmanned aerial vehicle to land at a preset height and deliver the materials to the rescue point.

[0101] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art to which the present invention pertains.

Claims

1. A method for delivering supplies for a heavy-duty unmanned aerial vehicle, characterized in that, Including: Obtain rescue requests within the disaster-stricken area, determine the location of the rescue point according to the rescue requests, and determine the material transportation path of the heavy-load UAV cluster according to the location of the rescue point; Control the heavy-load UAV cluster to transport materials to the air above the rescue point location according to the material transportation path, and perform image acquisition on the rescue point location by the heavy-load UAV to obtain the environmental image of the rescue point; Perform environmental recognition on the environmental image according to a pre-trained environmental recognition model to obtain the environmental state of the rescue point, where the environmental state includes a landable state and a non-landable state; When the environmental state is the landable state, control the heavy-load UAV to land at the rescue point according to the pre-trained landing control model; When the environmental state is the non-landable state, directly control the heavy-load UAV to land at a preset height and drop the materials at the rescue point; Determine the location of the rescue point according to the rescue request, and determine the material transportation path of the heavy-load UAV cluster according to the location of the rescue point, including: Obtain a preset list of material transportation paths, where the material transportation path includes at least one rescue point location and a heavy-load UAV base; Randomly select a material transportation path from the list of material transportation paths and add it to a preset current path list, and add the remaining material transportation paths in the list of material transportation paths to a preset remaining path list; Calculate the distance similarity between the material transportation path in the current path list and each material transportation path in the remaining path list; Select the material transportation path with the smallest distance similarity, add the material transportation path with the smallest distance similarity to the current path list, and delete the material transportation path with the smallest distance similarity from the remaining path list; Repeat the above operations until the number of rescue point locations in the current path list reaches a preset quantity threshold to obtain the path grouping of the heavy-load UAV cluster, and add the path grouping to a preset grouping list; Judge whether the remaining path list is empty. If it is empty, output each path grouping in the grouping list, where one heavy-load UAV in the heavy-load UAV cluster is responsible for one path grouping in the grouping list. Otherwise, repeat the above operations; The calculation formula of the distance similarity is as follows: ； Among them, represents the distance similarity, and represents a path and a path , and represent the number of rescue points and heavy-lift drone bases included in the path and the path . and represent the heavy-lift drone bases of the path and the path . represents a random number uniformly distributed in the range of 0 - 1, represents the rescue point and the rescue point the distance between them.

2. The material delivery method for a heavy-duty unmanned aerial vehicle according to claim 1, wherein The environmental recognition model adopts an improved YOLO model, and improves the backbone network of the YOLO model according to the feature extraction module and the feature fusion module. Among them, the feature extraction module includes four branches, each branch corresponding to a different receptive field size. Among them, the first branch adopts a residual structure, and before executing the remaining three branches, the number of channels of the input feature map is reduced from 2C to C by using a 1×1 convolution. The second branch extracts small receptive field features through a 3×3 convolution to obtain a first output feature map. The third branch uses a DRB module with K = 7 to extract medium receptive field features to obtain a second output feature map. The fourth branch uses a DRB module with K = 15 to extract large receptive field features to obtain a third output feature map. The first output feature map is added element-wise to the output feature map of the first branch to obtain a fourth output feature map. The second output feature map is added element-wise to the output feature map of the first branch to obtain a fifth output feature map. The third output feature map is added element-wise to the output feature map of the first branch to obtain a sixth output feature map.

3. The material delivery method for a heavy-duty unmanned aerial vehicle according to claim 2, characterized in that The feature fusion module uses a parallel structure of max pooling and average pooling to downsample the fourth output feature map to obtain a seventh output feature map, and uses the nearest neighbor interpolation method to upsample the sixth output feature map to obtain an eighth output feature map. The seventh output feature map, the fifth output feature map, and the eighth output feature map are divided into three equal sub-feature maps in the channel dimension. The selection weights are calculated by applying the Sigmoid function to the three equal channel sub-feature maps of the seventh output feature map, the fifth output feature map, and the eighth output feature map respectively through a feature adaptive selection fusion mechanism. The three equal channel sub-feature maps of the seventh output feature map, the fifth output feature map, and the eighth output feature map are fused according to the selection weights to obtain a first fusion feature, a second fusion feature, and a third fusion feature. The first fusion feature, the second fusion feature, and the third fusion feature are concatenated in the channel dimension to obtain an output fusion feature.

