Unmanned aerial vehicle grain depot fumigation agent delivery method and system

Abstract

The application discloses a kind of unmanned aerial vehicle granary fumigation agent delivery method and system, it is related to granary fumigation agent delivery technical field;Method includes: real-time acquisition obtains perception data;Perception data is preprocessed, and observation vector is obtained, and observation vector is encoded as belief state;Based on perception data, target medicine tray is determined by decision;According to target medicine tray, belief state is input into trained SAC strategy network, and continuous movement action is output to obtain;Unmanned aerial vehicle executes continuous movement action to reach target medicine tray, and triggers medicine delivery instruction when satisfying set medicine delivery condition;Unmanned aerial vehicle completes medicine delivery action according to medicine delivery instruction, and updates medicine tray map, determines next target medicine tray by decision, until completing all medicine tray medicine delivery action;The application can adapt to complex indoor environment of granary and realize accurate medicine delivery, and spraying uniformity is better, and operation efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of grain depot fumigation agent delivery technology, and more specifically, to a method and system for unmanned aerial vehicle (UAV) grain depot fumigation agent delivery. Background Technology

[0002] Grain fumigation is a key technology for controlling stored grain pests and ensuring food security. Traditional fumigation operations mostly use fixed pipeline circulation or manual application. Fixed pipeline systems are costly to build, lack flexibility, and have blind spots; manual application suffers from uneven application, low efficiency, and significant safety risks to workers exposed to toxic and harmful environments.

[0003] In recent years, drone technology has been widely used in agriculture due to its high mobility and flexibility, such as in pesticide spraying. However, the indoor fumigation environment of grain depots poses unique challenges to drone technology: (1) there is no GPS signal, and it relies on indoor positioning; (2) the environmental structure is complex and there are a lot of obstacles; (3) the operation requirements are high, requiring precise positioning of the pre-placed pesticide trays and application of pesticides, rather than large-area spraying.

[0004] Existing drone cruise algorithms are mostly based on Global path planning methods such as Reinforcement Learning (RRT) rely on pre-built high-precision maps and have poor adaptability in dynamic or highly uncertain environments. Reinforcement learning, as a learning paradigm that interacts with the environment and learns through trial and error, offers a new approach to solving such problems. However, directly applying end-to-end reinforcement learning to such complex tasks suffers from low training efficiency and poor performance in making decisions about discrete drug delivery actions.

[0005] In view of this, the present invention proposes a method and system for delivering fumigation agents for grain depots by drone to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a method for delivering fumigation agents for grain depots by unmanned aerial vehicles (UAVs), comprising:

[0007] A method for delivering fumigation agents to grain depots by drone, characterized in that it includes:

[0008] Real-time acquisition of sensing data, including environmental information and UAV status information;

[0009] The perceived data is preprocessed to obtain observation vectors, and the observation vectors are encoded as belief states;

[0010] Based on the perceived data, the decision-making process determines the target medicine tray;

[0011] Based on the target medicine tray, the belief state is input into the trained SAC policy network, and the output is a continuous movement action; the UAV executes the continuous movement action to reach the target medicine tray, and triggers the medicine delivery command when the set medicine delivery conditions are met.

[0012] The drone completes the drug delivery action according to the drug delivery instructions, updates the drug tray map, decides the next target drug tray, and continues until the drug delivery action of all drug trays is completed.

[0013] Furthermore, the environmental information includes environmental images, environmental depth, and environmental gas concentration; the environmental images also include visual markers on the medicine tray, and the pose of the medicine tray relative to the UAV is calculated based on the visual markers.

[0014] The drone status information includes the drone's acceleration, angular velocity, remaining pill quantity, remaining battery power, and medicine tray map; the medicine tray map is used to record the global location of discovered medicine trays and their corresponding delivery status; the delivery status is the ratio of the quantity delivered to the target quantity; the medicine tray map is initially empty, with no medicine tray locations.

[0015] Furthermore, methods for determining target medicine trays based on perceived data include:

[0016] The medicine tray is identified by matching the visual markers with the features of the preset medicine tray visual marker template.

[0017] By combining the acceleration and angular velocity of the UAV, the real-time global pose of the UAV is calculated.

[0018] Based on the real-time global poses of the identified medicine tray and the drone, the global position of the corresponding medicine tray is obtained.

[0019] If the global location of the corresponding medicine tray does not appear in the medicine tray map, the corresponding medicine tray is determined to be an undelivered medicine tray.

