Multi-input model compensation plant protection robot pesticide application control system and control method

Abstract

The invention discloses a multi-input model compensation plant protection robot pesticide application control system and control method, and belongs to the technical field of agricultural plant protection. The system comprises an upper computer input system, a lower computer controller and a pesticide application system; the upper computer input system is based on an image processing unit of an AI edge computing platform, performs target identification by using a deep learning model, and extracts key information of target crops or weeds in real time; the lower computer controller receives target information transmitted by the upper computer, adopts a dynamic front adjustment strategy of a spray boom virtual line, and dynamically controls an execution element according to the real-time traveling speed of a vehicle, sensor data of a pesticide application system and key information of target crops or weeds, so that accurate target pesticide spraying is realized; the pesticide applying system comprises a pesticide box, a plunger pump, a flow sensor, a pressure sensor, a switch valve, an electric ball valve, an overflow valve, a PWM proportional electromagnetic valve and a nozzle array. The pesticide applying precision is improved, the pesticide using amount is reduced, and the sustainability of agricultural production is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural plant protection, and in particular to a multi-input model compensation plant protection robot pesticide application control system and control method. Background Art

[0002] A plant protection robot is a robot used for the prevention and control of crop diseases and pests. It can fertilize, spray pesticides, and monitor diseases and pests according to the needs of the crops.

[0003] Traditional plant protection robots often have problems of drug waste and uneven application during the pesticide application process. They cannot accurately allocate the spraying amount according to geographic information and crop conditions, which not only increases costs but also has adverse effects on the environment. Therefore, it is necessary to design a multi-input model compensation plant protection robot pesticide application control system and control method. Summary of the invention

[0004] The object of the present invention is to provide a multi-input model compensation plant protection robot pesticide application control system and control method to solve the problems raised in the above background technology.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a multi-input model compensation plant protection robot pesticide application control system, including an upper computer input system, a lower computer controller and a pesticide application system;

[0006] The host computer input system is based on the image processing unit of the AI ​​edge computing platform, uses a deep learning model to perform target recognition, and extracts key information of the target in real time, including boundaries, center positions, density, size and growth status of crops, health status, etc.;

[0007] The lower computer controller receives the target information transmitted by the upper computer, adopts the dynamic pre-adjustment strategy of the spray boom virtual line, and dynamically controls the actuator according to the real-time travel speed of the vehicle, the sensor data of the spraying system, and the key information of the target crop or weeds, so as to achieve precise target spraying;

[0008] The pesticide application system comprises a medicine box, a plunger pump, a flow sensor, a pressure sensor, a switch valve, an electric ball valve, a relief valve, a PWM proportional solenoid valve and a nozzle array.

[0009] In a further embodiment, the dynamic advance adjustment strategy of the spray boom virtual line combines vehicle speed, spray valve opening delay and droplet landing time to dynamically optimize the spray trigger timing, ensuring that the target crops can be accurately covered under different operating conditions, maximizing the utilization rate of the liquid medicine and reducing waste; the host computer input system continuously calculates and adjusts the advance amount based on real-time data to ensure that the system can accurately trigger the spraying operation under various environmental conditions.

[0010] In a further embodiment, the image processing unit is managed by an AI edge computing platform, equipped with a deep learning model, and performs real-time image processing and target recognition based on visual recognition and dynamic control strategies. It also optimizes image analysis and data processing through parallel computing acceleration technology to ensure high real-time and high precision of image recognition, and adapt to the accurate recognition and dynamic control of different types of targets.

[0011] In a further embodiment, during target spraying, a high-precision positioning system is used in combination with real-time motion state feedback to dynamically adjust the timing of the spraying operation to ensure accurate spraying of the target object.

[0012] A multi-input model compensation plant protection robot pesticide application control method comprises the following steps:

[0013] S1. Construct a control model for flow and pressure;

[0014] S2. Control the water pump and electric ball valve;

[0015] S3. Perform PWM controlled nozzle flow and pressure stability control;

[0016] S4. A multi-level adaptive feedback control method based on model compensation takes the target values ​​of the prescription map input system and the visual target input system as the control benchmark, controls the three types of components: water pump speed, electric ball valve opening and PWM proportional solenoid valve duty cycle, achieves pressure stability and flow control for each nozzle, adjusts flow and pressure through closed-loop feedback, and ensures uniformity and efficiency of spraying.

[0017] In a further embodiment, in step S1, the control model of flow and pressure is constructed as follows:

[0018] 1) Construct nozzle flow model:

[0019] Nozzle flow rate Q nozzle Affected by the duty cycle D of the PWM proportional solenoid valve and the main line pressure P supply The double control is calculated as follows:

[0020]

[0021] Among them, k Q is the flow gain coefficient, which is related to the nozzle structure parameters; D is the duty cycle of the PWM proportional solenoid valve, ranging from [0,1]; P supply is the main pipeline pressure;

[0022] 2) Allocate system target flow and nozzle flow:

[0023] When there is no prescription map input system and visual target input system, the fixed per mu application rate mode is adopted, and the target flow rate Q requiredThe amount of pesticide applied per unit area Q target , boom width W and vehicle speed v, calculated as follows:

[0024] Q required =Q target ·W·v;

[0025] When there is a prescription map output system and a visual target input system, Q required Equal to the sum of all nozzle target flows.

[0026] In a further embodiment, in step S2, the water pump and the electric ball valve are controlled as follows:

[0027] 1) Modeling of the pesticide application system:

[0028] The relationship between the pump flow rate and the pump speed is:

[0029] Q pump =k pump ·n pump ;

[0030] Establish the electric valve return water tank flow and opening V valve and main line pressure P supply The nonlinear relationship between:

[0031] Q recirculated =k valve ·(V valve )α·(p supply ) β ;

[0032] The total flow of the main line is the pump flow minus the electric valve return tank flow. The total flow of the main line is:

[0033] Q total =k pump ·n pump -k valve ·(V valv e) α ·(P supply ) β ;

[0034] The electric valve opening V valve The main line pressure P is obtained through the sensor or control signal of the electric valve. supply Obtained by pressure sensor, Q total The total flow rate of the main line can be obtained by installing a flow sensor in the main line. If there is no flow sensor, the total flow rate of the main line can be calculated by the above formula.

