Orchard multi-plant protection vehicle path planning and dynamic scheduling method

Abstract

The invention discloses a path planning and dynamic scheduling method for multiple plant protection vehicles in an orchard. The method comprises the following steps: S1, generating an operation base map from aerial photography data of an unmanned aerial vehicle through an operation grid base map generation and operation point drawing module; s2, monitoring insect pests in real time through an insect pest monitoring module, and transmitting monitored insect pest point coordinates to a dispatching center; s3, calling an operation grid base map generation and operation point drawing module to draw operation points; s4, a path planning module performs path planning prediction on the N plant protection vehicles; s5, the N plant protection vehicles receive the operation path information and start to execute operation; s6, in the operation execution process, an obstacle avoidance scheme is provided through a vehicle-mounted local path planning module; and S7, new tasks are allocated to the plant protection vehicles completing the operation through the dynamic scheduling module, and when fault information occurs, the uncompleted tasks are allocated to the nearby plant protection vehicles until all the tasks are completed. According to the invention, path planning and dynamic scheduling of multiple plant protection vehicles in an orchard can be accurately and efficiently completed.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent agriculture and relates to a path planning and dynamic scheduling method for multiple plant protection vehicles in an orchard. Background Art

[0002] In the field of orchard plant protection operations, traditional plant protection methods are mainly relied on. In recent years, with the development of science and technology, plant protection vehicle technology has also begun to be applied to large-scale orchard operations in hills. However, in actual applications, both traditional plant protection methods and existing plant protection vehicle technology have many problems, which are specifically manifested as follows:

[0003] (1) Traditional plant protection methods rely on manual backpack sprayers or simple mechanical spray equipment. This method has serious limitations in hilly orchards. Hilly areas have large undulating terrain and complex and changeable topography. When operated manually, workers need to walk arduously between steep slopes, depressions and other different terrains. The labor intensity is extremely high and the physical exertion is extremely great, resulting in extremely low work efficiency. At the same time, it is difficult for humans to accurately determine the location and dosage of pesticides, and it is difficult to ensure uniform and accurate pesticide application effects. It is very easy to miss spraying areas, allowing pests and diseases to breed and spread in these areas, affecting the growth of fruit trees and fruit yields; and heavy spraying areas not only causes a waste of pesticides and increases costs, but may also cause environmental problems such as soil compaction and water pollution due to excessive use of pesticides, destroying the ecological balance and posing a threat to the sustainable development of orchards.

[0004] (2) Existing plant protection vehicle technology

[0005] (a) Lack of scientificity in path planning: Most plant protection vehicles lack an effective multi-starting point balanced traversal path planning method. When faced with the complex terrain, diverse distribution of fruit trees, and changing pest and disease conditions in hilly orchards, they are unable to intelligently and comprehensively consider the complex terrain, distribution of fruit trees, and pest and disease conditions in hilly orchards for intelligent path planning, resulting in unreasonable operation paths. Plant protection vehicles may drive in a roundabout way in the orchard, increasing unnecessary mileage, resulting in increased energy consumption, prolonged operation time, and significantly reduced operation efficiency. For example, in hilly areas, vehicles may frequently go up and downhill, consuming a lot of energy and easily wearing out vehicle parts. In addition, some plant protection vehicles do not fully consider the slope factor in hilly terrain when planning paths. When driving in steep slopes, vehicles need to overcome a large gravitational potential energy, which greatly increases energy consumption and is prone to vehicle failures such as engine overheating and brake failure, seriously affecting the continuity of operations. Once a vehicle fails, operations will stagnate during maintenance, missing the best time for pest and disease control, giving pests and diseases a chance to spread and harm the orchard.

[0006] (b) Poor coordination among plant protection vehicles: Existing plant protection vehicle technology often ignores the dynamic coordination among multiple plant protection vehicles. In large-scale hilly orchards, the operation of a single plant protection vehicle is difficult to meet the needs of fast and efficient plant protection, and the coordinated operation of multiple plant protection vehicles can significantly improve the operation efficiency. However, the existing plant protection vehicle technology lacks an effective orchard communication system and dynamic scheduling mechanism, and the plant protection vehicles cannot share location information and operation status in real time. This often results in some plant protection vehicles completing their tasks ahead of schedule and becoming idle, while some areas are seriously plagued by pests and diseases, causing the plant protection vehicles to be overburdened. When a plant protection vehicle fails or the operation progress is hindered, the operation plan cannot be adjusted in time, the overall operation efficiency is low, and it is difficult to give full play to the advantages of the coordinated operation of multiple plant protection vehicles and to achieve efficient plant protection in the orchard.