[0020] Get the delivery status of all undelivered medicine trays, and prioritize the medicine trays with the lowest proportion of delivered quantity to target delivery quantity as candidate target medicine trays.

[0021] If multiple candidate target pharmacy trays have the same delivery status ratio, calculate the straight-line distance from the UAV to each candidate target pharmacy tray, and select the candidate target pharmacy tray with the closest distance as the target pharmacy tray.

[0022] Furthermore, the reward function used when training the SAC policy network is a multi-objective weighted sum, and the objectives of the reward function include process terms, delivery terms, safety terms, and efficiency terms; the weight coefficients of each objective are set according to the actual achievement of each objective.

[0023] Furthermore, a dosing window is pre-set. When the proportion of the target medicine tray within the dosing window in the acquired environmental image reaches a set percentage, and the target medicine tray is directly below the drone, the dosing conditions are determined to be met.

[0024] Furthermore, continuous movement includes at least three-axis flight speed and flight angle.

[0025] Furthermore, based on the perception data, it can determine in real time whether a dangerous situation has occurred, and if a dangerous situation has occurred, it can trigger an emergency avoidance procedure.

[0026] Furthermore, methods for determining whether a dangerous situation has occurred include:

[0027] The system uses environmental depth as a basis for judgment. When the distance between the drone and an obstacle is less than a preset collision safety threshold, a collision is determined to be imminent. It also uses remaining battery power as a basis for judgment. When the remaining battery power is lower than a preset low battery power threshold, the battery is determined to be too low. Finally, it uses ambient gas concentration as a basis for judgment. When the ambient gas concentration exceeds a preset lower explosive limit threshold, the gas concentration is determined to be excessive.

[0028] Furthermore, emergency avoidance procedures include:

[0029] If a collision is imminent, the drone will be controlled to adjust its flight attitude and move away from the obstacle to avoid the risk of collision; if the battery is too low, the drone's backup power supply will be activated; if the gas concentration exceeds the limit, the drone's drug delivery mechanism will be controlled to stop the drug delivery action and close the drug delivery mechanism, and the drone will be controlled to return to the pre-set safe docking point.

[0030] A drone-based fumigation agent delivery system for grain depots includes:

[0031] The data acquisition module is used to collect and acquire sensing data in real time, including environmental information and UAV status information;

[0032] The data processing module is used to preprocess the sensed data, obtain the observation vector, and encode the observation vector into belief states;

[0033] The target decision-making module determines the target medicine tray based on perceived data.

[0034] The first execution module is used to input the belief state into the trained SAC policy network based on the target medicine tray and output continuous movement actions; the UAV executes continuous movement actions to reach the target medicine tray and triggers the medicine dispensing command when the set medicine dispensing conditions are met.

[0035] The second execution module is used to complete the drug delivery action according to the drug delivery instruction, update the drug tray map, determine the next target drug tray, and continue until the drug delivery action of all drug trays is completed.

[0036] Compared with the prior art, the technical effects and advantages of the UAV fumigation agent delivery method and system for grain depots of the present invention are as follows:

[0037] The system of this invention includes a data acquisition module, a data processing module, a target decision module, a first execution module, and a second execution module. The method involves real-time acquisition of perception data containing environmental information and UAV status information, preprocessing the perception data to obtain observation vectors and encoding them as belief states, determining the target medicine tray based on the perception data, inputting the belief states into a trained SAC policy network to output continuous movement actions, and having the UAV execute actions to reach the target medicine tray. When the preset medicine delivery conditions are met, a medicine delivery command is triggered. After completing the medicine delivery action according to the medicine delivery command, the medicine tray map is updated, and the next target medicine tray is determined until all medicine trays are delivered. During the process, dangerous situations are also judged in real time and emergency avoidance procedures are triggered. The SAC policy network is trained through offline simulation, and a reward function that takes into account multiple objectives such as process, delivery, safety, and efficiency is used during training.