[0035] 2) Adaptive control calculation:

[0036] Flow error E Qfor:

[0037] E Q =Q required -Q total ;

[0038] The control law of the pump speed is:

[0039]

[0040] K feedback It is a user-definable feedback factor;

[0041] The opening control law of the electric valve is:

[0042]

[0043] Among them, K robust is the robust gain, λ is the constant for adjusting the feedback rate, is the time derivative of the error, and robust feedback is added to optimize the opening control of the electric valve;

[0044] 3) Control gain update:

[0045] In order to make the gain k valve , k pump And the unknown nonlinear parameters α and β can be adjusted adaptively, and an error-based online update algorithm is designed. The specific gain update is as follows:

[0046] (1) Pump gain update:

[0047]

[0048] (2) Electric valve gain update:

[0049]

[0050] (3) Adaptive parameter update:

[0051] For α and β, their values ​​are dynamically adjusted by feedback of control error and flow rate, and the main line flow Q is adjusted based on α and β. total The sensitivity is as follows:

[0052]

[0053] in, and is the sensitivity coefficient about α and β, calculated as follows:

[0054]

[0055] Among them, γ1, γ2, γ3 and γ4 are adaptive gains used to adjust the learning rate.

[0056] In a further embodiment, in step S3, in the image analysis and data processing, there are control calculations for two input conditions: prescription map and vision-based target. In the process of prescription map application and vision-based target application, the target flow rate Q of each nozzle is nozzle,i,target1 and Q nozzle,i,target2 Adjust according to different inputs, as follows:

[0057] 1) Setting the target flow rate of each nozzle:

[0058] Prescription map application: According to the longitude and latitude of the nozzle location (x i ,y i ), the target flow rate Q of each nozzle is obtained from the prescription map nozzle,i,target1 ,Right now:

[0059] Q nozzle,i,target1 =f prescription (x i ,y i );

[0060] Vision-based target application: Based on the crop density and three-dimensional structure (ρ i ,h i ), calculate the target flow rate Q of each nozzle nozzle,i,target2 ,Right now:

[0061] Q nozzle,i,target2 =g visual (ρ i ,h i );

[0062] 2) Calculation of flow error for each nozzle:

[0063] For each nozzle, the flow error ΔQ is calculated in real time i :

[0064] ΔQ i =Q nozzle,i,target -Q nozzle,i,actual

[0065] Among them, Q nozzle,i,target It is calculated by the prescription map or the visual system, and Q nozzle,i,actual It can be estimated from pressure and duty cycle or measured directly;

[0066] 3) Calculation of duty cycle for each nozzle adjustment:

[0067] Based on the real-time flow error and target flow, the duty cycle D of the PWM proportional solenoid valve corresponding to the nozzle is adjusted through adaptive control. i,new,as follows:

[0068]

[0069] Among them, ΔQ i is the flow error; K p is the adaptive gain; P supply is the real-time main line pressure; P set is the set pressure;

[0070] In order to cope with the dynamically changing environment and system uncertainty, a compensation term based on pressure feedback is added and enhanced by combining robust control. The adjusted duty cycle is:

[0071]

[0072] Where K robust is the robustness gain, which allows the controller to be more adaptable to disturbances and parameter uncertainties;

[0073] 4) Automatic adjustment of gain parameters:

[0074] According to the flow error ΔQ i and pressure error ΔP i To update the gain K p,i,new :

[0075] K p,i,new =K p,i,old +μ p,i ·(ΔQ i ΔP i )

[0076] Among them, K p,i,new is the updated gain parameter; K p,i,old is the gain parameter before updating; μ p,i The learning rate adjusted for gain is used to control the size of the update step to ensure smooth updates and avoid system instability;

[0077] ΔQ i =Q target,i -Q actual,i , is the flow error of the i-th nozzle, the target flow Q target,i and the actual flow rate Q actual,i The difference between i =P set -P supply,i , is the pressure error of the i-th nozzle, and the set pressure P set and real-time pressure P supply,i The difference between

[0078] Gain-adjusted learning rate μ p,iAdjust according to the response of the system. Usually, a smaller learning rate can reduce system oscillation, while a larger learning rate can accelerate convergence. The learning rate is adjusted as follows:

[0079]

[0080] Among them, μ 0,i is the initial learning rate; α i is the learning rate attenuation coefficient, which is used to adjust the learning rate according to the size of the flow and pressure errors; ΔQ i and ΔP i are flow error and pressure error respectively;

[0081] 5) Updated gain calculation:

[0082] Gain K of each nozzle p,i,new The calculation is as follows:

[0083]

[0084] Estimated gain K p,i,new Substituting into the above controller formula, the duty cycle signal of the PWM proportional solenoid valve corresponding to each nozzle is obtained as follows:

[0085]

[0086] The above algorithm dynamically adjusts the difference between the target flow rate and the actual flow rate of each nozzle, and uses adaptive gain to adjust the duty cycle to ensure that the nozzle can accurately control the flow rate in different spraying scenarios to achieve the effect of precise spraying;

[0087] The actual flow rate of each nozzle is adjusted by the duty cycle of the PWM proportional solenoid valve, and combined with pressure feedback, dual control of flow and pressure is achieved.

