Unmanned aerial vehicle pesticide spraying control method and system fused with dynamic graph Transform

Abstract

The invention belongs to the technical field of unmanned aerial vehicle spraying, and provides an unmanned aerial vehicle pesticide spraying control method and system fused with a dynamic graph Transform, and the method comprises the steps: obtaining multi-source heterogeneous data at least comprising unmanned aerial vehicle state information and environment information of a to-be-sprayed region; planning a pesticide spraying path of the unmanned aerial vehicle based on the acquired state information of the unmanned aerial vehicle; according to the planned path, the environment information of the to-be-sprayed area and a short-time trajectory prediction model based on a characteristic dynamic attention long-short term memory network, predicting a flight path of the unmanned aerial vehicle for spraying the pesticide; and optimizing the predicted flight path by adopting a decision-making type dynamic graph structure fused with the environment information of the to-be-sprayed area, and controlling the unmanned aerial vehicle to spray the pesticide according to the optimized predicted flight path to complete the control of the pesticide spraying of the unmanned aerial vehicle.

Description

Technical Field

[0001] This invention belongs to the field of drone spraying technology, specifically relating to a drone pesticide spraying control method and system that integrates dynamic graph Transformer. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the acceleration of agricultural modernization, precision plant protection, as a core link in ensuring crop yields and reducing agricultural non-point source pollution, is facing an increasingly urgent need for technological upgrades. Traditional pesticide spraying methods often employ "flood irrigation" operations, relying on manual labor or ground machinery. This results in prominent problems such as low pesticide utilization, poor disease targeting, and sensitivity to environmental disturbances. It not only causes a large amount of pesticide waste and increases agricultural production costs, but also easily leads to ecological risks such as soil pollution, eutrophication of water bodies, and pesticide damage to crops.

[0004] Due to their high operational efficiency, wide coverage, and strong terrain adaptability, drones are gradually becoming the mainstream tool for agricultural plant protection. However, although commercial drone plant protection systems have achieved automated operations such as path planning and quantitative spraying, significant bottlenecks remain in terms of accuracy and safety in complex agricultural environments, making it difficult to meet the needs of refined and intelligent plant protection. On the one hand, there is insufficient fusion of multi-source information and trajectory stability. Existing systems mostly only collect their own position and speed information, lacking dynamic environmental perception and wind speed compensation. They are prone to deviating from their trajectory when encountering wind, and obstacle recognition relies on a single sensor, resulting in delayed obstacle avoidance. On the other hand, there is a disconnect between crop status perception and disease control. Existing systems struggle to obtain microscopic status and disease information, often applying pesticides with fixed parameters, leading to problems such as inadequate coverage of effective areas and over-application of pesticides in healthy areas. At the same time, spraying decisions lack spatiotemporal dynamic optimization and do not fully integrate dynamic data, making it difficult to improve disease control rates while reducing pesticide consumption and drift pollution.

[0005] Currently, technological improvements for drone-based plant protection mostly focus on single aspects, and a comprehensive, end-to-end integrated solution has not yet been formed. Driven by policies promoting green agricultural development and precision plant protection, there is an urgent need in this field for an intelligent system capable of integrating multi-source heterogeneous data, maintaining flight paths safely, and precisely spraying crops. This system should address the problems of low accuracy, high cost, and significant environmental risks associated with existing systems through dynamic environmental adaptation, differentiated disease identification, and adaptive optimization of spraying parameters, thus providing technical support for precision plant protection using agricultural drones. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a drone pesticide spraying control method and system integrating Dynamic Graph Transformer. Through real-time acquisition and processing of multi-source heterogeneous information, trajectory maintenance based on real-time wind speed compensation, automatic obstacle avoidance based on feature-driven dynamic attention LSTM, and a precision spraying decision optimization strategy based on Dynamic Graph Transformer and multi-objective reinforcement learning, this system enables safe and stable drone flight in complex agricultural environments. It achieves precise pesticide spraying for differentiated crop diseases, improves pesticide utilization and disease control effectiveness, reduces pesticide usage costs and environmental drift risks, and provides efficient and intelligent technical support for precision plant protection in agriculture.