[0038] This invention addresses the problems of high construction costs, poor flexibility, and fumigation blind spots associated with traditional fixed pipeline circulation systems; uneven and inefficient manual pesticide application; and significant safety risks to workers exposed to toxic and hazardous environments. It also solves the challenges of grain depots lacking GPS signals, having complex environments with obstacles, requiring precise positioning of the pesticide tray, and dealing with flammable and explosive fumigation gases. Furthermore, it overcomes the limitations of existing UAV navigation algorithms, which rely on high-precision maps, suffer from low efficiency in end-to-end reinforcement learning training, and exhibit poor decision-making performance for discrete pesticide application actions. Its beneficial effects and advantages include eliminating the need for workers to enter hazardous environments, effectively ensuring personnel safety; improved pesticide uniformity; increased operational efficiency; adaptability to complex indoor grain depot environments for precise pesticide application; emergency avoidance procedures further ensuring operational safety; and a SAC strategy network training method that makes UAV operations more stable and adaptable to different situations. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of a drone-based grain depot fumigation agent delivery system according to an embodiment of the present invention;

[0040] Figure 2 This is a flowchart of a method for delivering fumigation agents to a grain depot using a drone, according to an embodiment of the present invention.

[0041] Figure 3 This is a flowchart of the offline training phase method according to an embodiment of the present invention;

[0042] Figure 4 A flowchart illustrating the method for determining the target medicine tray in an embodiment of the present invention;

[0043] Figure 5This is a schematic diagram illustrating the periodic variation of the reward function in an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0045] Example 1:

[0046] Please see Figure 1 As shown in the figure, this embodiment discloses a drone-based grain depot fumigation agent delivery system, including a data acquisition module, a data processing module, a target decision module, a first execution module, and a second execution module. Each module is connected via wired and / or wireless means to achieve data transmission.

[0047] The data acquisition module is used to collect and acquire sensing data in real time, including environmental information and UAV status information.

[0048] Environmental information includes environmental images, environmental depth, and environmental gas concentration; the environmental images also include visual markers on the medicine tray, and the pose of the medicine tray relative to the UAV is calculated based on the visual markers; UAV status information includes the UAV's acceleration, angular velocity, remaining pill quantity, remaining battery power, and medicine tray map; the medicine tray map is used to record the global location of discovered medicine trays and their corresponding delivery status; the delivery status is the ratio of the quantity delivered to the target quantity; the medicine tray map is initially empty, with no medicine tray locations.

[0049] Among them, environmental images are acquired through visual sensors, environmental depth is acquired through lidar or depth cameras, and environmental gas concentration is acquired through gas sensors; the drone's acceleration and angular velocity are acquired through inertial measurement units, and the remaining number of pills, remaining battery power, and medicine tray map are maintained through the drone's internal state unit.

[0050] The data processing module is used to preprocess the sensed data to obtain observation vectors and encode the observation vectors into belief states.

[0051] The environmental image is denoised and distortion corrected to extract the visual marker features of the medicine tray, including shape and color features. The environmental depth is filtered to extract obstacle distance features. Acceleration and angular velocity data collected by the inertial measurement unit are smoothed to obtain the UAV's motion state features. Outlier removal of the ambient gas concentration is performed to obtain gas concentration data. The processed visual marker features, obstacle distance features, motion state features, gas concentration data, and the remaining number of medicine tablets and remaining battery power data in the internal state unit are normalized and integrated into a unified-dimensional observation vector. A recurrent neural network is used, taking the current observation vector and the historical observation vector sequence as input. Through the memory mechanism of the recurrent neural network, the dynamic change pattern of the observation vector over time is learned, and the output is a belief state that integrates current and historical information, representing the UAV's comprehensive cognition of the environment and its own state.

[0052] The target decision-making module determines the target medicine tray based on the perceived data.