[0088] In a further embodiment, in step S3, for the control of drug application when neither the prescription map nor vision is available, the calculation based on nozzle pressure compensation and local gain adjustment is as follows:

[0089] 1) Nozzle pressure feedback:

[0090] The pressure difference of each nozzle may cause different flow rates. Therefore, for each nozzle, the duty cycle is corrected using local pressure feedback. The pressure difference of each nozzle is expressed by P nozzle,i It means that the modified duty cycle of each nozzle is dynamically adjusted according to its pressure difference;

[0091] Introducing the global gain K global To adjust the control output of each nozzle to ensure that the flow of the entire system meets the target demand, this is achieved by weighting based on the local duty cycle, as follows:

[0092]

[0093] 2) Adaptive control calculation:

[0094] Adopt neural network adaptive algorithm to dynamically adjust the gain K global , a neural network model is introduced to estimate the pressure difference of the nozzle, and the gain is automatically adjusted according to the real-time flow error and pressure feedback. The gain is calculated as follows:

[0095]

[0096] Among them, K global is the current value of the global gain; α is the learning rate; is the gradient of the loss function with respect to the gain;

[0097] The loss function J is the sum of squares of flow errors or other appropriate performance indicators, calculated as follows:

[0098]

[0099] 3) The comprehensive control equation of nozzle flow and pressure stability is as follows:

[0100]

[0101] In a further embodiment, the duty cycle is adjusted by real-time flow error to ensure that the nozzle flow is close to the target value, and closed-loop flow control is performed. At the same time, based on the real-time pressure feedback of each nozzle, the duty cycle is dynamically adjusted to compensate for the pressure difference and optimize the nozzle injection effect. An adaptive algorithm is used to adjust the gain online to improve the adaptability and robustness of the system.

[0102] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: precise prescription map application accurately calculates the amount of medicine required by crops and accurately allocates the spraying amount according to geographic information and crop conditions; visual target application utilizes real-time image recognition technology to dynamically adjust the spraying amount to ensure that the medicine only acts on the target crop to avoid waste; by introducing model compensation and online adaptive algorithms, the system can dynamically adjust the control gain according to real-time monitored parameters such as flow and pressure to cope with the uncertainty and nonlinear changes in the amount of medicine applied under complex environments, which has promoted the advancement of precision application technology in the agricultural field, reduced the use of pesticides, reduced environmental pollution, and at the same time improved the efficiency and sustainability of agricultural production, providing important technical support for promoting the development of smart agriculture and green agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 It is a flow chart of the pesticide application control method of the present invention;

[0104] Figure 2It is a schematic diagram of the composition of the pesticide application system of the present invention;

[0105] Figure 3 It is a flow chart of the composition of the target spraying system of the present invention;

[0106] Figure 4 This is a schematic diagram of the field of view division of the target medication image of the present invention;

[0107] Figure 5 It is a schematic diagram of the target dosing collision decision algorithm of the present invention. DETAILED DESCRIPTION

[0108] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known in the art are not described.

[0109] See also Figure 1-Figure 5 ,The present invention provides a technical solution: a multi-input model compensation plant protection robot pesticide application control system, including an upper computer input system, a lower computer controller and a pesticide application system;

[0110] The host computer input system is based on the image processing unit of the AI ​​edge computing platform, and uses a deep learning model to perform target recognition and extract key information of the target in real time, including boundaries, center positions, density, size and growth status of crops, health status, etc. The target includes crops, weeds, pests, etc.

[0111] The lower computer controller receives the target information transmitted by the upper computer, adopts the dynamic pre-adjustment strategy of the boom virtual line, and dynamically controls the actuators according to the real-time vehicle speed, the sensor data of the spraying system, and the key information of the target crops or weeds to achieve precise target spraying. The actuators include three types of components: water pump speed, electric ball valve opening, and PWM proportional solenoid valve duty cycle;

[0112] The pesticide application system includes a medicine box, a plunger pump, a flow sensor, a pressure sensor, a switch valve, an electric ball valve, a relief valve, a PWM proportional solenoid valve and a nozzle array.

[0113] In a further embodiment, the dynamic advance adjustment strategy of the spray boom virtual line combines vehicle speed, spray valve opening delay and droplet landing time to dynamically optimize the spray trigger timing, ensuring accurate coverage of target crops under different operating conditions, maximizing liquid medicine utilization and reducing waste; the host computer input system continuously calculates and adjusts the advance amount based on real-time data to ensure that the system can accurately trigger the spraying operation under various environmental conditions.

[0114] The system divides the field of view of the binocular camera into N areas, such as Figure 4 As shown, each area corresponds to a nozzle, thereby realizing independent spraying control. The image acquisition device collects image data of the crop area in front in real time and transmits the data to the AI ​​edge computer for processing. The platform subscribes to the image data stream under the ROS framework, executes the target recognition model, and identifies the position and size of the target crop in each area in real time, providing accurate location information for spraying decisions.

[0115] like Figure 5 As shown in the figure, the virtual horizontal line represents the position of the spray boom in the binocular camera image coordinate system. The target detected in the image is matched with the nozzle on the spray boom to generate a solenoid valve control data frame, thereby accurately controlling the nozzle's spraying of the target. The virtual horizontal line divides the field of view into n+1 areas through n spraying area dividing lines. Each horizontal line is divided into several sections by the spraying area dividing line, where each spraying interval corresponds to the number of solenoid valves on the spray boom, and the length of each area is the average spraying width of each solenoid valve.

[0116] As the sprayer moves forward, the ground image moves downward, and the host computer (AI edge computer) detects and draws the target prediction box in real time through a deep learning target detection algorithm. The virtual horizontal line is used to determine whether the target enters the spraying area. When the prediction box intersects with the virtual horizontal line of the spray boom in the image, it means that the target enters the corresponding spraying interval. The spraying timing is determined based on the intersection of the virtual horizontal line of the spray boom in the image and the target crop position.

[0117] When the line connecting the front midpoints of the target crop prediction frame intersects with the virtual horizontal line of the pre-adjusted spray boom in the image, the system's deep learning model predicts the target information in each spraying area, namely, the boundary, center position, density, crop size and growth status, health status and other information. Through the communication interface, the pixel area data is transmitted to the lower computer controller. After receiving the data, the lower computer converts it into the nozzle target flow signal of the corresponding area. The target flow value of the nozzle is positively correlated with the target information. The mapping relationship between the two is obtained through experimental data regression. Through this mechanism, the system can adjust the spraying amount according to the crop size in real time to ensure accurate application of the pesticide.