[0007] According to some embodiments, the first aspect of the present invention provides a drone pesticide spraying control method that integrates dynamic graph Transformer, employing the following technical solution: A method for controlling pesticide spraying by unmanned aerial vehicles (UAVs) based on a dynamic graph Transformer includes: Acquire multi-source heterogeneous data, including at least drone status information and environmental information of the area to be sprayed; Path planning for drone pesticide spraying is performed based on the acquired drone status information. Based on the planned path, environmental information of the area to be sprayed, and a short-term trajectory prediction model based on a feature-driven dynamic attention long short-term memory network, the flight path of the drone spraying pesticides is predicted. The predicted flight path is optimized by using a decision-making dynamic graph structure that integrates environmental information of the area to be sprayed. The drone is then controlled to spray pesticides based on the optimized predicted flight path, thus completing the control of drone pesticide spraying.

[0008] As a further technical limitation, in the process of planning the path for drone pesticide spraying, the target direction vector of the drone is calculated based on the drone's target waypoint and the current position of the drone in the acquired drone status information. Combined with the drone's cruise speed, the expected ground velocity vector of the drone moving along the obtained target direction vector at the current moment is determined. The determined expected ground velocity vector is the planned drone pesticide spraying path.

[0009] As a further technical limitation, in the process of predicting the flight path of the drone spraying pesticides, the nearest neighbor target is selected from the preset obstacle feature set, the Euclidean distance between each obstacle in the preset obstacle feature set and the drone is calculated, and the obstacle index that minimizes the Euclidean distance is obtained; the time-series input sequence is constructed based on the feature vector of the nearest neighbor obstacle corresponding to the obstacle index that minimizes the Euclidean distance.

[0010] Furthermore, the temporal sequence g is updated based on the constructed temporal input sequence and dynamic attention weights to obtain temporal hidden state features; based on the obtained temporal hidden state features and the output layer of the long short-term memory network, the minimum obstacle distance of the UAV in the future time step is predicted; when the obtained minimum obstacle distance is within the preset safe distance threshold, the avoidance strategy based on the artificial potential field method is activated.

[0011] Furthermore, the avoidance strategy based on the artificial potential field method is as follows: for each obstacle that poses a threat, calculate the virtual repulsive force generated by it at the current position of the UAV; perform vector superposition of the obtained virtual repulsive forces to obtain the total repulsive force, and convert the obtained total repulsive force into an avoidance speed correction vector; and perform weighted fusion of the avoidance speed correction vector with the expected speed after wind speed compensation to generate the final flight speed of the UAV.

[0012] As a further technical limitation, a spatiotemporal graph Transformer fusion reinforcement learning is used to optimize the predicted flight path. A Markov decision process is used to model the drone pesticide spraying. Spatiotemporal coding and an Actor-Critic network are used to generate drone pesticide spraying decisions. The predicted flight path is iteratively optimized using the Proximal Policy Optimization (PPO) algorithm to obtain the final flight path of the drone spraying pesticides.

[0013] According to some embodiments, the second aspect of the present invention provides a drone pesticide spraying control system that integrates dynamic graph Transformer, employing the following technical solution: A drone pesticide spraying control system integrating dynamic graph Transformer includes: The acquisition module is configured to acquire multi-source heterogeneous data, including at least drone status information and environmental information of the area to be sprayed. The planning module is configured to plan the path for drone pesticide spraying based on the acquired drone status information. The prediction module is configured to predict the flight path of the drone spraying pesticides based on the planned path, environmental information of the area to be sprayed, and a short-term trajectory prediction model based on a feature dynamic attention long short-term memory network. The control module is configured to optimize the predicted flight path using a decision-oriented dynamic graph structure that integrates environmental information of the area to be sprayed, and control the drone to spray pesticides based on the optimized predicted flight path, thus completing the control of drone pesticide spraying.