[0053] Please see Figure 4As shown, the extracted visual marker features are matched with the preset medicine tray visual marker template features, and the matching degree is calculated. If the matching degree reaches the preset matching threshold, the medicine tray is identified. Acceleration and angular velocity data of the UAV are acquired. The acceleration data is denoised and gravity separated to obtain the linear acceleration of the UAV in the carrier coordinate system. The angular velocity data is denoised to obtain the angular velocity of the UAV in the carrier coordinate system. The processed angular velocity is integrated to obtain the change in the UAV's attitude angle. Combined with the UAV's initial attitude angle, the real-time attitude angle of the UAV in the global coordinate system is determined. The processed linear acceleration is transformed to the global coordinate system, and then integrated twice to obtain the change in the UAV's displacement. Combined with the UAV's initial position coordinates, the real-time position coordinates of the UAV in the global coordinate system are determined. The carrier coordinate system has the UAV's center of mass as its origin. The system uses a three-dimensional Cartesian coordinate system with the forward direction as the X-axis, the rightward direction as the Y-axis, and the vertical upward direction as the Z-axis. The global coordinate system uses a preset fixed point within the grain depot as the origin, with the length of the grain depot as the X-axis, the width as the Y-axis, and the vertical upward direction as the Z-axis. Based on the real-time global pose of the identified medicine trays and the drone, the global position of the corresponding medicine tray is obtained. The global position of the corresponding medicine tray is compared with the positions of completed medicine trays recorded in the medicine tray map. If the global position of the corresponding medicine tray does not appear in the medicine tray map, the corresponding medicine tray is determined to be an incomplete medicine tray. The delivery status of all incomplete medicine trays is obtained, and the medicine tray with the lowest proportion of delivered quantity to the target delivery quantity is selected as the candidate target medicine tray. If multiple candidate target medicine trays have the same delivery status proportion, the straight-line distance from the drone to each candidate target medicine tray is calculated, and the candidate target medicine tray with the closest distance is selected as the final target medicine tray.

[0054] The first execution module is used to input the belief state into the trained SAC policy network based on the target medicine tray and output continuous movement actions; the UAV executes the continuous movement actions to reach the target medicine tray and triggers the medicine delivery command when the set medicine delivery conditions are met.

[0055] Please see Figure 3 As shown, the methods for offline training of SAC policy networks include:

[0056] S101: Construct a virtual grain depot scene in a simulation environment, and randomly initialize the positions of medicine trays, obstacles, and drones;

[0057] S102: Initialize SAC policy network parameters and experience replay buffer;

[0058] S103: Real-time acquisition of sensing data ;

[0059] S104: Encode perceived data and historical perceived data into belief states. Input the SAC policy network to generate continuous movement actions. ;

[0060] S105: After the drone performs the action, the reward value is calculated according to the pre-designed reward function. It then transitions to a new state, generating new perceptual data. ;

[0061] S106: Transfer the empirical tuple Store the data in the experience replay buffer and sample the data to update the SAC policy network parameters;

[0062] S107: Repeat steps S103 to S106 until the SAC policy network converges. Save the trained SAC policy network parameters to obtain the trained SAC policy network.

[0063] The reward function For a multi-objective weighted sum, its expression is:

[0064] ;

[0065] In the formula, For process items, For delivery items, For safety reasons, For efficiency, All are weighting coefficients.

[0066] in, , Indicates time The Euclidean distance between the drone and the current target medicine tray. Indicates time -1 The Euclidean distance between the drone and the current target medicine tray. and The value is calculated visually. After the lidar is activated, a spatial model of the room is established, and the coordinates are transformed based on the drone's flight altitude and the pixel position of the target medicine tray within the camera's field of view. and When the difference is greater than zero, it indicates that the drone is approaching the target and a positive reward is given; otherwise, a negative reward is given. A successful dosing instruction function is defined if and only if dosing is successful, i.e., a dosing command is issued and the drone is within the dosing window. The value is positive; otherwise, it is 0. This is the erroneous dosing indication function. The value is [value] when an erroneous dosing occurs, i.e., a dosing command is issued but the drone is not within the dosing window. ,and It is negative if it is negative, otherwise it is 0. These are collision indication functions, distance indication functions, and concentration indication functions, which are triggered when a collision occurs, when the distance to an obstacle is less than a safe threshold, and when the gas concentration is detected to exceed the lower explosive limit, respectively. The corresponding weighting coefficients are... , All are negative values. A constant-time penalty is applied to each step, and . It is a constant penalty for each drug administration action, and Please see. Figure 5 As shown, Figure 5 The trend of the reward function changing with the training cycle in a preferred simulation result.

[0067] For example, the SAC algorithm is implemented using the PyTorch framework. Simulation training is performed on an NVIDIA GPU server, using the AirSim simulation platform to construct a grain depot environment. The SAC policy network structure consists of a policy network and a value network. and value network Both layers employ fully connected neural networks with two hidden layers, each containing 256 neurons, and using ReLU activation. The Actor network in the output layer uses the Tanh activation function, with an output range of [-1, 1], and is superimposed with a learnable standard deviation to construct a Gaussian policy. The Critic network outputs a single-valued Q-value as the value network. An additional network with the same structure generates the Q-value for the next state to assist the Critic network in parameter updates. Example hyperparameter settings are as follows: discount factor γ = 0.99, replay buffer size... Soft update coefficient The learning rate for all networks is Batch size is The target entropy is This corresponds to 4-dimensional continuous actions. After grid search, a set of effective weights is set as follows:

[0068]

[0069] After approximately 5 million steps of simulation training, the SAC policy network successfully converged.