[0118] In order to ensure the accuracy of the target, the system adopts a dynamic advance adjustment strategy for the virtual line of the spray boom. Due to the influence of the vehicle's driving speed, the virtual line needs to be dynamically adjusted in advance according to the vehicle speed, spray valve opening time and droplet landing time. As the vehicle speed increases, the advance amount of the virtual line increases accordingly to trigger the spraying operation in advance and avoid spraying delays. At the same time, the spray valve opening response time and the time required for the liquid to be sprayed out of the nozzle will also affect the adjustment of the advance amount to ensure that the liquid is accurately sprayed above the target crops.

[0119] Physical front distance Dphys is the physical distance of the boom relative to the camera, given by the vehicle's speed v car , spray valve opening time t open and droplet landing time t fall To calculate, at this time, the physical front distance refers to the actual distance from the reference point of the camera coordinate system to the bottom of the spray boom. The calculation formula is as follows:

[0120] D phys =v car ×(t open +t fall )

[0121] Among them, v car is the vehicle speed; t open is the spray valve opening time; t fall is the average droplet falling time;

[0122] In order to convert the physical front distance into the front pixel distance in the image coordinate system, the geometric relationship of the camera is used. Without relying on the depth map, it is assumed that the focal length of the camera is f and the ground projection width along the robot's forward direction, that is, the field of view width is W. view , then the front pixel distance D pix Distance from physical front D phys The relationship between them is:

[0123]

[0124] Where f is the focal length of the camera; D phys is the physical lead distance; W view is the width of the camera's field of view; D pix is the pixel distance of the front distance in image coordinates.

[0125] The virtual line's pre-adjustment mechanism combines vehicle speed, spray valve opening delay and droplet landing time to dynamically optimize the spraying trigger timing, ensuring that the target crops can be accurately covered under different operating conditions, maximizing the utilization rate of the liquid medicine and reducing waste. The upper computer continuously calculates and adjusts the pre-adjustment amount based on real-time data to ensure that the system can accurately trigger the spraying operation under various environmental conditions.

[0126] Through the above hardware design and control strategy, the system can achieve efficient and precise spraying control in complex field environments. The global decision-making of the upper computer is combined with the precise execution of the lower computer controller to ensure the real-time and accuracy of the spraying operation, reduce pesticide waste, and meet the high requirements of precision agriculture for operation accuracy and efficiency.

[0127] In a further embodiment, the image processing unit is managed by the AI ​​edge computing platform, which is equipped with a deep learning model and performs real-time image processing and target recognition based on visual recognition and dynamic control strategies. It also optimizes image analysis and data processing through parallel computing acceleration technology to ensure high real-time and high precision of image recognition, and adapt to the accurate recognition and dynamic control of different types of targets.

[0128] The system uses an Intel D435i binocular camera as an image acquisition device, which has high-resolution depth perception capabilities, a built-in IMU, and a wide dynamic range sensor. It can provide clear images in strong or weak light conditions. Through depth perception technology, the camera can accurately obtain the distance information of the target crop. The camera is installed parallel to the ground and dynamically adjusts the field of view according to the robot's travel path to ensure full coverage of the crop area and provide high-quality visual input for subsequent target recognition.

[0129] The target recognition model is based on a deep learning model, including YOLO, Unet, etc., which is developed in Python or C language and deployed on the host computer. The model efficiently identifies the target crops through real-time image processing to ensure spraying accuracy. During the training process, the deep learning model uses a large amount of labeled data for supervised learning, mainly including the labeling of crop images, soil and open space areas, to ensure that the model can accurately distinguish crops from backgrounds in complex environments.

[0130] The training data includes diverse samples of different lighting conditions, crop densities and shapes. After data enhancement processing, they are input into the deep learning model for training. The model training optimizes the recognition accuracy by gradually adjusting the weights of the convolutional neural network to ensure that it can cope with crop recognition tasks in different environments. The training process uses cross-validation and real-time feedback mechanisms to optimize the generalization ability of the model, enabling it to work stably under a variety of complex lighting and environmental conditions.

[0131] During the deployment phase, the deep learning model is integrated into the host computer software, mainly to realize real-time image processing, environmental adaptability adjustment and target data transmission. The host computer receives the real-time image stream from the binocular camera and uses the deep learning model to segment the image into target crops, soil and open space areas, and identifies the position, shape and size of the target crops in real time. The recognition results include feature information such as crop area, area and density, and are transmitted to the slave computer through the RS232 interface.

[0132] In order to cope with changes in lighting, the upper computer software implements automatic exposure adjustment and image contrast enhancement, ensuring recognition stability under strong or weak light conditions. The environmental adaptability processing enables the system to maintain high recognition accuracy under dynamically changing lighting conditions. The identified target data not only helps the spraying system to accurately locate crops, but also optimizes the spraying amount and coverage of the liquid medicine, reducing pesticide waste.

[0133] In a further embodiment, during target spraying, a high-precision positioning system is used in combination with real-time motion state feedback to dynamically adjust the timing of the spraying operation to ensure accurate spraying of the target object.

[0134] The hardware design of the lower computer controller is based on the DSP C2000 series MCU, integrating multiple functional modules to ensure high-precision spraying control and real-time data processing. The core functional modules include ADC module, EQEP module, PWM module, GPIO interface and RS232 communication interface. The modules work together to achieve accurate pressure and flow acquisition and control.

[0135] Pressure sensor data acquisition: The lower computer collects data from the pressure sensor in real time through the integrated ADC module. The MCU provides a 16-bit resolution ADC. The sampling speed of each ADC channel can reach 1.1MSPS and supports up to 12 external input channels. These high-precision data help the system monitor the pressure changes of the spraying system in real time to ensure the stability and uniformity of the spraying process. Accurate pressure monitoring provides reliable data support for controlling spraying accuracy.