[0007] (c) Weak obstacle identification and local path planning capabilities: In terms of obstacle identification and local path planning, the sensor systems of many plant protection vehicles are imperfect and cannot accurately identify various obstacles, such as people, stones, gullies, etc. Even if some plant protection vehicles are equipped with some sensors, they lack effective data fusion and analysis capabilities, resulting in low accuracy and reliability of obstacle monitoring. When encountering obstacles, it is impossible to quickly plan a reasonable local detour path, and the plant protection vehicle often has to stop working and wait for manual intervention, which can easily cause operation interruptions, seriously affect the progress of operations, miss the best time for pest and disease control, increase the risk of pest and disease spread, and thus have an adverse impact on the yield and quality of the orchard. Summary of the invention

[0008] In view of the problems existing in the prior art, the object of the present invention is to provide a path planning and dynamic scheduling method for multiple plant protection vehicles in an orchard.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard, comprising the following steps:

[0011] S1, generate an operation base map from drone aerial photography data through the operation grid base map generation and operation point drawing module;

[0012] S2. Real-time monitoring of pests is performed through the pest monitoring module, and the coordinates of the monitored pest points are transmitted to the dispatching center;

[0013] S3. The dispatch center calls the operation grid base map generation and operation point drawing module to draw the operation points according to the operation point information, and then allocates n operation areas to the n existing plant protection vehicles in the warehouse according to the drawn operation point map;

[0014] S4, the path planning module makes path planning predictions for N plant protection vehicles, returns the obtained path planning prediction result set to the dispatch center, and distributes the planned operation paths to the existing N plant protection vehicles;

[0015] S5, N plant protection vehicles receive the operation path information from the dispatch center and start from the warehouse to perform the operation;

[0016] S6. During the operation of the plant protection vehicle, the location information of each plant protection vehicle is tracked in real time and transmitted back to the dispatch center. When encountering obstacles, the plant protection vehicle provides an obstacle avoidance solution through the on-board local path planning module to enable the plant protection vehicle to avoid obstacles;

[0017] S7. When a plant protection vehicle completes the task in its operating area, it sends the completion information to the dispatch center. The dispatch center calls the dynamic dispatch module to assign new tasks to the plant protection vehicles that have completed the operation. When a plant protection vehicle has fault information, the dispatch center immediately notifies other plant protection vehicles that will be idle soon, and assigns its unfinished tasks to nearby plant protection vehicles based on the location and task status of the faulty plant protection vehicle until all tasks are completed.

[0018] Preferably, in step S1, generating an operation base map from drone aerial photography data by using an operation grid base map generation and operation point drawing module comprises the following steps:

[0019] S11. Use the monitoring equipment carried by the drone to scan and photograph the orchard to obtain drone aerial photography data;

[0020] S12. Import the acquired drone aerial photography data into GIS software to generate a digital elevation model (DEM) and orthophoto map of the orchard as the operating base map.

[0021] Preferably, in step S2, real-time monitoring of pests by the pest monitoring module and sending the monitored pest point coordinates back to the dispatching center include the following steps:

[0022] Insect traps and sensors are placed in the orchard to monitor the pest situation in real time. When a pest point (or tree) is detected, the coordinates of the pest point are sent back to the dispatch center.

[0023] Preferably, in step S3, calling the operation grid base map generation and operation point drawing module to draw the operation points includes the following steps:

[0024] S31, using GIS software to mark the points to be operated on the digital elevation model (DEM) and orthophoto map of the orchard, wherein the points to be operated are the positions of the traps or sensors that have captured the insects;

[0025] S32, creating a digital elevation model (DEM), an orthophoto map and a work point as different layers, each layer having independent properties and display settings; wherein the DEM layer is set to a three-dimensional terrain display mode to highlight the terrain and geomorphic features; the orthophoto layer is set to an image display mode to clearly display image information such as fruit tree distribution; the work point layer is set to an editable and annotated point element layer to facilitate the drawing, modification and annotation of the work point;

[0026] S33. After the setting is completed, a three-dimensional operation diagram is generated.

[0027] Preferably, in step S4, the path planning module performs path planning for N plant protection vehicles, including the following steps:

[0028] S41. Analyze the density and spatial layout of the operating points in the operating area, and take the minimum total distance of multiple plant protection vehicles and the minimum distance difference as the optimization goal. Suppose the number of plant protection vehicles is N, and the path length of each plant protection vehicle is L. i (i=1,2,…,N), the total distance is The objective function is: minimize(L total ) and minimize(σ(L i )), where σ(L i ) represents the standard deviation of the path length of each plant protection vehicle;

[0029] S42, using the long short-term memory network (LSTM) in deep reinforcement learning for path planning; specifically: construct a state space based on the collected state information, define the state space of the plant protection vehicle as the location information of the current operation point, the hilly terrain characteristics (slope, slope direction data), define the moving direction of the plant protection vehicle as the action space, and set the reward function to the objective function and energy consumption in step S41. The goal is to minimize the total distance of multiple plant protection vehicles and the difference in distance, and minimize the total energy consumption of the plant protection vehicle operation. At the same time, task completion is used as a positive reward, and encountering obstacles is used as a negative reward, both of which are linearly added to the reward function with fixed values;

[0030] S43, using the three-dimensional operation diagram stored in the dispatching center to perform model training;

[0031] S44, using the trained model to generate path predictions for n plant protection vehicles;

[0032] S45. Return the predicted path result set of the plant protection vehicle to the dispatch center.