[0014] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium, employing the following technical solution: A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the drone pesticide spraying control method with fused dynamic graph Transformer as described in the first aspect of the present invention.

[0015] According to some embodiments, the fourth aspect of the present invention provides an electronic device, which adopts the following technical solution: An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the drone pesticide spraying control method with fused dynamic graph Transformer as described in the first aspect of the present invention.

[0016] According to some embodiments, the fifth aspect of the present invention provides a computer program product, which adopts the following technical solution: A computer program product includes software code, wherein the program in the software code performs the steps in the drone pesticide spraying control method with fused dynamic graph Transformer as described in the first aspect of the present invention.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes an onboard processor to collect multi-source heterogeneous data in real time, and combines this with cloud-based path calculations to determine the desired moving target, providing a foundation for subsequent precise control. This solution addresses the problem of existing systems only collecting their own status data and lacking dynamic environmental awareness, achieving comprehensive perception of the operational environment and the drone's status, thus laying a data foundation for stable flight paths and precise spraying.

[0018] This invention employs a real-time wind speed compensation algorithm and an LSTM-based automatic obstacle avoidance algorithm with dynamic attention based on features. The former derives the required airspeed through vector derivation and converts it into flight control commands, correcting wind field interference to maintain flight path stability. The latter selects nearest-neighbor targets from the obstacle feature set to construct a temporal input sequence, dynamically generates attention weights using LSTM hidden states, adaptively adjusts the importance of input features, and predicts the minimum obstacle distance in future time steps after LSTM updates. Combined with the artificial potential field method, it generates avoidance commands, solving the problems of easy flight path deviation and obstacle avoidance lag, and improving flight safety in complex environments.

[0019] This invention constructs a spatiotemporal dynamic graph sequence of the operational area, modeling the problem of precision pesticide spraying by UAVs as a Markov decision process (MDP). It deeply fuses multi-source features through a spatiotemporal graph Transformer. The spatial encoding layer relies on an enhanced graph attention network (E-GAT) to construct local and global multi-scale graph representations, introducing cross-modal fusion vectors and learnable environment modulation parameters to achieve cross-modal enhanced fusion of multi-scale spatial features. The temporal encoding layer superimposes the spatial feature sequence and positional encoding before inputting it into a linear Transformer encoder, combining a gated recurrent unit (GRU) to handle temporal evolution and... Adaptive modeling of dynamic disturbances such as wind fields outputs a state representation that deeply integrates spatiotemporal context. Based on this, the Actor-Critic reinforcement learning framework is used to directly generate a hybrid action sequence of pesticide selection and spraying parameters. At the same time, the interaction experience is cached and uploaded to the cloud. The policy model is optimized and iteratively updated offline through the Proximal Policy Optimization (PPO) algorithm. Ultimately, precise targeted spraying of differentiated crop diseases in complex agricultural scenarios is achieved, which significantly improves pesticide utilization and disease control effects, reduces pesticide use costs and environmental drift risks, and provides an innovative technical paradigm for the intelligent and precise development of agricultural plant protection. Attached Figure Description

[0020] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0021] Figure 1 This is a flowchart of the drone pesticide spraying control method using a fused dynamic graph Transformer in Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of the drone pesticide spraying control method using a fused dynamic graph Transformer in Embodiment 1 of the present invention; Figure 3 This is a flowchart of the automatic obstacle avoidance method based on feature dynamic attention LSTM in Embodiment 1 of the present invention; Figure 4 This is a flowchart of the precision crop spraying strategy in Embodiment 1 of the present invention; Figure 5 This is a structural block diagram of the drone pesticide spraying control system based on the fused dynamic graph Transformer in Embodiment 2 of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0026] In this invention, terms such as "fixed connection," "connected," and "linked" should be interpreted broadly, indicating a fixed connection, an integral connection, or a detachable connection; a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can determine the specific meaning of these terms in this invention based on the specific circumstances, and they should not be construed as limitations on the invention.

[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] Example 1 Embodiment 1 of this invention introduces a method for controlling pesticide spraying by unmanned aerial vehicles (UAVs) by incorporating a dynamic graph Transformer.