[0136] Flow sensor data acquisition: The EQEP module (Enhanced Quadrature Encoder Pulse) is used to collect feedback data from the flow sensor. The EQEP module provides high-precision flow feedback to help the controller accurately adjust the spraying volume. Through this feedback, the system can adjust the nozzle flow in real time to ensure the consistency and accuracy of the spraying effect and reduce liquid waste.

[0137] Communication and control signal transmission: Through the RS232 interface, the lower computer and the upper computer transmit data to ensure the timely transmission of control signals. It is responsible for receiving the target recognition data transmitted by the AI ​​edge computer and accurately controlling the opening and closing of the nozzle according to the position and size of the target crop. The PWM control signal of each nozzle is precisely adjusted by the lower computer controller. The PWM module generates a high-precision control signal to adjust the switch of the nozzle and control the spraying amount of the liquid medicine. The fine PWM signal control allows the spraying amount to be adjusted in real time according to demand.

[0138] Proportional flow valve control: The system links the GPIO pin with the MOS tube amplifier circuit to accurately control the proportional flow valve. The GPIO pin receives the signal generated by the MCU, and the MOS tube amplifies the signal to drive the opening of the proportional flow valve, ensuring that the flow and pressure during the spraying process are accurately controlled, and ensuring that each nozzle is accurately adjusted according to real-time feedback data to avoid waste of liquid medicine and crop damage.

[0139] Enhanced PWM module: The controller is equipped with 24 enhanced PWM channels, which can accurately control the flow and pressure during the spraying process. By adjusting the PWM frequency and duty cycle, the system can finely control the spraying volume of the nozzle to ensure the uniformity and stability of the liquid distribution. The high-resolution PWM channel ensures the response speed and accuracy of the nozzle to adapt to the dynamically changing spraying needs.

[0140] A multi-input model compensation plant protection robot pesticide application control method comprises the following steps:

[0141] S1. Construct a control model for flow and pressure;

[0142] S2. Control the water pump and electric ball valve;

[0143] S3. Perform PWM controlled nozzle flow and pressure stability control;

[0144] S4. A multi-level adaptive feedback control method based on model compensation takes the target values ​​of the prescription map input system and the visual target input system as the control benchmark, controls the three types of components: water pump speed, electric ball valve opening and PWM proportional solenoid valve duty cycle, achieves pressure stability and flow control for each nozzle, adjusts flow and pressure through closed-loop feedback, and ensures uniformity and efficiency of spraying.

[0145] The flow of the nozzle is regulated by a PWM solenoid valve, and the duty cycle has a nonlinear functional relationship with the flow. The flow target of each nozzle is determined by the prescription map variable application or visual target application strategy. During the spraying process, the actual flow of each nozzle is measured in real time, and the flow error is calculated. The duty cycle of the nozzle is adjusted to make its flow close to the expected value. In addition, based on the real-time pressure of the main line, the duty cycle of the nozzle is dynamically adjusted to ensure that the flow is not affected by pressure fluctuations.

[0146] In a further embodiment, in step S1, the control model of flow and pressure is constructed as follows:

[0147] 1) Construct nozzle flow model:

[0148] Nozzle flow rate Q nozzle Affected by the duty cycle D of the PWM proportional solenoid valve and the main line pressure P supply The double control is calculated as follows:

[0149]

[0150] Among them, k Q is the flow gain coefficient, which is related to the nozzle structure parameters; D is the duty cycle of the PWM proportional solenoid valve, ranging from [0,1]; P supply is the main pipeline pressure;

[0151] 2) Allocate system target flow and nozzle flow:

[0152] When there is no prescription map input system and visual target input system, the fixed per mu application rate mode is adopted, and the target flow rate Q required The amount of pesticide applied per unit area Q target , boom width W and vehicle speed v, calculated as follows:

[0153] Q required =Q target ·W·v;

[0154] When there is a prescription map output system and a visual target input system, Q required Equal to the sum of all nozzle target flows.

[0155] In a further embodiment, in step S2, the water pump and the electric ball valve are controlled as follows:

[0156] 1) Modeling of the pesticide application system:

[0157] The relationship between the pump flow rate and the pump speed is:

[0158] Q pump =k pump ·n pump ;

[0159] Establish the electric valve return water tank flow and opening V valve and main line pressure P supply The nonlinear relationship between:

[0160] Q recirculated =k valve ·(V valve ) α ·(P supply ) β ;

[0161] The total flow of the main line is the pump flow minus the electric valve return tank flow. The total flow of the main line is:

[0162] Q total =k pump ·n pump -k valve ·(V valve ) α ·(P supply )β ;

[0163] The electric valve opening V valve The main line pressure P is obtained through the sensor or control signal of the electric valve. supply Obtained by pressure sensor, Q total The total flow rate of the main line can be obtained by installing a flow sensor in the main line. If there is no flow sensor, the total flow rate of the main line can be calculated by the above formula.

[0164] 2) Adaptive control calculation:

[0165] Flow error E Q For: E Q =Q required -Q total ;

[0166] The control law of the pump speed is:

[0167]

[0168] K feedback It is a user-definable feedback factor;

[0169] The opening control law of the electric valve is:

[0170]

[0171] Among them, K robust is the robust gain, λ is the constant for adjusting the feedback rate, The time derivative of the error is used to add robust feedback to optimize the opening control of the electric valve. Specifically, the opening control of the electric valve should not only consider model compensation (adaptive adjustment of gain), but also adjust the control input according to the flow error and error change to achieve the purpose of enhancing system stability and robustness.