[0070] The second execution module is used to complete the drug delivery action according to the drug delivery instruction, update the drug tray map, determine the next target drug tray, and continue until the drug delivery action of all drug trays is completed.

[0071] After the dosing command is triggered, the dosing mechanism first performs a self-check to confirm that there are enough tablets in the tablet storage compartment and that the dosing channel is not blocked. After the self-check is passed, the dosing mechanism starts the built-in drive motor to push the specified number of tablets down along the dosing channel and into the target medicine tray. At the same time, the counting sensor on the dosing mechanism detects the number of falling tablets in real time. When the detected number is consistent with the number required by the dosing command, the drive motor stops running, the dosing channel closes, and the single dosing action is completed.

[0072] Obtain the global location of the target medicine tray that has been successfully delivered. Retrieve the historical delivery quantity of the corresponding target medicine tray from the drone status information. Add the historical delivery quantity to the current delivery quantity to obtain the new delivery quantity. Calculate the ratio of the new delivery quantity to the target delivery quantity of the corresponding target medicine tray and update the delivery status of the corresponding target medicine tray. If the new delivery quantity is equal to the target delivery quantity, record the global location of the corresponding target medicine tray and the completed delivery mark in the medicine tray map. If the new delivery quantity is less than the target delivery quantity, only update the delivery quantity and corresponding delivery status of the corresponding target medicine tray in the medicine tray map, without adding a completed delivery mark.

[0073] The system processes the environmental image to identify all medicine trays in the current environment and obtain their global positions. It then compares the global positions of all medicine trays with the positions of completed medicine trays recorded in the medicine tray map, filtering out medicine trays not appearing on the map and those already on the map but with a delivery count less than the target delivery count, forming a set of medicine trays with incomplete delivery. The system calculates the delivery status of each medicine tray in this set, prioritizing the medicine tray with the lowest delivery count relative to the target delivery count as a candidate target medicine tray. If multiple candidate target medicine trays have the same delivery status ratio, the system calculates the straight-line distance from the UAV to each candidate target medicine tray using the depth sensor and the UAV's real-time global pose, selecting the closest candidate target medicine tray as the next target medicine tray.

[0074] After determining the next target medicine tray, the drone moves continuously to approach the target medicine tray based on its global position, repeating the steps of medicine delivery, medicine tray map update, and next target medicine tray decision. If the set of medicine trays that have not been delivered is empty, that is, the number of medicine trays delivered has reached the target number, then it is determined that the medicine delivery action of all medicine trays has been completed, the drone stops the medicine delivery-related operations, and executes the return procedure.

[0075] Example 2:

[0076] Please see Figure 2 As shown, this embodiment provides a method for delivering fumigation agents to grain depots using unmanned aerial vehicles (UAVs), including:

[0077] Real-time acquisition of sensing data, including environmental information and UAV status information;

[0078] The perceived data is preprocessed to obtain observation vectors, and the observation vectors are encoded as belief states;

[0079] Based on the perceived data, the decision-making process determines the target medicine tray;

[0080] Based on the target medicine tray, the belief state is input into the trained SAC policy network, and the output is a continuous movement action; the UAV executes the continuous movement action to reach the target medicine tray, and triggers the medicine delivery command when the set medicine delivery conditions are met.

[0081] The drone completes the drug delivery action according to the drug delivery instructions, updates the drug tray map, decides the next target drug tray, and continues until the drug delivery action of all drug trays is completed.