[0033] Preferably, in step S43, the model training using the three-dimensional operation diagram stored in the dispatching center includes the following steps:

[0034] S431. The plant protection vehicle starts to interact in the environment simulated by the three-dimensional operation map according to the initial random strategy, and designs a value network: composed of a target network and a prediction network, wherein the network structures of both networks adopt an LSTM network structure; the target network is a network with the same structure as the prediction network but with a lower update frequency, which is used to stabilize the training process;

[0035] S432, for each time step t, the epsilon strategy is used for action selection, the action is randomly selected with a probability of epsilon, the target network is used with a probability of 1-epsilon, and the action with the largest Q value in the state is selected. Epsilon is initialized to 1 and gradually decreases over time;

[0036] S433, execute the selected action a t Afterwards, you will get reward r t , observe the next state s t+1 , the experience tuple (s t , a t , r t ,s t+1 ) Playback buffer, set the buffer capacity to store 1000 experience tuples, when the buffer is not full, store new experience tuples in it; when the buffer is full, the new experience tuple will replace the earliest stored tuple;

[0037] S434, randomly extract a batch of experience tuples from the experience playback buffer, assuming that the batch size is B=100, for each experience tuple (s t , a t , r t ,s t+1 ), calculate the target Q value according to the target network, the calculation formula is:

[0038] y t =r t +γmax a′ Q(s t+1 , a′; θ - )

[0039] Among them, y j is the target Q value, r j is the reward, γ is the discount factor, set γ = 0.99, Q(s t+1 , a′; θ - ) is the target network (with parameters θ - ) for the next state s t+1 Estimate the Q value of taking action a′;

[0040] S435, the extracted state sequence (s tEach state in the prediction network is input into the prediction network for forward propagation, and the Q value estimate Q(s) for each action is output. t , a t ; θ), where θ is the parameter of the prediction network;

[0041] S436. Calculate the loss function based on the target Q value and the predicted Q value, and use the mean square error (MSE) as the loss function:

[0042] S437, based on the calculated loss function, calculate the gradient of the loss function to the prediction network parameter θ through the back propagation algorithm, use the Adam optimizer to update the network parameters according to the gradient, and the Adam optimizer will adjust the learning rate according to the first-order moment estimate and the second-order moment estimate of the gradient, and update the parameters so that the loss function gradually decreases;

[0043] S438, after each training of a certain number of time steps, a portion of the parameters of the prediction network is copied to the target network to update the parameters of the target network, so as to make the calculation of the target Q value more stable and avoid excessive fluctuations in the training process;

[0044] S439, repeat steps S432-S438 until a predetermined number of training rounds is reached or the network performance on the validation set no longer improves.

[0045] Preferably, in step S6, the on-board local path planning module on the orchard plant protection vehicle provides an obstacle avoidance solution, comprising the following steps:

[0046] S61. Use the sensor system composed of laser radar, camera, ultrasonic sensor, etc. equipped on the orchard plant protection vehicle to collect environmental information around the vehicle in real time. Process and analyze the data collected by the sensor to determine the position and range of obstacles in the orchard coordinate system;

[0047] S62, using the RRT* algorithm to plan the local path, input the obstacle location, shape, range information and the slope, slope direction, and terrain undulation information of the hill from the dispatch center, and output the obstacle avoidance path;

[0048] S63. The plant protection vehicle avoids obstacles according to the newly planned path.

[0049] Preferably, in step S62, the use of the RRT* algorithm to plan the local path, inputting the obstacle position, shape, range information and the slope, aspect, and terrain undulation information of the hills from the dispatch center, and outputting the obstacle avoidance path includes the following steps:

[0050] S621, set the current position as the starting point start and the reachable obstacle avoidance point next to the obstacle as the target point goal, and create an RRT tree T containing only start, and define the reachable obstacle avoidance point as the point that is reachable and can effectively avoid obstacles when the plant protection vehicle reaches the point, and define the cost function as the energy consumption of the plant protection vehicle, where the energy consumption represents a function of the hill slope, slope aspect, terrain, and path length, and the environment space is all reachable grid points except obstacles;

[0051] S622, random sampling: randomly sample a point x_rand in the environment space;

[0052] S623, find the nearest node x_near from the tree T, take x_near as the starting point, extend a certain distance in the direction, and obtain a new node x_new;

[0053] S624, detecting whether the path from x_near to x_new collides with an obstacle based on the obstacle position, shape, and range information; if a collision occurs, returning to step S622 to continue the next iteration;

[0054] S625. Find the node x_min closest to x_new in the tree T, and calculate the cost cost (x_min, x_new) from x_min to x_new.

[0055] S4626. For the node x_near_neighbors in the tree T whose distance to x_new is within a certain range, calculate the cost cost(x_near,x_new) of connecting to x_new through x_near, and select the connection method with the minimum cost;

[0056] S627, add x_new to the tree T, and update the connection relationship and cost information between nodes;

[0057] S628, optimizing some nodes in the tree T with the target point goal as the goal, improving the quality of the path by adjusting the connection relationship and cost information, and repeating steps S622-S628 until a path is found or the maximum number of iterations is reached;

[0058] S629: If the newly added node x_new is close to the target point goal, a feasible path is found and the algorithm ends.