4. A method for delivering materials for a heavy-duty unmanned aerial vehicle according to claim 3, characterized in that, The landing control model includes a Markov decision model and a deep reinforcement learning model. Among them, the Markov decision model includes a global state space set, an action space set of the heavy-load UAV cluster, a state transition probability, a discount factor, and a reward value. Among them, the global state space set includes a first state, a second state, and a third state. The action space set includes a height control action and a steering angle control action. The state transition probability represents the state transition probability that the global state jumps to the next state when the heavy-load UAV cluster takes an action. The discount factor represents the influence of subsequent states on the current action decision. The reward value includes a first reward, a second reward, and a third reward. The deep reinforcement learning model is used to solve the Markov decision model to obtain an optimal landing strategy.

5. A method for delivering materials for a heavy-duty unmanned aerial vehicle according to claim 4, characterized in that, Among them, The first state represents the relative position information between the heavy-duty drone and the rescue point. The first state includes the distance and relative direction angle between the heavy-duty drone and the rescue point. The second state represents the relative information of adjacent heavy-duty drones. The second state includes the distance and relative direction angle between the two heavy-duty drones with the shortest distance within the communication range. The third state represents the obstacle distance obtained by the lidar.

6. A method for delivering supplies for a heavy-duty unmanned aerial vehicle according to claim 5, characterized in that, The expression of the first reward is as follows: ； Among them, represents the first reward, and represents the preset weight, and , represents the set of rescue points near the heavy-lift drone, represents the distance between the heavy-lift drone and the target point at time t; The expression of the second reward is as follows: ; Among them, represents the second reward, represents the minimum distance between the heavy-load drone and the center of the influence range of surrounding obstacles, represents the influence radius of the obstacle, represents the obstacle avoidance buffer distance, represents the safety distance of the heavy-load drone, represents the allowable radius of the obstacle, represents the allowable straight-line distance between the heavy-load drone and the obstacle, represents the first preset reward value, represents the second preset reward value; The expression of the third reward is as follows: ; Among them, represents the third reward, represents the minimum distance between the heavy-lift drone and other heavy-lift drones, represents the minimum distance that should be maintained between heavy-lift drones.

7. A method for delivering supplies for a heavy-duty unmanned aerial vehicle according to claim 6, characterized in that, The deep reinforcement learning model is used to solve the Markov decision model to obtain the optimal landing strategy, including: Initialize the departure position and target position of the drone; Select an action uniformly at random from the global state space set with a preset probability; Execute the action, and at the same time obtain the state and reward at the next moment; Put the current trajectory into the experience replay buffer, randomly and uniformly collect multiple trajectories, and update the main network parameters according to the loss function for each trajectory, where the trajectory includes the current state, the current action, the state and reward at the next moment; Update the target network parameters using the main network parameters every preset number of training steps.

8. A material delivery system for a heavy-duty unmanned aerial vehicle, which is applicable to a material delivery method for a heavy-duty unmanned aerial vehicle as claimed in claim 7, characterized in that Including: A path planning unit (1), which is used to obtain rescue requests within the disaster area, determine the location of the rescue point according to the rescue request, and determine the material transportation path of the heavy-duty drone cluster according to the location of the rescue point; An environment acquisition unit (2), which is used to control the heavy-duty drone cluster to transport materials to the air above the rescue point location according to the material transportation path, and perform image acquisition on the rescue point location by the heavy-duty drone to obtain the environmental image of the rescue point; An environment recognition unit (3), which is used to perform environment recognition on the environmental image according to a pre-trained environment recognition model to obtain the environmental state of the rescue point, where the environmental state includes a landable state and a non-landable state; A landing control unit (4), which is used to control the heavy-duty drone to land at the rescue point according to a pre-trained landing control model when the environmental state is the landable state; A material delivery unit (5), which is used to directly control the heavy-duty drone to land at a preset height and deliver the materials to the rescue point when the environmental state is the non-landable state.

Images