[0029] like Figure 1 The method for controlling pesticide spraying by drones using a fusion dynamic graph Transformer, as shown, includes: Acquire multi-source heterogeneous data, including at least drone status information and environmental information of the area to be sprayed; Path planning for drone pesticide spraying is performed based on the acquired drone status information. Based on the planned path, environmental information of the area to be sprayed, and a short-term trajectory prediction model based on a feature-driven dynamic attention long short-term memory network, the flight path of the drone spraying pesticides is predicted. The predicted flight path is optimized by using a decision-making dynamic graph structure that integrates environmental information of the area to be sprayed. The drone is then controlled to spray pesticides based on the optimized predicted flight path, thus completing the control of drone pesticide spraying.

[0030] like Figure 2 As shown, this embodiment uses an onboard terminal processor of a drone as its core. Targeting the differentiated and precise control scenarios of downy mildew and powdery mildew in vineyards, it achieves precise pesticide spraying through cloud collaboration. The onboard processor integrates multi-source heterogeneous data, combines a real-time wind speed compensation algorithm with an obstacle avoidance algorithm based on feature-driven dynamic attention-gated LSTM, and ensures the safety and stability of the drone's flight. A crop precision spraying strategy based on dynamic graph Transformer and multi-objective reinforcement learning is adopted to optimize spraying decisions, improve pesticide utilization, control effectiveness, and reduce environmental risks.

[0031] As one or more real-time methods, the real-time perception and trajectory generation stage in this embodiment provides basic data input and initial motion targets for subsequent path planning and automatic obstacle avoidance.

[0032] The airborne processor acquires multi-source heterogeneous data in real time, including the UAV's real-time three-dimensional position vector. Attitude information, velocity vector Wind speed vector of the work area obstacle feature set ,in It is a single obstacle Features , It represents the number of obstacles detected at the current moment.

[0033] This embodiment is based on the desired path distributed from the cloud. Calculate the target direction vector Combined with preset cruise speed Generate the desired ground velocity vector ,Right now ; .

[0034] As one or more implementation methods, the track maintenance and safety assurance phase in this embodiment provides a stable, safe, and accurate execution platform for the precision spraying function, ensuring that the actual position of the UAV meets the mission requirements to the greatest extent possible.

[0035] The required airspeed for the drone is derived based on vector relationships. The input is fed into the underlying flight control system and converted into attitude angles and total thrust control values. The output corrected expected ground velocity is .

[0036] To ensure the safety and continuity of operations in complex environments, this embodiment employs a short-term trajectory prediction model based on a feature-driven dynamic attention long short-term memory network to achieve smooth and forward-looking obstacle avoidance.

[0037] like Figure 3 As shown, this embodiment uses the obstacle feature set Filter the nearest neighbor target ,calculate Every obstacle in Euclidean distance to the drone Find the distance The index of the smallest obstacle ;Right now ; ; in, Indicates the obstacle At that moment The relative position to the drone. If there are no obstacles in the field of view ( ( ), you can input a zero vector representing "safety".

[0038] This embodiment will index as The nearest neighbor obstacle feature vector is used as the time... Input To construct the historical information sequence required for time series forecasting, the past... The above filtering process is repeated at each time step to determine the nearest neighbor target at each time step, forming a time-series input sequence. ,Right now ; .

[0039] In this embodiment, at each time step This mechanism is based on the hidden state of the previous time step. , which is the input feature vector at the current time. An attention weight is dynamically generated for each dimension; an attention score is generated using the hidden state from the previous time step. As a query, the dimensions of the input features are calculated using a feedforward neural network. Same attention vector ,Right now ; in, and These are the weight matrix and bias of the first linear layer. It is the weight vector of the second linear layer.

[0040] In this embodiment, during the calculation of feature attention weights, the attention vector is... pass Function, generating dimension is Feature attention weight vector Each element in the vector Corresponding input features The importance of, and the sum of all weights is 1, that is ; in, .