[0059] Beneficial effects:

[0060] Compared with the prior art, the orchard plant protection vehicle path planning and dynamic scheduling method of the present invention has the following outstanding advantages:

[0061] (1) Accurate and efficient path planning: The present invention adopts a method that combines multi-starting point balanced traversal path planning with deep reinforcement learning. In the path planning process, the terrain characteristics of the hilly orchard are fully considered, and a more reasonable driving route can be selected by obtaining terrain data (such as slope, slope direction, etc.).

[0062] (2) Intelligent collaborative operation scheduling: By establishing an effective communication system and dynamic scheduling module for the orchard, real-time communication and collaborative operation between multiple plant protection vehicles are achieved. Plant protection vehicles can share location information and operation status in real time, and the scheduling center can reasonably allocate tasks according to actual conditions. When a plant protection vehicle completes a task or fails, the task allocation and operation plan can be adjusted quickly to ensure the continuity and efficiency of the operation, give full play to the advantages of collaborative operation of multiple plant protection vehicles, improve the efficiency of plant protection operations in the entire orchard, and achieve optimal allocation of resources.

[0063] (3) Powerful obstacle response capability: Equipped with advanced sensor systems and complete obstacle identification and local path planning modules, the vehicle integrates multiple sensors (such as lidar, cameras, ultrasonic sensors, etc.) to improve the accuracy and reliability of obstacle monitoring. The deep learning algorithm is used to analyze sensor data and continuously optimize the obstacle identification model, which can quickly and accurately identify various obstacles. Once an obstacle is found, the optimal detour path can be quickly planned, which effectively enhances the adaptability of the plant protection vehicle in complex hilly orchard environments, reduces the risk of operation interruption, ensures the smooth progress of operations, and improves the reliability of orchard plant protection operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the architecture of a path planning and dynamic scheduling system for multiple plant protection vehicles in an orchard;

[0065] Figure 2 It is a flow chart of a method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard;

[0066] Figure 3 A schematic diagram of the process of performing path planning for the path planning module;

[0067] Figure 4 A schematic diagram of the execution process of the dynamic scheduling module. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Reference Figure 1 The present invention provides a path planning and dynamic scheduling system for an orchard plant protection vehicle, including: an insect pest monitoring module, an operation grid base map generation and operation point drawing module, a path planning module, an orchard plant protection vehicle, a dynamic scheduling module, and a scheduling center; wherein,

[0070] The operation grid base map generation and operation point drawing module is used to generate an operation base map from drone aerial photography data, and import the generated operation base map into GIS software to generate a digital elevation model (DEM) and orthophoto map of the orchard.

[0071] The pest monitoring module is used to monitor pests in real time through monitoring equipment deployed in the orchard, and send the specific location (point coordinates) of the monitored pests back to the dispatching center.

[0072] The path planning module is used to make path planning predictions for n plant protection vehicles and return the path prediction result set of n plant protection vehicles to the dispatching center.

[0073] The dynamic scheduling module is used to: (1) receive information and operation information layout diagrams from the scheduling center that some plant protection vehicles have completed their assigned planned tasks, and according to the distribution of unfinished operation points, select operation points that are closer to the completed plant protection vehicles from the unfinished paths in reverse order for the idle plant protection vehicles and assign them to the plant protection vehicles. Update the operation information layout diagram, mark the newly assigned operation points on the diagram, and return the vehicle task information and updated operation information layout diagram information to the scheduling center; (2) receive information from the scheduling center that a plant protection vehicle has encountered a fault in a certain area. The scheduling center immediately notifies other plant protection vehicles that will be idle soon, and assigns its unfinished tasks to nearby plant protection vehicles based on the location and task status of the faulty plant protection vehicle, and finally returns the scheduling information to the scheduling center.

[0074] A plant protection vehicle is provided with a local path planning module, which is used to collect environmental information around the plant protection vehicle during its driving process, and plan a local path based on the collected information, so that the plant protection vehicle can avoid obstacles according to the newly planned path, i.e., the obstacle avoidance path.

[0075] Example 1

[0076] See also Figure 2-4 A method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard comprises the following steps:

[0077] S1. Generate an operation base map from the drone aerial photography data through the operation grid base map generation and operation point drawing module; the specific execution process includes the following steps:

[0078] S11. Use the laser radar and camera on the drone to scan and photograph the orchard and obtain the drone aerial data;

[0079] S12, importing the acquired drone aerial data into GIS software to generate a digital elevation model (DEM) and orthophoto map of the orchard;

[0080] S2. Real-time monitoring of pests is performed through the pest monitoring module, and the coordinates of the monitored pest points are transmitted to the dispatching center. The specific implementation process includes the following steps: real-time monitoring of pest conditions is performed through insect traps and sensors arranged in the orchard, and when a point (or tree body) with pests is detected, the coordinates of the pest point (location information) are sent back to the dispatching center.