[0172] 3) Control gain update:

[0173] In order to make the gain k valve , k pump And the unknown nonlinear parameters α and β can be adjusted adaptively, and an error-based online update algorithm is designed. The specific gain update is as follows:

[0174] (1) Pump gain update:

[0175]

[0176] (2) Electric valve gain update:

[0177]

[0178] (3) Adaptive parameter update:

[0179] For α and β, their values ​​are dynamically adjusted by feedback of control error and flow rate, and the main line flow Q is adjusted based on α and β. total The sensitivity is as follows:

[0180]

[0181] in, and is the sensitivity coefficient about α and β, calculated as follows:

[0182]

[0183] Among them, γ1, γ2, γ3 and γ4 are adaptive gains used to adjust the learning rate. Through the above adaptive update, the opening control law of the electric valve updates the gain parameters in real time, that is, the water pump gain k pump , Electric valve gain k valve The unknown nonlinear parameters α and β can be adjusted in real time according to the control error, thereby achieving precise flow control. With the introduction of robust feedback terms, the control system can better cope with external disturbances, system parameter uncertainties and model errors, thereby enhancing the robustness and stability of the system.

[0184] In a further embodiment, in step S3, in the image analysis and data processing, there are control calculations for two input conditions: prescription map and vision-based target. In the process of prescription map application and vision-based target application, the target flow rate Q of each nozzle is nozzle,i,target1 and Q nozzle,i,target2 Adjust according to different inputs, as follows:

[0185] 1) Setting the target flow rate of each nozzle:

[0186] Prescription map application: According to the longitude and latitude of the nozzle location (x i ,y i ), the target flow rate Q of each nozzle is obtained from the prescription map nozzle,i,target1 ,Right now:

[0187] Q nozzle,i,target1 =f prescription (x i ,y i );

[0188] Vision-based target application: Based on the crop density and three-dimensional structure (ρ i ,h i ), calculate the target flow rate Q of each nozzle nozzle,i,target2 ,Right now:

[0189] Q nozzle,i,target2 =gvisual (ρ i ,h i );

[0190] 2) Calculation of flow error for each nozzle:

[0191] For each nozzle, the flow error ΔQ is calculated in real time i :

[0192] ΔQ i =Q nozzle,i,target -Q nozzle,i,actual

[0193] Among them, Q nozzle,i,target It is calculated by the prescription map or the visual system, and Q nozzle,i,actual It can be estimated from pressure and duty cycle or measured directly;

[0194] 3) Calculation of duty cycle for each nozzle adjustment:

[0195] Based on the real-time flow error and target flow, the duty cycle D of the PWM proportional solenoid valve corresponding to the nozzle is adjusted through adaptive control. i,new ,as follows:

[0196]

[0197] Among them, ΔQ i is the flow error; K p is the adaptive gain; P supply is the real-time main line pressure; P set is the set pressure;

[0198] In order to cope with the dynamically changing environment and system uncertainty, a compensation term based on pressure feedback is added and enhanced by combining robust control. The adjusted duty cycle is:

[0199]

[0200] Where K robust is the robustness gain, which allows the controller to be more adaptable to disturbances and parameter uncertainties;

[0201] 4) Automatic adjustment of gain parameters:

[0202] According to the flow error ΔQ i and pressure error ΔP i To update the gain K p,i,new :

[0203] K p,i,new =K p,i,old +μ p,i ·(ΔQ i ΔP i)

[0204] Among them, K p,i,new is the updated gain parameter; K p,i,old is the gain parameter before updating; μ p,i is the learning rate adjusted by the gain, which is used to control the size of the update step to ensure smooth updates and avoid system instability; ΔQ i =Q target,i -Q actual,i , is the flow error of the i-th nozzle, the target flow Q target,i and the actual flow rate Q actual,i The difference between i =P set -P supply,i , is the pressure error of the i-th nozzle, and the set pressure P set and real-time pressure P supply,i The difference between

[0205] Gain-adjusted learning rate μ p,i Adjust according to the response of the system. Usually, a smaller learning rate can reduce system oscillation, while a larger learning rate can accelerate convergence. The learning rate is adjusted as follows:

[0206]

[0207] Among them, μ 0,i is the initial learning rate; α i is the learning rate attenuation coefficient, which is used to adjust the learning rate according to the size of the flow and pressure errors; ΔQ i and ΔP i They are flow error and pressure error respectively; through this method of dynamically adjusting the learning rate, the gain update can be made more flexible, and the update step size can be automatically increased when the flow and pressure errors are large to improve the response speed of the system.

[0208] 5) Updated gain calculation:

[0209] Gain K of each nozzle p,i,new The calculation is as follows:

[0210]

[0211] Estimated gain K p,i,new Substituting into the above controller formula, the duty cycle signal of the PWM proportional solenoid valve corresponding to each nozzle is obtained as follows:

[0212]

[0213] The above algorithm is used to dynamically adjust the difference between the target flow and actual flow of each nozzle, and the adaptive gain is used to adjust the duty cycle to ensure that the nozzle can accurately control the flow in different application scenarios to achieve the effect of precise application of pesticides; the actual flow of each nozzle is adjusted by the duty cycle of the PWM proportional solenoid valve, combined with pressure feedback to achieve dual control of flow and pressure.

[0214] In a further embodiment, in step S3, for the control of spraying under the condition that neither the prescription map nor the vision is available, considering the difference in pressure of each nozzle, an algorithm based on nozzle pressure compensation and local gain adjustment is designed to make the control of each nozzle more accurate, as follows:

[0215] 1) Nozzle pressure feedback:

[0216] The pressure difference of each nozzle may cause different flow rates. Therefore, for each nozzle, the duty cycle is corrected using local pressure feedback. The pressure difference of each nozzle is expressed by P nozzle,i It means that the modified duty cycle of each nozzle is dynamically adjusted according to its pressure difference;

[0217] Introducing the global gain K global To adjust the control output of each nozzle to ensure that the flow of the entire system meets the target demand, this is achieved by weighting based on the local duty cycle, as follows:

[0218]

[0219] 2) Adaptive control calculation:

[0220] Considering that the nozzle pressure may change in actual operation and the pressure difference between each nozzle may be large, a neural network adaptive algorithm is used to dynamically adjust the gain K global In order to better cope with complex and changing working conditions, a neural network model is introduced to estimate the pressure difference of the nozzle, and the gain is automatically adjusted according to the real-time flow error and pressure feedback. The gain update formula is:

[0221]

[0222] Among them, K global is the current value of the global gain; α is the learning rate; is the gradient of the loss function with respect to the gain; the loss function J can be selected as the sum of squares of flow errors or other suitable performance indicators:

[0223]

[0224] 3) The comprehensive control equation of nozzle flow and pressure stability is as follows:

[0225]

[0226] In a further embodiment, the duty cycle is adjusted by real-time flow error to ensure that the nozzle flow is close to the target value, and closed-loop flow control is performed. At the same time, based on the real-time pressure feedback of each nozzle, the duty cycle is dynamically adjusted to compensate for the pressure difference and optimize the nozzle injection effect. An adaptive algorithm is used to adjust the gain online to improve the adaptability and robustness of the system.

[0227] The preferred specific embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above specific embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical scheme of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A multi-input model compensation plant protection robot pesticide application control system, characterized in that: It includes upper computer input system, lower computer controller and pesticide application system; The host computer input system is based on the image processing unit of the AI ​​edge computing platform, uses a deep learning model to perform target recognition, and extracts key information of the target in real time, including boundaries, center positions, density, size and growth status of crops, health status, etc.; The lower computer controller receives the target information transmitted by the upper computer, adopts the dynamic pre-adjustment strategy of the spray boom virtual line, and dynamically controls the actuator according to the real-time travel speed of the vehicle, the sensor data of the spraying system, and the key information of the target crop or weeds, so as to achieve precise target spraying; The pesticide application system comprises a medicine box, a plunger pump, a flow sensor, a pressure sensor, a switch valve, an electric ball valve, a relief valve, a PWM proportional solenoid valve and a nozzle array.

2. A multi-input model compensation plant protection robot pesticide application control system according to claim 1, characterized in that: The dynamic advance adjustment strategy of the spray boom virtual line combines vehicle speed, spray valve opening delay and droplet landing time to dynamically optimize the spray trigger timing, ensuring that the target crops can be accurately covered under different operating conditions, maximizing the utilization rate of the liquid medicine and reducing waste; the upper computer input system continuously calculates and adjusts the advance amount based on real-time data to ensure that the system can accurately trigger the spraying operation under various environmental conditions.

3. A multi-input model compensation plant protection robot pesticide application control system according to claim 1, characterized in that: The image processing unit is managed by the AI ​​edge computing platform, which is equipped with a deep learning model and performs real-time image processing and target recognition based on visual recognition and dynamic control strategies. It also optimizes image analysis and data processing through parallel computing acceleration technology to ensure high real-time and high precision of image recognition, and adapt to the accurate recognition and dynamic control of different types of targets.

4. A multi-input model compensation plant protection robot pesticide application control system according to claim 1, characterized in that: In target spraying, a high-precision positioning system is used in combination with real-time motion status feedback to dynamically adjust the timing of the spraying operation to ensure accurate spraying of the target.

5. A multi-input model compensation plant protection robot pesticide application control method, characterized in that: The steps include: S1. Construct a control model for flow and pressure; S2. Control the water pump and electric ball valve; S3. Perform PWM controlled nozzle flow and pressure stability control; S4. A multi-level adaptive feedback control method based on model compensation takes the target values ​​of the prescription map input system and the visual target input system as the control benchmark, controls the three types of components: water pump speed, electric ball valve opening and PWM proportional solenoid valve duty cycle, achieves pressure stability and flow control for each nozzle, adjusts flow and pressure through closed-loop feedback, and ensures uniformity and efficiency of spraying.

6. A multi-input model compensation plant protection robot pesticide application control method according to claim 5, characterized in that: In step S1, the control model of flow rate and pressure is constructed as follows: 1) Construct nozzle flow model: Nozzle flow rate Q nozzle Affected by the duty cycle D of the PWM proportional solenoid valve and the main line pressure P supply The double control is calculated as follows: Among them, k Q is the flow gain coefficient, which is related to the nozzle structure parameters; D is the duty cycle of the PWM proportional solenoid valve, ranging from [0,1]; P supply is the main pipeline pressure; 2) Allocate system target flow and nozzle flow: When there is no prescription map input system and visual target input system, the fixed per mu application rate mode is adopted, and the target flow rate Q required The amount of pesticide applied per unit area Q target , boom width W and vehicle speed v, calculated as follows: Q required =Q target ·W·v; When there is a prescription map output system and a visual target input system, Q required Equal to the sum of all nozzle target flows.

7. A multi-input model compensation plant protection robot pesticide application control method according to claim 6, characterized in that: In step S2, the water pump and the electric ball valve are controlled as follows: 1) Modeling of the pesticide application system: The relationship between the pump flow rate and the pump speed is: Q pump =k pump ·n pump ; Establish the electric valve return water tank flow and opening V valve and main line pressure P supply The nonlinear relationship between: Q recirculated =k valve ·(V valve ) α ·(p supply ) β ; The total flow of the main line is the pump flow minus the electric valve return tank flow. The total flow of the main line is: Q total =k pump ·n pump -k valve ·(V valve ) α ·(P supply ) β ; The electric valve opening V valve The main line pressure P is obtained through the sensor or control signal of the electric valve. supply Obtained by pressure sensor, Q total The total flow rate of the main line can be obtained by installing a flow sensor in the main line. If there is no flow sensor, the total flow rate of the main line can be calculated by the above formula. 2) Adaptive control calculation: Flow error E Q for: E Q =Q required -Q total ; The control law of the pump speed is: K feedback It is a user-definable feedback factor; The opening control law of the electric valve is: Among them, K robust is the robust gain, λ is the constant for adjusting the feedback rate, is the time derivative of the error, and robust feedback is added to optimize the opening control of the electric valve; 3) Control gain update: In order to make the gain k valve , k pump And the unknown nonlinear parameters α and β can be adjusted adaptively, and an error-based online update algorithm is designed. The specific gain update is as follows: (1) Pump gain update: (2) Electric valve gain update: (3) Adaptive parameter update: For α and β, their values ​​are dynamically adjusted by feedback of control error and flow rate, and the main line flow Q is adjusted based on α and β. total The sensitivity is as follows: in, and is the sensitivity coefficient about α and β, calculated as follows: Among them, γ1, γ2, γ3 and γ4 are adaptive gains used to adjust the learning rate.