[0041] This embodiment will calculate the attention weight vector. Compared with the original input feature vector Performing the Hadamard product yields a new input feature vector after attention-gated adjustment. ,Right now .

[0042] Feature vectors processed by attention gating mechanism The data is fed into a standard LSTM gating structure to calculate and update the cell state, resulting in the final hidden state. .

[0043] This embodiment will show the final hidden state. The input is fed into a fully connected output layer, where regression predicts the future. Minimum obstacle distance after time step ,Right now ; in, and These are the weights and biases of the output layer.

[0044] When the predicted minimum obstacle distance Below the preset safety threshold At that time, the avoidance strategy based on the Artificial Potential Field (APF) method is activated; Targeting each obstacle that poses a threat Calculate its current position in the drone The generated virtual repulsive force The repulsive force is directed along the obstacle towards the drone, and its magnitude is related to the distance. Negative correlation, i.e. ; in, The repulsive gain coefficient is... The distance gradient represents the direction of the repulsive force.

[0045] The total repulsive force is obtained by vector superposition of the repulsive forces generated by all threatening obstacles. And convert it into an evasion velocity correction vector. ; Adjust the evasion velocity vector Compared with the expected speed after wind speed compensation The weighted data is then fused to generate the final ground speed command issued to the flight control system. ,Right now ; in, It is a fusion gain coefficient used to adjust the smoothness of the obstacle avoidance response.

[0046] like Figure 4 As shown, the crop precision spraying strategy stage based on dynamic graph Transformer and multi-objective reinforcement learning in this embodiment is aimed at the differentiated precision control scenario of downy mildew and powdery mildew in vineyards. By constructing a decision-making dynamic graph structure that integrates vineyard-specific scenario data, and using spatiotemporal graph Transformer and reinforcement learning methods, precision spraying is modeled as a Markov decision process. Spatiotemporal coding and Actor-Critic network are used to generate application decisions, and the model is iteratively optimized by the cloud-based PPO algorithm, ultimately achieving precise and intelligent application of pesticides for differentiated diseases in vineyards.

[0047] To achieve differentiated medication, state characterization needs to include detailed information that can distinguish different diseases, and construct a spatiotemporal dynamic graph sequence that can accurately describe the dynamic characteristics and spatial correlations of the working environment. .

[0048] In dynamic graph sequence structures It is a collection of nodes, representing a local working area. Discrete grid cells; It is the node feature matrix, where the first node is... row vector Represents a node At any moment of Dimensional features; that is ; in, This indicates vector concatenation.

[0049] Environmental characteristics It includes local wind speed vectors, temperature, and humidity.

[0050] Crop status characteristics It includes the normalized vegetation index, leaf area index, and disease probability vector. Disease probability vector It uses a pre-trained lightweight CNN model to process airborne multispectral images. Obtained through real-time analysis, i.e. ; in, These represent the probabilities of the unit being healthy, suffering from downy mildew, and suffering from powdery mildew, respectively.

[0051] Characteristics of historical assignments This includes the cumulative flow of pesticide A (for downy mildew) and pesticide B (for powdery mildew) sprayed at this node throughout its history. Among them, real-time feedback features It includes the current pesticide coverage rate provided by the real-time coverage estimation module, which identifies the pesticide coverage area by analyzing crop images captured by the airborne camera through U-Net identification and analyzes the pesticide coverage area and outputs the real-time pesticide coverage rate; It is a dynamically weighted adjacency matrix, whose elements Characterizes at time node For nodes The intensity of the impact. The construction integrates multi-source dynamic information, namely ; Spatial proximity factor Use a Gaussian kernel function to measure the Euclidean distance between nodes Modeling is performed to show that the influence intensity decays with distance, i.e. ; in, This is a distance scaling hyperparameter.

[0052] Wind field directionality factor Used to simulate the directional effects of drug droplet drift, i.e. ; in, For the node point to vector With real-time wind direction vector The angle between them. This factor makes the weight of the downwind side close to 1 and the weight of the upwind side close to 0.