[0081] S3. The dispatch center calls the operation grid base map generation and operation point drawing module to draw the operation points according to the operation point information, and then allocates n operation areas to the n existing plant protection vehicles in the warehouse according to the drawn operation point map; the specific execution process includes the following steps:

[0082] S31, using GIS software to mark the points to be operated on the digital elevation model (DEM) and orthophoto map of the orchard, wherein the points to be operated are the positions of the traps or sensors that have captured the insects;

[0083] S32. Create the digital elevation model (DEM), orthophoto map and work points as different layers respectively, each layer has independent properties and display settings, and generate a three-dimensional work map after the settings are completed; wherein, the DEM layer is set to the three-dimensional terrain display mode to highlight the terrain and geomorphic features; the orthophoto layer is set to the image display mode to clearly display the image information such as the distribution of fruit trees; the work point layer is set to an editable and annotated point feature layer to facilitate the drawing, modification and annotation of the work points.

[0084] S33. After the setting is completed, a three-dimensional operation diagram is generated, and the generated three-dimensional operation diagram is saved to the dispatching center.

[0085] S4, the path planning module makes path planning predictions for N plant protection vehicles, returns the obtained path planning prediction result set to the dispatch center, and then distributes the planned operation paths to the existing N plant protection vehicles; the specific execution process includes the following steps:

[0086] S41. Analyze the density and spatial layout of the operating points in the operating area, and take the minimum total distance of multiple plant protection vehicles and the minimum distance difference as the optimization goal. Suppose the number of plant protection vehicles is N, and the path length of each plant protection vehicle is L. i (i=1,2,…,N), the total distance is The objective function is: minimize(L total ) and minimize(σ(L i )), where σ(Li ) represents the standard deviation of the path length of each plant protection vehicle;

[0087] S42, using the long short-term memory network (LSTM) in deep reinforcement learning for path planning;

[0088] Specifically, a state space is constructed according to the collected state information, the state space of the plant protection vehicle is defined as the location information of the current operation point, the hilly terrain characteristics (slope, slope direction data), the moving direction of the plant protection vehicle is defined as the action space, and the reward function is set to the objective function and energy consumption in step S300. The goal is to minimize the total distance of multiple plant protection vehicles and the distance difference, and minimize the total energy consumption of the plant protection vehicle operation. At the same time, task completion is used as a positive reward, and encountering obstacles is used as a negative reward, both of which are linearly added to the reward function with fixed values;

[0089] S43, using the three-dimensional operation diagram stored in the dispatching center to perform model training; the specific execution process includes the following steps:

[0090] S431. The plant protection vehicle starts to interact in the environment simulated by the three-dimensional operation map according to the initial random strategy, and designs the value network: it consists of a target network and a prediction network, where the network structures of both use the LSTM network structure. The target network is a network with the same structure as the prediction network but with a lower update frequency, which is used to stabilize the training process;

[0091] S432, for each time step t, the epsilon strategy is used for action selection, the action is randomly selected with a probability of epsilon, the target network is used with a probability of 1-epsilon, and the action with the largest Q value in the state is selected. Epsilon is initialized to 1 and gradually decreases over time;

[0092] S433, execute the selected action a t Afterwards, you will get reward r t , observe the next state s t+1 , the experience tuple (s t , a t , r t ,s t+1 ) Playback buffer, set the buffer capacity to store 1000 experience tuples, when the buffer is not full, store new experience tuples in it; when the buffer is full, the new experience tuple will replace the earliest stored tuple;

[0093] S434, randomly extract a batch of experience tuples from the experience playback buffer, assuming that the batch size is B=100, for each experience tuple (s t , a t , r t ,st+1 ), calculate the target Q value according to the target network, the calculation formula is:

[0094] y t =r t +γmax a′ Q(s t+1 , a′; θ - )

[0095] Among them, y j is the target Q value, r j is the reward, γ is the discount factor, set γ = 0.99, Q(s t+1 , a′; θ - ) is the target network (with parameters θ - ) for the next state s t+1 Estimate the Q value of taking action a′;

[0096] S435, the extracted state sequence (s t Each state in the prediction network is input into the prediction network for forward propagation, and the Q value estimate Q(s) for each action is output. t , a t ; θ), where θ is the parameter of the prediction network;

[0097] S436. Calculate the loss function based on the target Q value and the predicted Q value, and use the mean square error (MSE) as the loss function:

[0098] S437, based on the calculated loss function, calculate the gradient of the loss function to the prediction network parameter θ through the back propagation algorithm, use the Adam optimizer to update the network parameters according to the gradient, and the Adam optimizer will adjust the learning rate according to the first-order moment estimate and the second-order moment estimate of the gradient, and update the parameters so that the loss function gradually decreases;

[0099] S438, after each training of a certain number of time steps, a portion of the parameters of the prediction network is copied to the target network to update the parameters of the target network, so as to make the calculation of the target Q value more stable and avoid excessive fluctuations in the training process;

[0100] S439, repeat steps S432-S438 until a predetermined number of training rounds is reached or the network performance on the validation set no longer improves.