8. The multi-input model compensation plant protection robot pesticide application control method according to claim 7, characterized in that: In step S3, in the image analysis and data processing, there are two input conditions of control calculation: prescription map and vision-based target. In the process of prescription map application and vision-based target application, the target flow rate Q of each nozzle is nozzle,i,target1 and Q nozzle,i,target2 Adjust according to different inputs, as follows: 1) Setting the target flow rate of each nozzle: Prescription map application: According to the longitude and latitude of the nozzle location (x i ,y i ), the target flow rate Q of each nozzle is obtained from the prescription map nozzle,i,target1 ,Right now: Q nozzle,i,target1 =f prescription (x i ,y i ); Vision-based target application: Based on the crop density and three-dimensional structure (ρ i ,h i ), calculate the target flow rate Q of each nozzle nozzle,i,target2 ,Right now: Q nozzle,i,target2 =g visual (ρ i ,h i ); 2) Calculation of flow error for each nozzle: For each nozzle, the flow error ΔQ is calculated in real time i : ΔQ i =Q nozzle,i,target -Q nozzle,i,actual Among them, Q nozzle,i,target It is calculated by the prescription map or the visual system, and Q nozzle,i,actual Through pressure and duty cycle estimation or direct measurement by flow sensor; 3) Calculation of duty cycle for each nozzle adjustment: Based on the real-time flow error and target flow, the duty cycle D of the PWM proportional solenoid valve corresponding to the nozzle is adjusted through adaptive control. i,new ,as follows: Among them, ΔQ i is the flow error; K p is the adaptive gain; P supply is the real-time main line pressure; P set is the set pressure; In order to cope with the dynamically changing environment and system uncertainty, a compensation term based on pressure feedback is added and enhanced by combining robust control. The adjusted duty cycle is: Where K robust is the robustness gain, which allows the controller to be more adaptable to disturbances and parameter uncertainties; 4) Automatic adjustment of gain parameters: According to the flow error ΔQ i and pressure error ΔP i To update the gain K p,i,new : K p,i,new =K p,i,old +μ p,i ·(ΔQ i ·ΔP i ) Among them, K p,i,new is the updated gain parameter; K p,i,old is the gain parameter before updating; μ p,i The learning rate adjusted for gain is used to control the size of the update step to ensure smooth updates and avoid system instability; ΔQ i =Q target,i -Q actual,i , is the flow error of the i-th nozzle, the target flow Q target,i and the actual flow rate Q actual,i The difference between i =P set -P supply,i , is the pressure error of the i-th nozzle, and the set pressure P set and real-time pressure P supply,i The difference between Gain-adjusted learning rate μ p,i Adjust according to the response of the system. Usually, a smaller learning rate can reduce system oscillation, while a larger learning rate can accelerate convergence. The learning rate is adjusted as follows: Among them, μ 0,i is the initial learning rate; α i is the learning rate attenuation coefficient, which is used to adjust the learning rate according to the size of the flow and pressure errors; ΔQ i and ΔP i are flow error and pressure error respectively; 5) Updated gain calculation: Gain K of each nozzle p,i,new The calculation is as follows: Estimated gain K p,i,new Substituting into the above controller formula, the duty cycle signal of the PWM proportional solenoid valve corresponding to each nozzle is obtained as follows: The above algorithm dynamically adjusts the difference between the target flow rate and the actual flow rate of each nozzle, and uses adaptive gain to adjust the duty cycle to ensure that the nozzle can accurately control the flow rate in different spraying scenarios to achieve the effect of precise spraying; The actual flow rate of each nozzle is adjusted by the duty cycle of the PWM proportional solenoid valve, and combined with pressure feedback, dual control of flow and pressure is achieved.

9. A multi-input model compensation plant protection robot pesticide application control method according to claim 8, characterized in that: In step S3, for the control of drug application when neither the prescription map nor the vision is available, the calculation based on the nozzle pressure compensation and the local gain adjustment is as follows: 1) Nozzle pressure feedback: The pressure difference of each nozzle may cause different flow rates. Therefore, for each nozzle, the duty cycle is corrected using local pressure feedback. The pressure difference of each nozzle is calculated by P nozzle,i It means that the modified duty cycle of each nozzle is dynamically adjusted according to its pressure difference; Introducing the global gain K global To adjust the control output of each nozzle to ensure that the flow of the entire system meets the target demand, this is achieved by weighting based on the local duty cycle, as follows: 2) Adaptive control calculation: Adopt neural network adaptive algorithm to dynamically adjust the gain K global , a neural network model is introduced to estimate the pressure difference of the nozzle, and the gain is automatically adjusted according to the real-time flow error and pressure feedback. The gain is calculated as follows: Among them, K global is the current value of the global gain; α is the learning rate; is the gradient of the loss function with respect to the gain; The loss function J is the sum of squares of flow errors or other appropriate performance indicators, calculated as follows: 3) The comprehensive control equation of nozzle flow and pressure stability is as follows:

10. A multi-input model compensation plant protection robot pesticide application control method according to claim 9, characterized in that: The duty cycle is adjusted through real-time flow error to ensure that the nozzle flow is close to the target value and to perform closed-loop flow control. At the same time, based on the real-time pressure feedback of each nozzle, the duty cycle is dynamically adjusted to compensate for the pressure difference and optimize the nozzle injection effect. Adaptive algorithm is used to adjust the gain online to improve the adaptability and robustness of the system.

Images