[0053] Environmental and crop status correction factors Correcting for differences in humidity and crop health status, i.e. ; in, and They are nodes Differences in humidity and crop health index between regions These are adjustable weight parameters.

[0054] In this embodiment, precision spraying is modeled as a Markov decision process (MDP), defined as a tuple. ;in, state space S :state From the past The system is constructed using a sequence of spatiotemporal dynamic graphs at each time step to capture its temporal dynamics. ; Hybrid motion space :action This is the spraying decision vector that the drone can execute. Discrete part This indicates the choice of pesticide, representing no spraying, spraying pesticide A, and spraying pesticide B respectively; continuous part These represent spraying parameters, specifically flow rate and spray width; State transition probability The model does not require explicit learning; it is jointly determined by the physical and biological dynamics of the environment. Multi-objective reward function ,Right now ; in, Proportional to the decrease in the probability of the target disease, The uniformity and effectiveness of coverage in the target area are proportional to the standard deviation and mean of the coverage. It is directly proportional to the consumption of medicines and the energy consumption of drones. . With drift risk assessment function The value is directly proportional to the value; Discount factor .

[0055] To enhance the spatiotemporal modeling capabilities of the Transformer structure, this embodiment introduces a dynamic multi-scale cross-modal attention mechanism.

[0056] In the spatial coding layer, for each time step Constructing local scale maps and global scale map The multi-scale graph representation employs an Enhanced Graph Attention Network (E-GAT) to aggregate information. For each scale... and nodes Updated feature representation ;Right now ; ; ; in, It is the attention coefficient, calculated through a cross-modal query-key mechanism; It is a cross-modal fusion vector. These are learnable environmental modulation parameters.

[0057] go through After layer E-GAT, multi-scale embeddings are fused: .

[0058] In the temporal coding layer, the spatial feature sequence With position encoding After addition, the input is given to a linear Transformer encoder, whose core attention mechanism is: ; in, It is a random feature mapping, approximation To achieve linear complexity, a dynamic graph evolution module is inserted. After each Transformer layer, a GRU is used to predict the graph state for the next potential time step. ; ; in, It is the predicted adjacency matrix.

[0059] Finally, the Transformer outputs the final state representation of the deeply fused spatiotemporal context. .

[0060] This embodiment utilizes an actor network. Will Mapped to a hybrid action space, the discrete action head outputs the probability distribution of selecting various pesticides. The mean of the Gaussian distribution of the continuous motion head output and standard deviation The actual actions are obtained by sampling from this distribution. ,Right now ; ; Critics Network Evaluate the value of the state and output it, i.e. .

[0061] In this embodiment, the onboard processor will sense in real time... Enter the deployed and Model, generates optimal action ,in ;action Parsed into low-level hardware control instructions This drives the corresponding drug pumps and nozzle actuators; closed-loop correction and model iteration are performed. After the drone performs its actions, the environment evolves into... And calculate the reward ; Empirical tuple The parameters are cached and sent to the cloud server; the cloud uses Proximal Policy Optimization (PPO) to optimize the model parameters. and Asynchronous updates are performed, and their core objective function is: ; ; in It is the strategy probability ratio. It is the advantage function estimate.

[0062] This embodiment uses an onboard processor to collect multi-source heterogeneous data in real time, and combines this with cloud-based path calculations to determine the desired moving target, providing a foundation for subsequent precise control. This solution addresses the problem of existing systems only collecting their own status data and lacking dynamic environmental awareness, achieving comprehensive perception of the operating environment and the drone's status, thus laying a data foundation for stable flight paths and precise spraying.

[0063] This embodiment employs a real-time wind speed compensation algorithm and an LSTM-based automatic obstacle avoidance algorithm with dynamic attention based on features. The former derives the required airspeed through vector derivation and converts it into flight control commands, correcting wind field interference to maintain flight path stability. The latter selects nearest-neighbor targets from the obstacle feature set to construct a temporal input sequence, dynamically generates attention weights using LSTM hidden states, adaptively adjusts the importance of input features, and predicts the minimum obstacle distance in future time steps after LSTM updates. Combined with the artificial potential field method, it generates avoidance commands, solving the problems of easy flight path deviation and obstacle avoidance lag, and improving flight safety in complex environments.