[0101] S44, using the trained model to generate path predictions for n plant protection vehicles;

[0102] S45. Return the predicted path set of the plant protection vehicle to the dispatch center.

[0103] The present invention adopts an intelligent path planning technology that integrates multiple factors, and integrates multi-starting point balanced traversal and deep reinforcement learning: a multi-starting point balanced traversal path planning method is adopted, and combined with the long short-term memory network (LSTM) in deep reinforcement learning, multiple factors in the operation area are fully considered for path planning. When constructing the state space, the operation point of the orchard is taken as the basis, and the current operation point location and hilly terrain characteristics (such as slope and slope direction data) are integrated to define the movement direction action space of the plant protection vehicle. In the optimization problem, the minimum total distance of multiple plant protection vehicles and the minimum distance difference are taken as the optimization goals, and the energy consumption of the plant protection vehicle operation is taken into consideration, and the reward function is set to include the above-mentioned objective function and energy consumption factors.

[0104] S5, N plant protection vehicles receive the operation path information from the dispatch center and start from the warehouse to perform the operation;

[0105] S6. When the plant protection vehicle encounters an obstacle during the operation, an obstacle avoidance scheme is provided by the on-board local path planning module of the plant protection vehicle, so that the plant protection vehicle can avoid the obstacle. The specific execution process includes the following steps:

[0106] S61. Use the sensing equipment such as laser radar, camera, ultrasonic sensor, etc. equipped on the orchard plant protection vehicle to collect the environmental information around the vehicle in real time, and process and analyze the environmental information data collected by the sensing equipment to determine the position and range of obstacles in the orchard coordinate system.

[0107] S62, using the RRT* algorithm to plan the local path, input the obstacle location, shape, range information and the slope, slope direction, and terrain undulation information of the hill from the dispatch center, and output a newly planned path, i.e., the obstacle avoidance path; wherein the specific execution process of using the RRT* algorithm to plan the local path includes the following steps:

[0108] S621, set the current position as the starting point start and the reachable obstacle avoidance point next to the obstacle as the target point goal, and create an RRT tree T containing only start, and define the reachable obstacle avoidance point as the point that is reachable and can effectively avoid obstacles when the plant protection vehicle reaches the point, and define the cost function as the energy consumption of the plant protection vehicle, where the energy consumption represents a function of the hill slope, slope aspect, terrain, and path length, and the environment space is all reachable grid points except obstacles;

[0109] S622, random sampling: randomly sample a point x_rand in the environment space;

[0110] S623, find the nearest node x_near from the tree T, take x_near as the starting point, extend a certain distance in the direction, and obtain a new node x_new;

[0111] S624, detecting whether the path from x_near to x_new collides with an obstacle based on the obstacle position, shape, and range information; if a collision occurs, returning to step S622 to continue the next iteration;

[0112] S625. Find the node x_min closest to x_new in the tree T, and calculate the cost cost (x_min, x_new) from x_min to x_new.

[0113] S626. For the node x_near_neighbors in the tree T whose distance to x_new is within a certain range, calculate the cost cost(x_near,x_new) of connecting to x_new through x_near, and select the connection method with the minimum cost;

[0114] S627, add x_new to the tree T, and update the connection relationship and cost information between nodes;

[0115] S628, optimizing some nodes in the tree T with the target point goal as the goal, and improving the quality of the path by adjusting the connection relationship and cost information;

[0116] S629: If the newly added node x_new is close to the target point goal, a feasible path is found and the algorithm ends.

[0117] S63. The plant protection vehicle avoids obstacles according to the obstacle avoidance path.

[0118] The present invention performs local path planning for obstacle avoidance: by fusing and processing data from multiple sensors, various types of obstacles are identified, including people, stones, gullies, trees and other different types of obstacles, and deep learning algorithms are used to conduct in-depth analysis of sensor data, continuously optimize the obstacle recognition model, and improve recognition speed and accuracy. When an obstacle is detected, the RRT algorithm is used to quickly plan the optimal detour path in combination with the slope and direction of the hill.

[0119] S7. When a plant protection vehicle completes the task in its own operation area, it sends the completion information to the dispatch center. The dispatch center calls the dynamic dispatch module to assign new tasks to the plant protection vehicles that have completed the operation. When a plant protection vehicle has a fault, the dispatch center immediately notifies other plant protection vehicles that will be idle soon, and assigns its unfinished tasks to nearby plant protection vehicles according to the location and task status of the faulty plant protection vehicle until all tasks are completed. The dynamic dispatch collaborative operation mechanism, the specific execution process, includes the following two aspects:

[0120] (1) Receive information and operation information layout diagrams from the dispatch center on the completion of the assigned planning tasks of certain plant protection vehicles, and select operation points that are closer to the completed plant protection vehicles from the unfinished paths in reverse order according to the distribution of unfinished operation points, and assign them to the plant protection vehicles; update the operation information layout diagram, mark the newly assigned operation points on the diagram, and return the vehicle task information and the updated operation information layout diagram information to the dispatch center;

[0121] (2) Upon receiving the fault information about a plant protection vehicle in a certain area from the dispatch center, the dispatch center will immediately notify other plant protection vehicles that will be idle soon, and assign its unfinished tasks to nearby idle vehicles based on the location and task status of the faulty plant protection vehicle. The work area division between the operating bodies will be dynamically adjusted based on the actual situation, and the dispatch information will be returned to the dispatch center.