[0082] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0083] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for delivering fumigation agents to grain depots using unmanned aerial vehicles (UAVs), characterized in that, include: Real-time acquisition of perception data, including environmental information and UAV status information; environmental information includes environmental images, environmental depth, and environmental gas concentration; environmental images also include visual markers on the medicine tray, and the pose of the medicine tray relative to the UAV is calculated based on the visual markers; The drone's status information includes the drone's acceleration, angular velocity, remaining pill count, remaining battery power, and a medicine tray map. The medicine tray map records the global location of discovered medicine trays and their corresponding delivery status. The delivery status is the ratio of the number of medicine trays delivered to the target number of medicine trays to be delivered. The medicine tray map is initially empty, with no medicine tray locations. The perceived data is preprocessed to obtain observation vectors, and the observation vectors are encoded as belief states; Based on the perceived data, the decision-making process determines the target medicine tray; Methods for determining target medicine trays based on perception data include: The medicine tray is identified by matching the visual markers with the features of the preset medicine tray visual marker template. By combining the acceleration and angular velocity of the UAV, the real-time global pose of the UAV is calculated. Based on the real-time global poses of the identified medicine tray and the drone, the global position of the corresponding medicine tray is obtained. If the global location of the corresponding medicine tray does not appear in the medicine tray map, the corresponding medicine tray is determined to be an undelivered medicine tray. Get the delivery status of all undelivered medicine trays, and prioritize the medicine trays with the lowest proportion of delivered quantity to target delivery quantity as candidate target medicine trays. If multiple candidate target pharmacy trays have the same delivery status ratio, calculate the straight-line distance from the UAV to each candidate target pharmacy tray and select the candidate target pharmacy tray with the closest distance as the target pharmacy tray; Based on the target medicine tray, the belief state is input into the trained SAC policy network, and the output is a continuous movement action; the UAV executes the continuous movement action to reach the target medicine tray, and triggers the medicine delivery command when the set medicine delivery conditions are met. The drone completes the drug delivery action according to the drug delivery instructions, updates the drug tray map, decides the next target drug tray, and continues to complete the drug delivery action for all drug trays.

2. The method for delivering fumigation agents to a grain depot by unmanned aerial vehicle (UAV) according to claim 1, characterized in that, The reward function used when training the SAC policy network is a multi-objective weighted sum. The objectives of the reward function include process, delivery, safety, and efficiency. The weight coefficients of each objective are set according to the actual achievement of each objective.

3. The method for delivering fumigation agents to a grain depot by unmanned aerial vehicle (UAV) according to claim 1, characterized in that, A dosing window is pre-set. When the target medicine tray occupies a set percentage within the dosing window in the acquired environmental image, and the target medicine tray is directly below the drone, the dosing conditions are considered met.

4. The method for delivering fumigation agents to grain depots by unmanned aerial vehicles according to claim 1, characterized in that, Continuous movement includes at least three-axis flight speed and flight angle.

5. The method for delivering fumigation agents to a grain depot by unmanned aerial vehicle (UAV) according to claim 1, characterized in that, Based on the sensing data, it can determine in real time whether a dangerous situation has occurred, and if a dangerous situation has occurred, it will trigger an emergency avoidance procedure.

6. A method for delivering fumigation agents to a grain depot using a drone according to claim 2, characterized in that, Methods for determining whether a dangerous situation has occurred include: The system uses environmental depth as a basis for judgment. When the distance between the drone and an obstacle is less than a preset collision safety threshold, a collision is determined to be imminent. It also uses remaining battery power as a basis for judgment. When the remaining battery power is lower than a preset low battery power threshold, the battery is determined to be too low. Finally, it uses ambient gas concentration as a basis for judgment. When the ambient gas concentration exceeds a preset lower explosive limit threshold, the gas concentration is determined to be excessive.

7. The method for delivering fumigation agents for grain depots by unmanned aerial vehicles according to claim 1, characterized in that, Emergency evacuation procedures include: If a collision is imminent, the drone will be controlled to adjust its flight attitude and move away from the obstacle to avoid the risk of collision; if the battery is too low, the drone's backup power supply will be activated; if the gas concentration exceeds the limit, the drone's drug delivery mechanism will be controlled to stop the drug delivery action and close the drug delivery mechanism, and the drone will be controlled to return to the pre-set safe docking point.

8. A drone-based grain depot fumigation agent delivery system, used to implement the drone-based grain depot fumigation agent delivery method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect and acquire sensing data in real time, including environmental information and UAV status information; The data processing module is used to preprocess the sensed data, obtain the observation vector, and encode the observation vector into belief states; The target decision-making module determines the target medicine tray based on perceived data. The first execution module is used to input the belief state into the trained SAC policy network based on the target medicine tray and output continuous movement actions; the UAV executes continuous movement actions to reach the target medicine tray and triggers the medicine dispensing command when the set medicine dispensing conditions are met. The second execution module is used to complete the drug delivery action according to the drug delivery instruction, update the drug tray map, determine the next target drug tray, and continue until the drug delivery action of all drug trays is completed.

Images