[0064] This embodiment constructs a spatiotemporal dynamic graph sequence of the operation area, modeling the problem of precision pesticide spraying by UAVs as a Markov decision process (MDP). It deeply fuses multi-source features through a spatiotemporal graph Transformer. The spatial encoding layer relies on an enhanced graph attention network (E-GAT) to construct local and global multi-scale graph representations, introducing cross-modal fusion vectors and learnable environment modulation parameters to achieve cross-modal enhanced fusion of multi-scale spatial features. The temporal encoding layer superimposes the spatial feature sequence and positional encoding and inputs them into a linear Transformer encoder, combining a gated recurrent unit (GRU) to analyze the temporal evolution and... Adaptive modeling of dynamic disturbances such as wind fields outputs a state representation that deeply integrates spatiotemporal context. Based on this, the Actor-Critic reinforcement learning framework is used to directly generate a hybrid action sequence of pesticide selection and spraying parameters. At the same time, the interaction experience is cached and uploaded to the cloud. The policy model is optimized and iteratively updated offline through the Proximal Policy Optimization (PPO) algorithm. Ultimately, precise targeted spraying of differentiated crop diseases in complex agricultural scenarios is achieved, which significantly improves pesticide utilization and disease control effects, reduces pesticide use costs and environmental drift risks, and provides an innovative technical paradigm for the intelligent and precise development of agricultural plant protection.

[0065] Example 2 Embodiment 2 of the present invention introduces a drone pesticide spraying control system that integrates dynamic graph Transformer.

[0066] like Figure 5 The illustrated UAV pesticide spraying control system based on a fusion dynamic graph Transformer includes: The acquisition module is configured to acquire multi-source heterogeneous data, including at least drone status information and environmental information of the area to be sprayed. The planning module is configured to plan the path for drone pesticide spraying based on the acquired drone status information. The prediction module is configured to predict the flight path of the drone spraying pesticides based on the planned path, environmental information of the area to be sprayed, and a short-term trajectory prediction model based on a feature dynamic attention long short-term memory network. The control module is configured to optimize the predicted flight path using a decision-oriented dynamic graph structure that integrates environmental information of the area to be sprayed, and control the drone to spray pesticides based on the optimized predicted flight path, thus completing the control of drone pesticide spraying.

[0067] The detailed steps are the same as those of the drone pesticide spraying control method with fused dynamic graph Transformer provided in Example 1, and will not be repeated here.

[0068] Example 3 Embodiment 3 of the present invention provides a computer-readable storage medium.

[0069] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the drone pesticide spraying control method with fused dynamic graph Transformer as described in Embodiment 1 of the present invention.

[0070] The detailed steps are the same as those of the drone pesticide spraying control method with fused dynamic graph Transformer provided in Example 1, and will not be repeated here.

[0071] Example 4 Embodiment 4 of the present invention provides an electronic device.

[0072] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the drone pesticide spraying control method with fused dynamic graph Transformer as described in Embodiment 1 of the present invention.

[0073] The detailed steps are the same as those of the drone pesticide spraying control method with fused dynamic graph Transformer provided in Example 1, and will not be repeated here.

[0074] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0075] A computer program product includes software code, wherein the program in the software code performs the steps of the drone pesticide spraying control method with fused dynamic graph Transformer as described in Embodiment 1 of the present invention.

[0076] The detailed steps are the same as those of the drone pesticide spraying control method with fused dynamic graph Transformer provided in Example 1, and will not be repeated here.