[0122] 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 technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.

Claims

1. A method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard, characterized in that: The following steps are involved: S1, generate an operation base map from drone aerial photography data through the operation grid base map generation and operation point drawing module; S2. Real-time monitoring of pests is performed through the pest monitoring module, and the coordinates of the monitored pest points are transmitted to the dispatching center; S3. The dispatch center calls the operation grid base map generation and operation point drawing module to draw the operation points according to the operation point information, and then allocates n operation areas to the n existing plant protection vehicles in the warehouse according to the drawn operation point map; S4, the path planning module makes path planning predictions for N plant protection vehicles, returns the obtained path planning prediction result set to the dispatch center, and then distributes the planned operation paths to the existing N plant protection vehicles; S5, N plant protection vehicles receive the operation path information from the dispatch center and start from the warehouse to perform the operation; S6. When the plant protection vehicle encounters an obstacle during operation, an obstacle avoidance plan is provided through the on-board local path planning module of the plant protection vehicle, so that the plant protection vehicle can avoid the obstacle; S7. When a plant protection vehicle completes the task in its operating area, it sends the completion information to the dispatch center. The dispatch center calls the dynamic dispatch module to assign new tasks to the plant protection vehicles that have completed the operation. When a plant protection vehicle has fault information, the dispatch center immediately notifies other plant protection vehicles that will be idle soon, and assigns its unfinished tasks to nearby plant protection vehicles based on the location and task status of the faulty plant protection vehicle until all tasks are completed.

2. The method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard according to claim 1, characterized in that: In step S1, the operation grid base map generation and operation point drawing module generates an operation base map from drone aerial photography data, including the following steps: S11. Use the monitoring equipment carried by the drone to scan and photograph the orchard to obtain drone aerial photography data; S12. Import the acquired drone aerial photography data into GIS software to generate a digital elevation model and orthophoto map of the orchard as the operating base map.

3. A method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard according to claim 2, characterized in that: In step S2, the pest monitoring module is used to monitor the pests in real time, and the coordinates of the monitored pest points are sent back to the dispatch center, including the following steps: the pest situation is monitored in real time through insect traps and sensors arranged in the orchard, and when a pest point is monitored, the coordinates of the pest point are sent back to the dispatch center.

4. The method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard according to claim 1, characterized in that: In step S3, calling the operation grid base map generation and operation point drawing module to draw the operation points includes the following steps: S31, using GIS software to mark the points to be operated on the digital elevation model and orthophoto map of the orchard, wherein the points to be operated are the positions of the traps or sensors that have captured the insects; S32, creating the digital elevation model, orthophoto map and operation point as different layers, each layer having independent attributes and display settings; wherein the digital elevation model layer is set to a three-dimensional terrain display mode to highlight the terrain and geomorphic features; the orthophoto layer is set to an image display mode to clearly display image information such as fruit tree distribution; the operation point layer is set to an editable and annotated point element layer to facilitate the drawing, modification and annotation of the operation point; S33. After the setting is completed, a three-dimensional operation diagram is generated, and the generated three-dimensional operation diagram is saved to the dispatching center.

5. The method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard according to claim 1, characterized in that: In step S4, the path planning module performs path planning for N plant protection vehicles, including the following steps: S41. Analyze the density and spatial layout of the operating points in the operating area, and take the minimum total distance of multiple plant protection vehicles and the minimum distance difference as the optimization goal. Suppose the number of plant protection vehicles is N, and the path length of each plant protection vehicle is L. i (i=1,2,…,N), the total distance is The objective function is: minimize(L total ) and minimize(σ(L i )), where σ(L i ) represents the standard deviation of the path length of each plant protection vehicle; S42, using the long short-term memory network in deep reinforcement learning for path planning; specifically: constructing a state space based on the collected state information, defining the state space of the plant protection vehicle as the location information of the current operation point and the hilly terrain characteristics, defining the moving direction of the plant protection vehicle as the action space, and setting the reward function to the objective function and energy consumption in step S41, the goal is to minimize the total distance of multiple plant protection vehicles and the difference in distance, and minimize the total energy consumption of the plant protection vehicle operation, and at the same time, task completion is used as a positive reward, and encountering obstacles is used as a negative reward, both of which are linearly added to the reward function with fixed values; S43, using the three-dimensional operation diagram stored in the dispatching center to perform model training; S44, using the trained model to generate path predictions for n plant protection vehicles; S45. Return the predicted path result set of the plant protection vehicle to the dispatch center.

6. A method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard according to claim 5, characterized in that: In step S43, the model training is performed using the three-dimensional operation diagram stored in the dispatching center, including the following steps: S431, the plant protection vehicle starts to interact in the environment simulated by the three-dimensional operation map according to the initial random strategy, and designs the value network: it consists of a target network and a prediction network, and the network structures of both adopt the LSTM network structure; S432, for each time step t, the epsilon strategy is used for action selection, the action is randomly selected with a probability of epsilon, the target network is used with a probability of 1-epsilon, and the action with the largest Q value in the state is selected. Epsilon is initialized to 1 and gradually decreases over time; S433, execute the selected action a t Afterwards, you will get the reward r t , observe the next state s t+1 , the experience tuple (s t , a t , r t ,s t+1 ) Playback buffer, set the buffer capacity to store 1000 experience tuples, when the buffer is not full, store new experience tuples in it; when the buffer is full, the new experience tuple will replace the earliest stored tuple; S434, randomly extract a batch of experience tuples from the experience playback buffer, assuming that the batch size is B=100, for each experience tuple (s t , a t , r t ,s t+1 ), calculate the target Q value according to the target network, the calculation formula is: y t =r t +γmax a′ Q(s t+1 ,a′;θ - ) Among them, y j is the target Q value, r j is the reward, γ is the discount factor, set γ = 0.99, Q(s t+1 , a′; θ - ) is the target network's response to the next state s t+1 Estimate the Q value of taking action a′; S435, input the extracted state sequence into the prediction network for forward propagation, and output the Q value estimate Q(s) for each action t , a t ; θ), where θ is the parameter of the prediction network; S436. Calculate the loss function based on the target Q value and the predicted Q value, and use the mean square error as the loss function: S437, based on the calculated loss function, calculate the gradient of the loss function to the prediction network parameter θ through the back propagation algorithm, use the Adam optimizer to update the network parameters according to the gradient, and the Adam optimizer will adjust the learning rate according to the first-order moment estimate and the second-order moment estimate of the gradient, and update the parameters so that the loss function gradually decreases; S438, after each training of a certain number of time steps, a portion of the parameters of the prediction network is copied to the target network to update the parameters of the target network, so as to make the calculation of the target Q value more stable and avoid excessive fluctuations in the training process; S439, repeat steps S432-S438 until a predetermined number of training rounds is reached or the network performance on the validation set no longer improves.

7. The method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard according to claim 1, characterized in that: In step S6, when the plant protection vehicle encounters an obstacle during the operation, an obstacle avoidance scheme is provided by the on-board local path planning module of the plant protection vehicle, so that the plant protection vehicle avoids the obstacle, including the following steps: S61. Using the sensor equipment equipped on the orchard plant protection vehicle, collect environmental information around the vehicle in real time, and process and analyze the environmental information data collected by the sensor equipment to determine the position and range of the obstacle in the orchard coordinate system; S62, using the RRT* algorithm to plan the local path, input the obstacle location, shape, range information and the slope, slope direction, and terrain undulation information of the hill from the dispatch center, and output the obstacle avoidance path; S63. The plant protection vehicle avoids obstacles according to the obstacle avoidance path.

8. The method for path planning and dynamic scheduling of multiple plant protection vehicles in an orchard according to claim 7, characterized in that: In step S62, the local path is planned using the RRT* algorithm, the obstacle position, shape, range information and the slope, slope direction, and terrain undulation information of the hill from the dispatch center are input, and the obstacle avoidance path is output, including the following steps: S621, set the current position as the starting point start and the reachable obstacle avoidance point next to the obstacle as the target point goal, and create an RRT tree T containing only start, and define the reachable obstacle avoidance point as the point that is reachable and can effectively avoid obstacles when the plant protection vehicle reaches the point, and define the cost function as the energy consumption of the plant protection vehicle, where the energy consumption represents a function of the hill slope, slope aspect, terrain, and path length, and the environment space is all reachable grid points except obstacles; S622, random sampling: randomly sample a point x_rand in the environment space; S623, find the nearest node x_near from the tree T, take x_near as the starting point, extend a certain distance in the direction, and obtain a new node x_new; S624, detecting whether the path from x_near to x_new collides with an obstacle based on the obstacle position, shape, and range information; if a collision occurs, returning to step S622 to continue the next iteration; S625. Find the node x_min closest to x_new in the tree T, and calculate the cost cost (x_min, x_new) from x_min to x_new. S4626. For the node x_near_neighbors in the tree T whose distance to x_new is within a certain range, calculate the cost cost(x_near,x_new) of connecting to x_new through x_near, and select the connection method with the minimum cost; S627, add x_new to the tree T, and update the connection relationship and cost information between nodes; S628, optimizing some nodes in the tree T with the target point goal as the goal, improving the quality of the path by adjusting the connection relationship and cost information, and repeating steps S622-S628 until a path is found or the maximum number of iterations is reached; S629: If the newly added node x_new is close to the target point goal, a feasible path is found and the algorithm ends.

Images