[0077] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0078] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0079] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0081] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0082] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0083] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for controlling pesticide spraying by unmanned aerial vehicles (UAVs) integrating dynamic graph Transformer, characterized in that, include: Acquire multi-source heterogeneous data, including at least drone status information and environmental information of the area to be sprayed; Path planning for drone pesticide spraying is performed based on the acquired drone status information. Based on the planned path, environmental information of the area to be sprayed, and a short-term trajectory prediction model based on a feature-driven dynamic attention long short-term memory network, the flight path of the drone spraying pesticides is predicted. The predicted flight path is optimized by using a decision-making dynamic graph structure that integrates environmental information of the area to be sprayed. The drone is then controlled to spray pesticides based on the optimized predicted flight path, thus completing the control of drone pesticide spraying.

2. The method for controlling pesticide spraying by a drone using a fusion dynamic graph Transformer as described in claim 1, characterized in that, In the process of planning the path for drone pesticide spraying, the target direction vector of the drone is calculated based on the drone's target waypoint and the current position of the drone in the acquired drone status information. Combined with the drone's cruise speed, the expected ground velocity vector of the drone moving along the obtained target direction vector at the current moment is determined. The determined expected ground velocity vector is the planned drone pesticide spraying path.

3. The method for controlling pesticide spraying by a drone using a fusion dynamic graph Transformer as described in claim 1, characterized in that, In the process of predicting the flight path of a drone spraying pesticides, the nearest neighbor target is selected from the preset obstacle feature set, and the Euclidean distance between each obstacle in the preset obstacle feature set and the drone is calculated to obtain the obstacle index that minimizes the Euclidean distance. The time-series input sequence is constructed based on the feature vector of the nearest neighbor obstacle corresponding to the obstacle index that minimizes the Euclidean distance.

4. The method for controlling pesticide spraying by a drone using a fusion dynamic graph Transformer as described in claim 3, characterized in that, The temporal sequence is updated based on the constructed temporal input sequence and dynamic attention weights to obtain temporal hidden state features; Based on the obtained temporal hidden state features and the output layer of the long short-term memory network, the minimum obstacle distance of the drone in the future time step is predicted; when the obtained minimum obstacle distance is within the preset safe distance threshold, the avoidance strategy based on the artificial potential field method is activated.

5. The method for controlling pesticide spraying by a drone using a fusion dynamic graph Transformer as described in claim 4, characterized in that, The avoidance strategy based on the artificial potential field method is as follows: for each obstacle that poses a threat, calculate the virtual repulsive force generated by it at the current position of the UAV; superimpose the obtained virtual repulsive forces to obtain the total repulsive force, and convert the obtained total repulsive force into an avoidance speed correction vector; and weight and fuse the avoidance speed correction vector with the expected speed after wind speed compensation to generate the final flight speed of the UAV.

6. The method for controlling pesticide spraying by a drone using a fusion dynamic graph Transformer as described in claim 1, characterized in that, The predicted flight path is optimized by using a spatiotemporal graph Transformer fusion reinforcement learning. The drone pesticide spraying is modeled using a Markov decision process. The drone pesticide spraying decision is generated by spatiotemporal coding and an Actor-Critic network. The predicted flight path is iteratively optimized by combining a proximal policy algorithm to obtain the final flight path of the drone spraying pesticides.

7. A drone pesticide spraying control system integrating dynamic graph Transformer, characterized in that, include: The acquisition module is configured to acquire multi-source heterogeneous data, including at least drone status information and environmental information of the area to be sprayed. The planning module is configured to plan the path for drone pesticide spraying based on the acquired drone status information. The prediction module is configured to predict the flight path of the drone spraying pesticides based on the planned path, environmental information of the area to be sprayed, and a short-term trajectory prediction model based on a feature dynamic attention long short-term memory network. The control module is configured to optimize the predicted flight path using a decision-oriented dynamic graph structure that integrates environmental information of the area to be sprayed, and control the drone to spray pesticides based on the optimized predicted flight path, thus completing the control of drone pesticide spraying.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the drone pesticide spraying control method of fused dynamic graph Transformer as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the drone pesticide spraying control method of fused dynamic graph Transformer as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code performs the steps of the drone pesticide spraying control method of the fused dynamic graph Transformer as described in any one of claims 1-6.

Images